Climber behavioral gate

Companion to /audit-adaptivity (which audits the code): this skill regenerates the deterministic adversarial traces behind .claude/docs/hill-climber.md §3 and re-runs the real cache (product.Caffeine) against fresh LRU / static-window-ceiling anchors. Run it after any change to WindowClimber or the window-resize logic in BoundedLocalCache. This directory is also the durable home of the trap generators — the trace files themselves are disposable and regenerate bit-identically (verified against the 2026-07 study's battery).

Harness

• Generate traces into a scratch dir; every generator is seeded and deterministic.

• The simulator drives the real library: ./gradlew simulator:run -q -Dcaffeine.simulator.maximum-size=<N> -Dcaffeine.simulator.files.paths.0=<fmt>:<path> -Dcaffeine.simulator.policies.0=product.Caffeine -Dcaffeine.simulator.policies.1=linked.Lru -Dcaffeine.simulator.admission.0=Always -Dcaffeine.simulator.report.format=csv. Always pass admission.0=Always — the default admission list includes Clairvoyant, which materializes the whole trace.

• probe.py <trace> --size N is the per-cell verdict instrument: fresh LRU / Belady / best-static-window anchors, the real cache's hit rate, and a TRAP? flag when the gap to the ceiling is under −3pp. Run it from the repo root (or set CAF_TREE).

• The reactive arm is a third anchor, not just an ablation. LRU says "better than doing nothing"; the static-window ceiling says "how much is on the table"; the reactive arm says "is the density machine earning its complexity on this row" — which is the only one of the three that can retire code. These families were constructed to attack the density signal, so most of them are not traps for the hit-rate law at all and "far above LRU" is a bar that hides that. Measured 2026-08-04: on the deterministic families the reactive arm is level or ahead (whisper 66.04 vs 66.78, mixture d025@32k 56.46 vs 58.48, mixture d050 long 58.44 vs 58.18), and where the machine is clearly better the gain is often a spread rather than a level — straywall2 reactive 48.12 ±19.84 against 55.07±0.03, demoflood 61.83 ±9.87 against 68.56±0.05. A single draw on those two reads 54.71 / 68.40, i.e. near-parity, and is wrong; never quote a reactive comparison from one run. Like LRU and the ceiling, the value is a property of the trace instance rather than of the change under test, so it is measured once and recorded in the table's reactive column — a routine change-verification run does not need to re-measure it. Re-measure when the instance is regenerated by a new generator revision, or when a change touches shared climber state (Step, Sample, the sample period, the resize path) rather than the density tier alone. Refresh it with gate.py <traces> hybrid,reactive out.csv, which pairs the arms inside each run; it needs the harness worktree, since reactive is a harness knob.

• The static-window anchor is weighted-capable (2026-08-02; before that date weighted runs silently dropped it — "no curve" in under a second): sketch.WindowTinyLfu declares WEIGHTED and budgets its regions by entry weight, mirroring BoundedLocalCache's candidate/victim eviction loop; on unit-weight traces it is bit-for-bit the classic single-step form (verified on the bundled lirs cells). Weighted terrain/ceiling scans sweep percent-main exactly like unweighted ones — but read the right metric: the panel reports the weighted (byte) hit rate for weighted cells, while the climber's own goal metric is the per-request object rate (its sample counters are unweighted), so judge climber terrain on the Hit Rate column. A weighted ceiling taken before 2026-08-07 is void. The policy's admission filter was sized once, when the cache first filled, and never re-sized; the cache retracks it on every addition. The sketch's reset period scales with its capacity, so a frozen one ages at near-zero period and flatters small windows, where nearly every admission is filtered. On metaCDN_rprn@4G that put the reported static-window optimum at a 1% window when it is at 80%. Nothing recorded was affected (no weighted ceiling had been taken); the filter now retracks per addition and unweighted output is unchanged bit-for-bit.

• run.py <trace> --size N [--anchors] [--runs N] [--variants ...] [--dump PREFIX] runs one cell (LRU + static-window anchors, N repeats, mean ± spread), and gate.py <traces-dir> [variants] [out.csv] runs this table's whole battery at the table's per-cell N — the runnable form of the gate; update the two together. gate.py rotates its arms inside each run rather than running them in blocks, which keeps the separate processes temporally adjacent. It does not make them share JVM state or admission draws (they ran in blocks until 2026-08-02, which made slow environmental drift harder to distinguish).

• Characterizing a holdout: pass an EMPTY --variants. run.py --anchors and probe.py both used to loop a variant list that was never allowed to be empty, so every anchor run also fired one product.Caffeine pass. §6 selects a holdout by LRU-only characterization and a cell is spent once the product has touched it, so the tool silently burned the cell it was characterizing — measured 2026-08-02, cost arc/P1@65536. Both now accept --variants '' and skip the product entirely; verify with the printed output (an anchors-only run prints the LRU=… ceiling=… line and nothing else).

• -Dcaffeine.climber.seed=<n> (harness) makes each arm deterministic. BoundedLocalCache.admit's 1-in-128 HashDoS draw is the simulator's only live nondeterminism (the policy runs .executor(Runnable::run)), so seeding it makes an arm reproduce bit for bit and a lottery basin can be compared seed by seed instead of by unseeded draw counting. Equal seeds initialize equal streams, but different contest schedules can attach the kth draw to different requests. This is paired-by-seed evidence, not request-indexed common-random-number pairing, unless both the draw count and request-index digest match. Added 2026-08-03, where it settled two contested rows immediately: a −5.77 on phases_d050 was pure basin draw (arms bit-identical on all 8 seeds) while a −2.12 on crashnoise_a12 was real (−1.43 on 7 of 8). Prefer it for any single-cell adjudication. It does NOT replace the unseeded N≥8 policy for claims about the distribution of basins, and a seeded sweep is only as broad as its seed list — use 8 seeds, not one.

• harness.py wires in the adversarial-round instrumentation — harness.py apply <worktree> (also verify and strip), distilled from rounds 2–4 so each round stops rebuilding it. NEVER commit the result: it prints to stderr and carries variant knobs, and ErrorProne flags the SystemOut by design. It adds -Dcaffeine.climber.debug=true (one mode-labeled stderr line per density sample — the trajectory instrument that found every behavioral defect to date), -Dcaffeine.climber.seed=<n> (a seeded admission tiebreak), -Dcaffeine.climber.auditbar=<f> (the audit crash bar's fraction of the rate frozen at arm; 0 restores the pre-fix absolute 5pp), and -Dcaffeine.climber.variant= with ablation arms: noaudit (audit layer off), reactive (hit-rate tier at every size), nocal (no cold-start calibration), hardreset / starvwrite / prefix (restore the pre-round-3 clock reset, the pre-round-3 starvation schedule write, or both), absprobe (the pre-adv3 absolute walk-interior bar for starvation probes), precrash (the pre-crash-semantics machine: shared crash ledger, escalated crashes double the audit clock, one-sample audit aborts), noretest (a veto's return never re-tests the claim that sent it, the pre-2026-08-19 arrival), flatroom (the pre-2026-08-03 room rule, measuring the flat restart magnitude while the walk strode up to 4× it), and staleclaim (the pre-2026-08-03 stand-down, which left the goal metric smoothing across a regime shift) — the A/B arms for landed fixes without checking out history. One arm is a candidate rather than history: parkbound (end an audit's park with its shield instead of holding to the next audit — battery mean +0.08 and real-corpus neutral). It was recorded as the live lead on the R4-F1 give-back, worth "+1.3..2.0 uniformly on all three shieldtrap seeds". That does not reproduce under seeded paired runs and the lead is withdrawn (2026-08-04, N=8 admission seeds 1–8, arms rotated inside each run): s7 78.06 → 76.92, s11 77.19 → 77.60, s13 77.44 → 77.83 — mean −0.11, not +1.3..2.0, and it closes none of the 1.37pp the reactive arm leads this family by (78.93). What it actually does is collapse the basin structure to a fixed ~77.6: on s11 its spread is 0.13 against hybrid's 2.94, so it gains where hybrid draws low and gives back where hybrid draws high. On s7 that trade is bad — two of eight seeds fall to 74.37/74.68 against hybrid's 79.01/79.00, a −4.6 tail the unseeded mean hid. saw_p40 (the family's other sentinel) is +0.49, which does not pay for it. Re-open only with a mechanism for the s7 tail. The rearm* family (rearm, rearmstill, rearmcold, rearmreset, rearmboth, rearmheld, rearmheld2, rearmheldcold; one flag table in harness.py) is the stand-down re-arm of item 4's latency face, priced dead in every form 2026-08-22: a discarding stand-down restarts the audit schedule, gated on held parks or not, with the stillness run restarted or kept. Kept so the square can be re-priced after an alternation change; hill-climber.md §5 has the numbers.

• It is snippets, not a diff, and that is load-bearing. It was experiment-harness.patch until 2026-08-04, when it broke three times in one day: a unified diff encodes the source it was cut from, carrying javadoc as context, so an unrelated comment reword breaks it — and both failure modes are silent. git apply is all-or-nothing per file, so one stale hunk refuses every hunk for that file and yields a build with no instrumentation that still prints plausible numbers (an ablation arm then measures the shipped machine). A fuzzy patch lands the rest and reverts whatever the stale context encodes — twice reverting a landed fix inside densityClimb. Both harnesses assert their edit tables are safe before touching a file: presence of the applied form is the idempotency test, the verify test and the strip target, so two edits whose applied forms are equal — or where one contains the other — make the second invisible. Apply skips it as already wired, verify passes, and the instrumentation is silently partial while still printing plausible numbers. That is not hypothetical: the admission harness hit it on its first run (two edits inserting the same counter line at different anchors) and lost 276,764 of 558,186 hits. harness.py anchors on code only, so doc edits cannot move it; a moved anchor names itself and stops, applying is idempotent, and strip round-trips to pristine. If a variant arm reads suspiciously like the baseline, run harness.py verify before believing it is a finding. It earned this on its first use: verify reported 5 of 27 not wired in a study that had applied the old patch with --fuzz, and two of the five were the admission-seeding pair — so every "seeded, bit-reproducible" adjudication in that study would have been fiction. The study reset both files, re-applied, and then proved seeding empirically (one cell, one seed, run twice, bit-identical) rather than reasoning past the warning. Do that check whenever verify is not clean. Point run.py/gate.py at the worktree via CAF_TREE. A strategy extraction can invalidate the harness anchors. State exactly which variants a replacement adapter supports: verified hybrid/reactive/noaudit runs do not validate unported historical knobs. Check native behavior and reached variant branches as well as source splices.

• Check the core's Java target as well as the simulator's runtime. Compile production-like overlays with the repository's core release target (currently Java 11), and verify class-file versions and actual loaded class origins. A newer simulator JVM does not make newer source syntax acceptable in the library. The later-retention prototype needed two dispatch checks rewritten as Java 11 casts; its density methods were instruction-identical, then all 717 candidate seed cases were replayed to establish exact behavior. Preserve old artifact identities and distinguish equivalence-supported reuse from direct runs of the corrected build. Rebind runtime benchmarks to that final build too.

• A probe you add yourself must print climb -prefixed lines, and must assert its own denominator. run.py keeps only stderr lines starting with "climb ", so an ad-hoc System.err.println("MYCOUNTER ...") is discarded before it reaches the --dump file and the count reads zero — from a probe that compiled in and ran on every sample. Measured 2026-08-09, where it turned 3 phantom anchor plants into a "no measured exposure" conclusion that was about to be written up. The line must also carry win= and mode=, since summarize parses both out of every dumped line. The check that catches it is free: pair the new counter against a quantity the harness already counts (drain evaluations must equal mode=undo samples) and refuse the numerator until the denominator matches. Same family as the verify incident above — a silent probe and a wired probe produce identical-looking output.

• exposure.py (harness-dependent) answers "can this trace ever audit?": stillness fraction, longest run vs the wait, audits/probes armed — the instrument behind the corda blind-hold and w097 at-the-bar findings. blindlock.py (harness-dependent) answers the neighbouring question, "how much exploration is the blind-corner branch refusing?" — it reads a --dump trajectory and counts blind-corner samples (labelled hold/ARM*) whose stillness run has already reached auditWait. It is what found the 2026-08-04 lockout: metronome refused 116 due re-tests, its longest unbroken run being 65, and four other cells 29 apiece. Its columns changed meaning when holdOrAudit shipped — a due clock now pre-empts the refractory hold, so the count is starvation probes outranking an ordered audit and no longer includes holds; read its docstring before comparing against a pre-2026-08-06 number.

• census.py and corner.py (harness-dependent, 2026-08-15) read trajectory dumps. census.py run dumps one trajectory per (cell, seed) for gate.py's cells plus real.py's, and census.py parse emits a row per audit confirm: direction, base and confirmed window, whether the confirm sample was blind or the grown region starved, the park's length and how many of its samples starve, how it ended, and the hit rate over the park against the samples before the arm. It is what showed a park on a starved position is 3 in 642 confirms, that 204 of 254 audits crashing out of a park release it, that 34 of 92 veto returns arrive with the rate up and release the hold, and how often the 5pp shift trigger discards the anchor per cell. corner.py reports the corner starvation probes over the same dumps (rung at arm, ending, depth, the ending sample's cost): 302 upper-corner probes ending 163 crash / 115 fail / 11 confirm against the floor's 1,098 ending 785 confirm. Re-run both before proposing a rule about parks, corners or the stand-down.

• The tier gate moves the law and the cadence together, and the 2×2 separates them. Above 4096 the climber switches steering law and sample period (10× maximum → 4×), so neither can be attributed alone. harness.py carries all four corners: reactive / density force a tier at every size, and react4x / dens10x are the cross arms (hit-rate law on the short cadence, and density law on the long one). The 2026-08-04 D2 study needed this — the dominant term flips with terrain, so a single isDense switch attributes a tier cliff to the wrong half.

• byworkload.py counts workloads, not cells; pair.py compares seeds, not spreads. Both encode a mistake that is easy to make with these corpora. Epoch slices of one trace are not independent evidence — six agreeing epochs are replication within one workload, and pooling them inflates the base of a reachability claim (D2's "8 movers" were 3 workloads until it was widened to 9). And a seeded sweep is not a repeated measurement of one basin: each seed defines an independently reproducible run, so arms are compared seed by seed and an arm's own column spread says nothing about whether a difference is real (phases_d050 spreads 12.9pp across eight seeds with the code held fixed). This pairing controls seed identity only; it does not align draws to requests without the parity proof above. sweep.py is the canonical seeded long-format producer: it interleaves arms within each seed and aborts a cell rather than writing a partial arm vector. Its output carries parallel seeds and runs fields; pair.py rejects legacy positional rows, duplicate seeds, incomplete arms, and unequal seed sets rather than truncating or shifting the comparison. byworkload.py enforces the same identity contract when comparing the 4096/4097 tier pair.

• lcs2lirs.py projects a libCacheSim oracleGeneral trace to unit-weight LIRS epochs. The tier boundary is 4096 in the maximum's native units, so on a byte-weighted trace it means 4096 bytes; projecting to unit weight is what makes such a trace usable at the density tier at all.

• run.py --seeds 1,2,..,8 runs one pass per admission seed instead of --runs repeats, so each arm is reproducible and a contested cell is compared by seed rather than by unseeded draw counting. Variants run adjacently within each seed, and the output prints the exact seeds=s1:hit-rate,... vector for every arm. Before 2026-08-04 no committed tool accepted -Dcaffeine.climber.seed, and three workspaces had each hand-rolled their own; use this. It is what showed a −6.32 slowswap_step, a −2.60 phases_d050 and a −0.59 bandtrap2 were all basin draws (bit-identical on all eight seeds).

• real.py sweeps the real corpus (ten cloud-physics cells plus three ARC) and floors.py the thin-signal floor cells (w50@123038, S1/S2/S3, DS1@1051635) — the runnable form of the gate's thin-signal floors row, and the cost side of anything that touches the refractory or the probe machine. Both skip the frozen merge-gate holdout and arc/P3. The ARC cells declare the arc reader: they were lirs until 2026-08-04, which yields nothing silently, so every ARC row written before then was blank and the "real corpus 10/10 ties" bar on record is the ten cloud-physics cells alone.

• kvcache.py runs the LLM prefix-cache rows: four recency-favorable cells and two frequency-favorable controls from the Qwen Bailian traces (~/Documents/traces/kvcache/), read by bailian, seeded 1–8 by default. Point CAF_TREE at the harness worktree, or the seeds are ignored and each run is a fresh draw. Its longer rows repeat the same keys, as repeat.py does; this adds decisions but also preserves frequency-sketch history. An unchanged LRU rate does not establish a neutral extension. Keep those recorded rows and bars intact. For new warm-up claims, compare with verified fresh-key namespaces per pass, preserving sharing within each pass, and report first-pass and later-pass scores separately. --anchors takes LRU, ARC and static windows over the same replay; any new key transformation must be applied to every comparison arm.

• Keep prefix utility separate from the legacy block metric. With verified request boundaries, report usable leading-prefix hits alongside independent-block hits on the same paired replay; state the denominator and weighting. A later block hit after the first miss may be unusable. Do not assume LRU's equality on one replay is universal, stop the access stream at the first miss, or replace the existing block-hit bars. Keep boundary markers outside policy/sketch/statistics updates. Supplemental scoring needs its own verified instrumentation and paired references; it is not supplied merely by running kvcache.py. See methodology.

• marginal.py answers "where does a steering signal rest, versus where does the hit rate peak?" — signal anatomy with no controller involved. Wire a worktree with harness.py apply (NEVER commit it: it adds region-attributed hit counters and an LRU-tail band to the static-window anchor, maintained by an O(1) midpoint marker), point CAF_TREE at it, and sweep. The tail band catches only a small share of window hits (~0.4% on attack_whisper@8192), so on a small or short cell it rounds to zero and the marginal error goes epsilon-driven: cs@512 reads ht=0/hq=0 at every window while attack_whisper@8192 reads ht=2009/hq=2.18M. Confirm ht/hq are non-zero on the cell you are measuring before trusting a marginal number from it. It reports, per cell, the peak window and the zero-crossing of both the average-value error ln((H_w/C_w)/(H_m/C_m)) — what the density arm steers on — and the marginal-value error ln((H_tail/C_tail)/(H_prob/C_prob)). This is the instrument that measured the density arm's give-back to its cause in 2026-08-02 (average rest point above the peak on 7/7 crossing cells, losing 0.71pp mean vs the marginal form's 0.09pp); TAIL=<delta> sweeps the band fraction to check an answer is not an artifact of it. Its window list stops at 0.80 because that is the window's reachable ceiling (increaseWindow donates only the protected allocation), which is the right range for a rest-point question; set WINDOWS=0.9,0.95,... when the question is instead where the hit rate peaks, since six of twenty-one cells measured 2026-08-08 peak above the ceiling and a sweep stopping at 0.80 reports the wrong peak for them.

• admission_harness.py + admission.py answer "how often is an admission right, and how much does it earn when it is?" — the measure taken from Merlin's Figure 20, which prices admission decisions directly instead of inferring their quality from the end hit rate. Same worktree convention (admission_harness.py apply <worktree>, never commit), then CAF_TREE=<wt> admission.py <out.csv> <manifest> sweeps static windows and reports, per region, precision (share of admissions ever hit) and yield (mean hits when hit). It exists because density = precision x yield x (entries per capacity) and the steering signal sees only the product: at P3@152508's rest point the two regions look comparable by density (15.4 vs 12.3) while the window is right 1.1% of the time for 1.03 hits against main's 59.2% for 10.78 — a 1% window's turnover carries a region almost nothing is hit in. Every row is conservation-checked (window hits + main hits must equal the reported hit count) and a row that fails is refused, which is not decoration: the harness's first version dropped every protected-region hit and still printed plausible precisions. Diagnostic only — do not re-derive equalization. Precision- and yield-equalization were both checked against the peak on four cells: precision ratio spans 0.01–2.55 and yield ratio 0.03–0.26, against density's 0.58–0.95.

• -Dcaffeine.climber.startwin=<frac> (harness) plants the initial window, and it is the only instrument here that asks whether the climber DESCENDS. Every other row starts the cache where the product starts it, at 1%. On a frequency-favorable trace that is already the answer, so the row passes without the machine moving, and a broken descent would read clean on the whole real corpus. startwin.py sweeps the plant over 5/10/20/40/80% and reports, per cell, the frozen static window at the plant (the do-nothing floor), the same cell from the shipped start (the recovery target) and recov, the fraction of the gap between them the run wins back. Four things about it are load-bearing:

  • The plant holds probation at its shipped 19.8% instead of re-splitting main 80/20, which is the geometry increaseWindow itself produces (it conserves window and protected). That is what puts an 80% plant at the window's reachable ceiling rather than above it. --split resize gives the other geometry, which a setMaximum resize leaves but the climber cannot reach.
  • It also means run.py's static sweep is not a like-for-like frozen reference at a large plant: it varies percent-main alone, so its 80% point has 16% protected where the plant has 0.2%. frozen_matched.py re-measures the frozen point with the plant's own split.
  • Recovery is a function of density samples, not of cache size. A sample is 4 × maximum requests and a full descent from 80% needs 15–20 of them, so the plant's cost in hit rate scales with the maximum while the descent does not. Read the sample count before reading a deficit: on ds1_4M an 80% plant ends at an 80.1% window because two samples is two commands, which prices arithmetic and not the machine.
  • ds1_256k is not a member of this family, and neither is any other DS1 cell below ~1M: DS1's static optimum moves with size, a 30% window at 262144 against 1–2% at 1M, so a plant there is not a descent test at all.

  Four instruments come with it. descent.py reports the median falling step per sample and the undone column; repeat.py replays a trace N times, which is what separates "too short for this maximum" from "cannot descend"; screen.py picks further workbench cells from anchors alone, so it spends nothing; and mechanism.py recomputes |error| × 0.03 × maximum per sample and prints it beside the motion observed, which is how "the descent is gain-limited" was established rather than assumed.

• Noise: unseeded admission gives ±0.1–0.8pp run to run. Gate verdicts are coarse (an escape vs a ~28pp pin), so 1–2 runs suffice; take N≥5 before claiming anything under 2.5pp. Keep candidate-vs-reference arms in the same sweep (interleaved runs, append+resume CSVs) rather than comparing across sweeps, and compare seeded arms seed by seed. Interleaving limits drift; equal seeds are not request-indexed common random numbers without draw-count and digest parity.

• A paired decision gain is not an adoption gate. Attribution and later-retention diagnostics use the decision-replay protocol. They require equivalent baseline prefixes, independently counted misses at predeclared horizons, and a discriminator that separates helpful from harmful actions using decision-time evidence. The later prefixes must not inherit an earlier alternative park. A surviving rule then runs end-to-end on its own evolving states; do not estimate its gain by summing the diagnostic event table. Preserve this battery, real corpus, floors, repeated and planted controls, and fresh holdout requirements. Retention work must price the existing park/release and return/schedule controls (phases, widepin, demoflood, ghostclaim, the moat rows and planted mainsat), not only the witness it keeps. Any added history or lifecycle observation also owes retained-state, allocation and hot-hit/mixed/adaptation-heavy throughput measurements with logging disabled. A diagnostic alone has not passed or failed the unrun adoption gates. Price whole continuations: even a single later park can gain through 256C and lose over a longer registered trace. Define a practical runtime loss budget before classifying cost; baseline-noise flags and isolated sample timings do not supply that budget or establish whole-cache acceptability.

The gate (§3 families — the must-always-re-run set)

Generate (seed 7 is canonical; add --seed 11/13 when a verdict is contested):

S=<scratch>; G=.claude/skills/climber-gate
python3 $G/gen.py mixture   --max 8192  --dfrac 0.10 --out $S/mixture_8192_d010.lirs
python3 $G/gen.py mixture   --max 8192  --dfrac 0.25 --out $S/mixture_8192_d025.lirs
python3 $G/gen.py mixture   --max 32768 --dfrac 0.25 --out $S/mixture_32768_d025.lirs
python3 $G/gen.py mixture   --max 8192  --dfrac 0.50 --out $S/mixture_8192_d050.lirs
python3 $G/gen.py mixture   --max 8192  --dfrac 0.50 --lengthmult 576 --out $S/mixture_8192_d050_long.lirs
python3 $G/gen.py mixture   --max 8192  --dfrac 0.10 --lengthmult 640 --out $S/mixture_8192_d010_long.lirs
python3 $G/gen.py mixture   --max 8192  --dfrac 0.25 --lengthmult 640 --out $S/mixture_8192_d025_long.lirs
python3 $G/gen.py mixture   --max 32768 --dfrac 0.25 --lengthmult 640 --out $S/mixture_32768_d025_long.lirs
python3 $G/gen.py phases    --max 8192  --dfrac 0.50 --out $S/phases_8192_d050.lirs
python3 $G/gen.py phases    --max 32768 --dfrac 0.50 --out $S/phases_32768_d050.lirs
python3 $G/gen.py deadphase --max 8192  --out $S/deadphase_8192.lirs
python3 $G/gen.py widepin   --max 8192  --out $S/widepin_8192.lirs
python3 $G/gen2.py straywall2 --max 8192  --dfrac 0.50 --out $S/straywall2_8192_d050.lirs
python3 $G/gen2.py straywall2 --max 16384 --dfrac 0.50 --out $S/straywall2_16384_d050.lirs
python3 $G/demoflood.py --max 8192 --out $S/demoflood_8192.lirs
python3 $G/gen_attacks.py $S          # trickle s7/s11/s13 + lowmix gate rows
python3 $G/blind_gen.py $S            # blind_balloonflip + blind_nullchurn gate rows (seed 41)
python3 $G/blind_gen.py $S v2         # blind_bandtrap2 gate row
(cd $S && python3 $G/whisper.py attack && python3 $G/whisper.py quarter)   # whisper gate rows
# F1-era rows (the goal-metric layer's own dose instruments and sentinels):
python3 $G/whisper_mod.py --amp 0.08 --period 12 --out $S/whisper_a0.08_p12.lirs
python3 $G/whisper_mod.py --amp 0.08 --period 6  --out $S/whisper_a0.08_p6.lirs
python3 $G/whisper_mod.py --amp 0.12 --period 12 --out $S/whisper_a0.12_p12.lirs
python3 $G/gen_adv.py mixmod --max 32768 --dfrac 0.25 --lengthmult 400 --amp 0.00 --out $S/mixmod_32768_a000.lirs
python3 $G/gen_adv.py mixmod --max 32768 --dfrac 0.25 --lengthmult 400 --amp 0.10 --period 6 --out $S/mixmod_32768_a010.lirs
python3 $G/gen_adv.py metronome --max 8192 --ratio 1 --out $S/metronome_8192_r1.lirs
python3 $G/gen_adv.py regimeramp --max 8192 --out $S/regimeramp_8192.lirs
python3 $G/escgen.py jam --max 8192 --out $S/esc_jam.lirs
python3 $G/a1_tenant.py --max 8192 --share 0.10 --dist 400 --out $S/tenant_s10.lirs
# Round-3 rows (the position-jam family; whisper base, dose-matched modulation):
python3 $G/gen_jam.py --mode flat --out $S/posjam_flat_8192.lirs
python3 $G/gen_jam.py --mode jam  --jitter 0.5 --out $S/jam_j50.lirs
python3 $G/gen_jam.py --mode jam  --period-req 81920 --out $S/jam_d25.lirs
python3 $G/gen_jam.py --mode jam  --out $S/jam_d0.lirs
# Round-2 instruments, promoted to rows by round 4 (the audit-amplification watch):
python3 $G/gen_adv.py shieldtrap --max 8192 --out $S/shieldtrap_8192.lirs
python3 $G/gen_adv.py shieldtrap --max 8192 --seed 11 --out $S/shieldtrap_s11.lirs
python3 $G/gen_adv.py shieldtrap --max 8192 --seed 13 --out $S/shieldtrap_s13.lirs
python3 $G/gen_adv.py loopcliff  --max 8192 --out $S/loopcliff_8192.lirs
python3 $G/gen_adv.py climbtrend --max 8192 --out $S/climbtrend_8192.lirs
python3 $G/gen_adv.py climbtrend --max 8192 --direction down --out $S/climbtrend_dn_8192.lirs
python3 $G/gen_adv.py climbtrend --max 8192 --hot-lo 0.55 --hot-hi 0.55 --modamp 0.12 --modperiod 40 --out $S/saw_p40.lirs
# adv3 rows (the walk-interior exit bar; whisper/mixture bases, amplitude-dosed):
python3 $G/crashnoise.py --shape flat --amp 0.00 --out $S/cn_flat.lirs
python3 $G/crashnoise.py --shape sine --amp 0.12 --period 12 --out $S/cn_sine_a12.lirs
python3 $G/mixnoise.py --max 8192 --shape flat --out $S/mn8_flat.lirs
python3 $G/mixnoise.py --max 8192 --shape sine --amp 0.10 --out $S/mn8_sine_a10.lirs
# crash-semantics rows (Terra H4-C1 pulse train + adv4 moat terrain; audit crash pricing):
python3 $G/gen_h4c1.py --phase 4 --order high-low --samples 260 --out $S/h4c1_p4_attack_long.lirs
python3 $G/gen_h4c1.py --phase 4 --order low-high --samples 260 --out $S/h4c1_p4_control_long.lirs
python3 $G/gen_moat.py --hotn 7800 --bandd 4000 --n 12000000 --out $S/moat_h7800_long.lirs
python3 $G/gen_moat.py --hotn 3000 --bandd 4000 --out $S/moat_h3000_b4000.lirs
python3 $G/gen_moat.py --hotn 4000 --bandd 4000 --out $S/moat_h4000_b4000.lirs
python3 $G/gen_moat.py --hotn 5000 --bandd 4000 --out $S/moat_h5000_b4000.lirs
# 2026-08-03 rows (the audit's room rule and the stale away-anchor claim):
python3 $G/slowswap.py --ramp 1  --p3 139 --out $S/slowswap_r1.lirs
python3 $G/slowswap.py --ramp 20 --p3 120 --out $S/slowswap_r20.lirs
# 2026-08-04 rows (the blind-corner lockout; one terrain, the rider flips only hasBlindCorner):
python3 $G/gen_blindlock.py --max 8192 --whisper 0     --out $S/blindlock_blind.lirs
python3 $G/gen_blindlock.py --max 8192 --whisper 0.004 --out $S/blindlock_sighted.lirs
# 2026-08-04 rows (the D2 tier straddle; run the PAIR at 4096 and 4097):
python3 $G/gen_straddle.py --max 4096 --period 8 --out $S/strad_p8.lirs
python3 $G/gen_straddle.py --max 4096 --period 8 --lo 0.325 --hi 0.325 --out $S/strad_stat.lirs
# 2026-08-15 sentinels (from the LIRS family study; regenerate its traces bit-identically):
python3 $G/gen_norank.py repeat --refs 6 --width 4096 --out $S/rep_r6_w4096.lirs
python3 $G/gen_norank.py flood --jrate 1.0 --prologue --out $S/flood_j100.lirs
# 2026-08-15 /audit-regret round 1 (the spec lives with that skill; the witness verbatim, dreq in requests):
A=.claude/skills/audit-regret
python3 $A/workload.py $A/specs/shallowmoat.json --out $S/shallowmoat_8192.lirs
python3 $A/workload.py $A/specs/scarburst.json --out $S/scarburst_8192.lirs      # reactive-anchor sentinel
python3 $A/workload.py $A/specs/flashpark.json --out $S/flashpark_8192.lirs      # shieldtrap dose note; no row yet
# 2026-08-16 /audit-regret round 2 (the pulsed lure; the period-16 witness at 128 samples, the period-8 form at 256):
python3 $A/workload.py $A/specs/absolve.json --out $S/absolve_8192.lirs
python3 $A/workload.py $A/specs/absolve_p8.json --out $S/absolve_p8_8192.lirs
# 2026-08-17 /audit-regret round 3 (the stale-claim family's away-anchor witness; two phases, 128 samples):
python3 $A/workload.py $A/specs/ghostclaim.json --out $S/ghostclaim_8192.lirs
# 2026-08-17 /audit-regret round 4 (the audit verdict's position on the walk, 132 samples; and round 3's
# hazefloor note rowed at 320 samples as the top-corner retention cycle's sentinel):
python3 $A/workload.py $A/specs/crestpast.json --out $S/crestpast_8192.lirs
python3 $A/workload.py $A/specs/hazefloor.json --out $S/hazefloor_8192.lirs
# 2026-08-21 /audit-regret round 6 (the phase-chase family: the 240-sample base at seeds 1-8 and
# its deterministic 60-sample first-cycle witness):
python3 $A/workload.py $A/specs/parkveil.json --out $S/parkveil_8192.lirs
python3 $A/workload.py $A/specs/parkveil_min.json --out $S/parkveil_min_8192.lirs
# 2026-08-21 /audit-regret round 7 (the audit clock's backoff against a late prize: the 262-sample base
# and its 187-sample never-found witness; and the main-side masked-signal sentinel at 72 samples):
python3 $A/workload.py $A/specs/latebloom.json --out $S/latebloom_8192.lirs
python3 $A/workload.py $A/specs/latebloom_min.json --out $S/latebloom_min_8192.lirs
python3 $A/workload.py $A/specs/mainsat.json --out $S/mainsat_8192.lirs
# 2026-08-22 /audit-regret round 8 (the opening side choice, deterministic; the edge-landing confirm, N=8):
python3 $A/workload.py $A/specs/sidecliff.json --out $S/sidecliff_8192.lirs
python3 $A/workload.py $A/specs/jumpslide.json --out $S/jumpslide_8192.lirs
python3 $A/workload.py $A/specs/lowbar.json --out $S/lowbar_8192.lirs          # the moat family at a fraction-bar dose, N=8
# 2026-09-04 /audit-regret (the return retest's live margin, 72 samples; seeded latency sentinels):
python3 $A/workload.py $A/specs/ramp_retest.json --out $S/ramp_retest_8192.lirs
python3 $A/workload.py $A/specs/ramp_retest.json --max 16384 --out $S/ramp_retest_16384.lirs
# 2026-09-24 /audit-regret (a dead sample releases a shielded audit park; one 2-sample burst at s36):
python3 $A/workload.py $A/specs/deadpark.json --out $S/deadpark_8192.lirs
# 2026-09-17 /audit-regret round 10 (a narrow static peak, 288 samples; sweep the anchors densely over 41-60%):
python3 $A/workload.py $A/specs/needle.json --out $S/needle_8192.lirs
python3 $A/workload.py $A/specs/needle_wide.json --out $S/needle_wide_8192.lirs

Standing rule: every adversarial round's attack instruments become gate rows in the same change that lands its fix — an instrument without a row is not a regression test (round 4's finding: the round-2 generators had no rows, so the round-3 fix was never measured against them and shipped a −1..−2.1 give-back unmeasured).

Two kinds of rows: a pass bar must hold or the change is rejected; a sentinel is a known-fail frontier cell whose bar is "no NEW drift beyond the recorded value ±noise". Recorded values are the 2026-07-31 baseline at the full adversary-round stack — F1 streak confirm + fresh-park shield, F3 calibration audit, the parkFreshLeft latch fix, round 2's auditBeatBase gate, and round 3's audit-clock decay + audit-owned schedule, plus round 5's audit-confirm zero-step and crash-cadence clock (2026-08-01; per-cell CSVs in the local climber-adv2 workspace, the round-5 sweep as *_confirmzero.csv; rows re-based by rounds 3/5 say so inline), plus the 2026-08-02 crash-semantics stack — per-layer crash ledgers, crash-never-doubles-the-audit-clock, escalated time persistence with the tolerated walk holding direction, and the bridge writes (local climber-crash-semantics workspace, full-battery CSV gate_esc.csv; rows re-based by it say so inline, and it added the h4c1 and moat rows). The adv3 round's records (the walk-interior bar family, its candidate table, and the whisper_mod a0.12 re-base) are in the local climber-adv3 workspace, with the same-day pricing study's screens, battery (gate_devprobecap.csv), widepin draws, and spent-holdout scoring beside them; the rows re-based by the starvation-probe pricing (mixnoise a10, metronome's trajectory signature) say so inline. The round-2-era per-cell CSV is the local climber-adversary workspace's gate_final.csv, with gate_f3b.csv / gate_f3.csv / gate_f1.csv / gate_f4.csv / gate_auditroom.csv beside it as the earlier records; the prior shipped-design record is the climber-failure-modes workspace's gate_v3dev_port.csv. All values are on these generators' instances; average-verdict ablation values are kept in parentheses where the two verdict forms differ. Bimodal families (widepin, phases, straywall2@8192's rare low draw, crashnoise_a12 — added 2026-08-02, previously recorded as deterministic on N=2 — and blind_balloonflip since the half-full gate) adjudicate on an N=8 mean — three separate rounds have now caught a single-run or N=3 "improvement" or "regression" that was basin luck; when a bar fails on a code change, run the paired reference binary before believing it. Seeds vary admission draws, not a trace's timing: on a trap whose regime changes at a fixed sample, a change that moves the audit schedule meets that change at the same phase on every seed, so an eight-seed loss can still be one timing draw (the slowswap row measures how large).

How a bar must be stated (2026-08-05 audit of all 52 rows, which found defects in 37). Four distinct defects had each been found by accident rather than by looking, so every row was re-measured seeded at N=8 against the LRU/ceiling anchors. The rules that came out of it:

1. A bar must be able to fail for the reason the row exists. A level under LRU cannot — trickle_s7's ≥66 sat 7.8pp below the do-nothing baseline, blindlock_blind's ≥48 by 11.0, phases_d050's ≥50 by 22.2. Where the machine clears LRU, state the bar as LRU ("above LRU on every seed") and record the measured margin.
2. Where the machine is below LRU, an above-LRU bar is unreachable, so bar the row against itself: "no drift below mean − 2sd", with the LRU gap stated in the same breath because that gap is the thing the row tracks. Do not leave a level that permits pp of silent decay (the old ≥48 on mixture_d050 allowed 2.7 on top of an 8.3pp deficit).
3. A bar inside the row's own seeded spread is a coin flip. moat_h7800's ≥45 sat inside draws of 44.32–46.23 and the row fails it on 3 of 8 seeds today; crashnoise_a12's ≥61.2 sat inside 60.75–63.63. Bimodal rows take a bar on the N=8 mean, never on a draw.
4. Levels rot; margins do not. Five recorded values no longer reproduced (balloonflip 79.8 → 78.04, and its recorded 79.5–83.0 range no longer contains the value at all; mixture_d010 61.9 → 63.00; regimeramp 53.5 → 54.03; trickle_s11 73.0 → 72.46; mixture_d025@32k 58.0 → 58.48). Prefer "above LRU", "within noise of the ceiling", or "no drift below X" over a bare number.
5. Check the row has enough decisions to bar at all — see the trace-length table below.

Family	Cell	Recorded	Bar
corda + 5×loop + corda (bundled real)	512..8192 incl. 4097	near static ceiling	pass: no tier-boundary cliffs
mixture d010	8192	63.00 ± 0.02 at N=8 (2026-08-05 re-measure; the recorded 61.9 had rotted 1.1 low; LRU 59.83)	pass: above LRU on every seed (LRU 59.83; N=8 min 62.98, margin +3.15) — escaped
mixture d010 --lengthmult 640	8192	63.58 ± 0.01 at N=8 (2026-08-05, 159 density samples against the parent's 47; LRU 59.85, ceiling 65.98 @10%, reactive 61.77)	pass: above LRU on every seed (margin +3.71 at the N=8 min). Steady-state companion to the 47-sample parent, which resolves only warmup. Trajectory oscillates between the optimum and a high excursion (window by sixth 25/15/44/15/43/44% against a 10% optimum) at a flat 63–65% hit rate, so the level error is real but costs little in-run
mixture d025	8192	59.67 ± 0.10 at N=6 (LRU 59.58 on the instance; was 60.5 pre-round-3)	pass: N≥6 MEAN above LRU — the margin sits at the noise floor (+0.09), so a single below-LRU run is noise, only a below-LRU mean is real, and a fall toward the ~35–40 floor pin is the unmistakable regression. The ~0.9 give-back is the calibration audit newly reaching the family (a starvation confirm no longer cancels it) and spending two failed-audit excursions on a 47-sample trace; don't "fix" it by re-suppressing the audit. 2026-08-05: this row prices warmup — its final third reads 62.78 against LRU 59.58, so the noise-floor margin is a convergence artifact; the --lengthmult 640 companion above is the quality number
mixture d025 --lengthmult 640	8192	61.69 ± 0.01 at N=8 (2026-08-05, 159 samples; LRU 59.60, ceiling 65.85 @20%, reactive 61.24)	pass: above LRU on every seed (margin +2.07 at the N=8 min). This row is why the parent's +0.09 margin is not a result: the parent's whole-trace 59.67 is its own first third (59.83), and steady state is 62.5, i.e. +2.9 over LRU. Oscillates in a 51–68% band against a 20% optimum
mixture d050	8192	50.8 (2026-08-04 re-base of 34.7 by the blind-corner fix; seeded N=8 34.71 → 50.78, per-seed +16.01..+16.14; LRU 59.10, ceiling 64.17 @40%)	sentinel: no drift below 50.4 (N=8 50.78 ± 0.02); 8.32pp below LRU 59.10, so an above-LRU bar is unreachable and that gap is what the row tracks. The old ≥48 allowed a further 2.7pp of silent decay — the dip-length trace used to end before the ladder's ~round-3 escape, and the audit now escapes it instead. The 8.3pp below LRU is warmup, not a pin (2026-08-05): the final third reads 62.85 against LRU 59.10, i.e. +3.75 above, and the --lengthmult 576 companion's steady state is 62–63. Read this row as convergence cost; a fall toward 34.7 means the blind-corner ordering regressed
mixture d050 --lengthmult 576	8192	58.2 (2026-08-04 re-base of 52.9 by the same fix, +5.00; LRU 59.1)	sentinel: no drift below 57.8 (N=8 58.18 ± 0.01); 0.95pp below LRU 59.13. The old ≥56 allowed 2.2pp of silent decay — the ladder-escalation escape, now completed by the audit rather than by the ladder alone
mixture d025	32768	58.0 (parent reactive 56.4; LRU 59.9)	sentinel: no drift below 58.1 (N=8 58.48 ± 0.01, re-based from the recorded 58.0); 1.37pp below LRU 59.85. The old ≥55 allowed 3.5pp of silent decay — the F4-class flagship, closed: the room rule un-walled the first audit (31.2 → 39.3) and the cold-start calibration audit removed the pinned prefix (39.3 → 58.0, above the parent arm). The --lengthmult 640 variant reaches 61.0 (parent 60.8)
mixture d025 --lengthmult 640	32768	61.37 ± 0.00 at N=8 (2026-08-05, 159 samples; LRU 59.88, ceiling 65.77 @20%, reactive 60.89)	pass: above LRU on every seed (margin +1.48 at the N=8 min). The parent reads 1.37 below LRU and its final third reads 64.85, i.e. +5.0 above — the parent is a convergence-cost row
phases d050	8192	2026-08-21 corner-deletion re-base: seeded 1–8 60.93 / 61.82 / 61.08 / 60.50 / 60.93 / 61.24 / 48.26 / 48.15 (mean 57.86; six seeds up +0.1..+1.1, the two low-basin seeds bit-identical). 2026-08-15 wedge re-base: seeded 1–8 60.47 / 60.88 / 59.98 / 60.00 / 60.39 / 61.43 / 48.26 / 48.15 (mean 57.45; pre-change 59.47 / 60.58 / 57.32 / 58.14 / 57.57 / 56.71 / 57.36 / 50.14, mean 57.16), unseeded N=8 gate 59.07 (pre 55.96); before that 51.9–58.6 batch means, range 45.0–60.5 (N=8; the 58.6 batch is the round-3 clock at N=8)	sentinel: no drift of the N=8 mean below 51.5 (N=8 57.17 ± 2.86, draws 50.14–60.08 — bimodal, adjudicate the mean, never a draw); 15.02pp below LRU 72.19, and this is the group's one genuine pin: its final third still reads 64.16, 8.03pp under LRU, where every other short row reaches parity or better given the samples (2026-08-05 thirds sweep). The old mean ≥50 sat at the low draw — basin lottery; the earlier ≥55 bar was an N=3 high-draw calibration, restated 2026-07-31 against a paired reference build
phases d050	32768	2026-08-21 corner-deletion re-base: seeded 1–8 61.15 / 59.97 / 59.53 / 57.76 / 46.68 / 61.50 / 57.71 / 53.60 (mean 57.24; seven seeds up +0.6..+1.6, seed 5 bit-identical). 2026-08-15 wedge re-base: seeded 1–8 60.48 / 58.40 / 58.45 / 56.84 / 46.68 / 60.38 / 57.10 / 52.95 (mean 56.41; pre-change 45.01 / 44.73 / 53.78 / 56.84 / 46.68 / 55.99 / 57.10 / 52.95, mean 51.63: four seeds up, four bit-identical), unseeded N=5 gate 57.20 (pre 49.08); before that 47.9 mean (N=5)	sentinel: no drift of the N=8 seeded mean below 51.6; the sample-aligned defeat persists at this scale on the unmoved seeds; cadence family, adjacent to F2/F3
deadphase	8192	2026-08-15 wedge re-base 49.00 (seeded 1–8, −0.16 on every seed against 49.16): each dead phase's probe now walks ×2/×4 and the live phase's opening sample lands farther out, an accepted price. Seeded 1–8 on 2026-10-08: 48.97–49.05, unchanged. The static terrain is a 49.6 shelf from a 1% to a 30% window (49.66 @1%, 49.63 at the 2% density floor) that falls to 48.62 @80%; LRU 48.13. The row first recorded the ceiling as 49.1 (2026-08-06), which does not reproduce: the generator is unchanged since, and the 08-15 reactive refresh already read 49.68	pass: no drift of the N=2 mean below 48.9. The 0.66 below the shelf is the price of returning from the dead phases' exploration (the exploration itself is ~free); a rise toward 49.6 means the return got cheaper
widepin	8192	2026-08-21 corner-deletion re-base, seeded 1–8: 56.69 / 56.66 / 56.85 / 56.49 / 51.34 / 56.84 / 57.43 / 51.58, mean 55.49 (five down by ≤2.4, three up by ≤1.3): the high basin gives back part of what the corner probe's churn found here, the deletion's one battery cost. 2026-08-16 audit-undo re-base, seeded 1–8: 58.29 / 59.09 / 58.52 / 56.24 / 50.85 / 58.82 / 59.12 / 50.24, mean 56.40 (six up by +2..+11.6, two down by ≤0.6): an audit's retreat no longer re-arms the starvation refractory, so the low basin's escape comes sooner. 2026-08-15 wedge re-base, seeded 1–8: 58.87 / 50.68 / 50.66 / 44.64 / 44.89 / 56.73 / 59.24 / 44.84, mean 51.32 (seven up, seed 2 −5.6: a basin flip on the coin-flip family). Pre-change seeded (--seeds 1,2,..,8): 56.2 / 56.3 / 49.3 / 42.9 / 43.1 / 56.7 / 48.6 / 43.5, mean 49.58. LRU 71.8. Re-derived 2026-08-05 over 24 unseeded runs: true mean 50.63 ± 5.25, range 42.5–56.3, basins 12/24 high (>52) and 12/24 low — a 50/50 coin, with N=8 batch means reading 52.04 / 49.52 / 50.34. The former record (52.4–54.9 batch means, bar "mean ≥50") was an unrepresentative high sample: reaching 52.4 needs 6+ high draws in 8, and the ≥50 bar sat at the distribution's own mean, so a random batch failed it about half the time and any verdict taken from it was a coin flip. Adjudicate this family seeded only, as slowswap already is	pass: no per-seed drift below the recorded values −noise, seeds 1–8, arms rotated inside each seed. Never an unseeded batch mean. Limits family softened by the marginal verdict and the anchor; the rest needs a lag-free signal
straywall2 d050	8192, 16384	55.1 attractor (2026-08-04 re-base by the blind-corner fix, +9.2 seeded N=8 from 45.89; battery N=4 reads 55.5; the pre-fix record was 46.1 with one 35.1 draw in 8) / 54.99 ± 0.21 @16384 (2026-08-07 seeded N=8 re-base from the stale pre-fix 45.7; draws 54.62–55.13; LRU 57.58, ceiling 64.70 @40%)	sentinel: no drift below 54.8 on the attractor at 8192 (N=8 55.12 ± 0.15) and below 54.5 at 16384; 2.39 / 2.59pp below LRU 57.51 / 57.58 — the same audit-completed escape geometry at both sizes (floor start, ~50% window landing, 2 audits 1 confirm). The old ≥53 allowed 2.1pp of silent decay — the rung-scaled stride escape, now completed by the audit
blind_balloonflip	8192	2026-10-08 re-base after the half-full gate: seeded 1–8 80.49 / 80.41 / 80.30 / 83.93 / 81.03 / 80.47 / 80.09 / 80.31 (mean 80.88, bimodal: s4 lands in an 84 basin the other seven miss; reactive 85.09). The gate ("hold the adaptive window until the cache is half full", 2026-09-08) lifted seeds 1–4 from 77.92 to 83.19 in an A/B across it; the fill's samples no longer count (59 decisions, was 68). Before that, the 2026-08-16 audit-undo re-base: seeded 1–8 77.77 / 77.78 / 77.85 / 77.89 / 77.72 / 77.82 / 77.81 / 77.76 (mean 77.80, −0.26 on every seed against 78.06: an accepted price of the faster post-audit retry). Before that 78.04 ± 0.14 at N=8 (2026-08-05 re-measure: the recorded 79.5–83.0 scatter / ~79.8 mean no longer contains the value, which is now tight rather than scattered (sd 0.14 across 8 seeds, range 77.85–78.30) — so the row's bimodality claim is rotted too and it has been dropped from the bimodal list above; treat the old range as rot, not as noise; LRU 52.1; ceiling 86.57 @1% window, reactive arm 87.01 — the layer is 7–8pp behind BOTH, see below)	pass: far above LRU on an N=8 mean — but read that bar knowing it is 29pp slack: the reactive anchor prices the real gap. The trajectory is the ε-asymmetry balloon (hill-climber.md §5) doing exactly what it is designed to do — on seed 1, 17 of 59 samples have main earning zero, the window holds 1–2% through s11 and rests at 18%, then balloons 18% → 80.2% in three samples at s32 and holds, mean window 35.2% against a 1% optimum, and the probe machinery does not recover it (1 probe, 2 audits, 0 confirms). Not a regression to chase; the cell that prices the zero-main prior, and the row to re-run if the ε-asymmetry is ever revisited — the family went spready (R4-F1 class), so the parkFreshLeft-latch signature (a fall to ~79.2 from a starvation confirm's stale shield inherited by an early veto park, recovered by the latch fix) is no longer readable from a single draw; the calibration audit itself costs ~nothing here
blind_nullchurn	8192	2.48 ≈ LRU	pass: perpetual probing is harmless at equilibrium
blind_bandtrap2	8192	2026-08-15 wedge re-base: seeded 1–8 75.51 / 76.25 / 76.31 / 76.00 / 76.17 / 76.39 / 76.29 / 76.28 (mean 76.15, +4.05 on every seed against 72.10); unseeded N=2 gate 76.16. Every one of its 293 starvation confirms (8 seeds) was a wedge that reset the ladder; the escalated ×4 walk overshoots to the top corner and the audit's descent confirms at the band (~31%). Before that 72.6 (ablation 69.1; LRU 76.5)	pass: no drift below 75.5 at N=8 seeded, and now 0.4pp below LRU 76.53 (was 4.43 below; the old bar of 71.5 stands as the pre-change record). The B1 discard geometry is repaired
trickle (s7/s11/s13)	8192	2026-08-15 wedge re-base: s7 seeded 1–8 73.52 on every seed but s7 (73.57), mean 73.53 against 70.70; s11 73.40 / 73.41 (mean 73.40 against 72.46); unseeded N=2 gate 73.52 / 73.40. Before that 69.9–72.2 / 73.0 / 72.5–73.1 (ablation 56.1; LRU 73.8)	sentinel: no per-seed drift below 73.0 (s7) / 73.2 (s11), now 0.26 / 0.39pp below LRU 73.79 (was 3.09 / 1.35: the wedge reset kept the ladder from reaching the trickle band); pre-change bars 68.0 / 72.2 (N=8 70.70 ± 1.26, draws 69.40–73.03 / 72.46 ± 0.01). The old ≥66 sat 7.8pp under the do-nothing baseline and could not fail for the reason the row exists — the F3 prize, captured
lowmix (s7/s11/s13)	16384	s7 5.94 (2026-08-04 re-base by the blind-corner fix, +3.17 seeded N=8) / s11 2.89 / s13 2.94 — s11 and s13 not re-measured (pre-fix s7 2.75; round-3 re-base; pre-round-3 3.11 / 3.23 / 3.11; average-verdict ablation 6.94 / 2.03 / 6.96; LRU 7.55)	sentinel: the marginal verdict's NAMED TRADE — a bistable 2–7%-HR family where the frozen baseline false-vetoes an LRU-ward escape; bar = no drift below the recorded values −noise. The round-3 re-base is one calibration audit now reaching the family (decay clock + preserved schedule) and landing the bistable low basin — same excursion class as zigzag's −0.7
demoflood	8192	68.6 (2026-08-04 re-base by the blind-corner fix, +2.23 seeded N=8 from 66.36; LRU 16.7)	pass: no drift below 68.0 with confirms (N=8 68.58 ± 0.02; LRU 16.66, so an above-LRU bar is 52pp of slack and the row must be barred against itself). The old ≥66 allowed 2.5pp of silent decay — the freeze-earned construct; a fall toward ~59-with-zero-confirms means the frozen baseline broke (that is the live-adjudication pin)
whisper (attack)	8192, 16384	66.8 / 66.6 (LRU 64.6)	pass: above LRU on every seed (LRU 64.57; N=8 66.78 ± 0.00, margin +2.21). The old ≥64 sat 0.57 below LRU and could not fail for the reason the row exists — the sharpest F4 repro, 0.17% of requests just above the starvation bar keeps the window "sighted" so no starvation probe arms; owned and repaired by the equilibrium audit, and the calibration audit removed its pinned prefix (64.1 → 66.8, now clear of LRU) — a fall toward the old 55.5 pin means the audit clock or its retry floor broke
whisper (quarter)	8192	66.9 (LRU 64.5; average-verdict ablation 55.6)	pass: above LRU on every seed (LRU 64.48; N=8 66.82 ± 0.00, margin +2.33; the old ≥64 sat below LRU) — the same trace with the whisper dose below the bar: the probe arms and the marginal verdict confirms the escape the average form fails
whisper + modulation a0.08/p12	8192	65.3 (pre-streak 60.5; LRU 64.5)	pass: above LRU on every seed (LRU 64.50; N=8 65.15 ± 0.00, margin +0.64; the old ≥63 sat 1.5 below LRU) — the F1 dose repair: window-irrelevant ±8pp rate modulation must not turn the audit off; the streak confirm holds LRU parity through it
whisper + modulation a0.08/p6	8192	64.4 (crash-semantics re-base: escalated persistence owns −1.5 — post-crash audit retries tolerate the period-6 dips two samples longer and the retry cadence de-phases from the modulation; was 66.0 after the round-5 zero-step's −1.15; pre-streak 56.4; parent reactive 58.1)	sentinel: no drift below 64.3 (N=8 64.48 ± 0.03), which is LRU parity −0.06 (LRU 64.54). The old ≥58 allowed 6.5pp of silent decay — above the reactive arm inside the lag-limit band; the study's journaled trade row
whisper + modulation a0.12/p12	8192	64.0 (crash-semantics re-base of 62.2: escalated TIME persistence lets post-crash audit retries survive the dose and confirm; parent 64.7; LRU 64.4)	sentinel: no drift of the N=8 mean below 62.8 (2026-08-18 re-base (the crestpast repair: a confirm's verdict is the best sample of its confirming run, and a park's retreat and its landing are the park's own test). N=8 63.28 ± 1.15: 63.89–63.92 on six draws with a 61.41 basin on two, where the row was unimodal before. Same F1 exposure as crashnoise_a12. Pre-repair N=8 64.00 ± 0.12); 0.37pp below LRU 64.37. The old ≥62 sat 2.4 under LRU and the −noise clause carried the row alone — formerly the AUDIT-side walk-interior sentinel (adv3 F1's open half), closed 2026-08-02: the depth-priced counterfactual's 66.75 remains unreachable without the holdout-fatal pricing trade (the R4-F1 dial), and the persistence heal banks +1.8 of it. Park retention remains the open thread for the w097 weather family, where audits DO confirm
mixmod a010 / a000	32768	56.6 dosed / 58.8 undosed	pass: dosed above LRU on every seed (LRU 55.32; N=8 56.61 ± 0.02, margin +1.25; the old ≥55 sat 0.3 below LRU) — the F1 dose instrument; a fall toward 48 means a deviation-priced confirm came back (the ~2pp dose cost is the modulation slowing the early escape, not the audit dying)
posjam flat / j50 / d25	8192	66.9 / 66.4 / 66.5 (LRU 64.9; ceiling 68.1; pre-round-3 the jams pinned at 56.0 = the audit-free value, −8.9 below LRU)	pass: above LRU on every seed, all three (LRU 64.93/64.90/64.93; N=8 66.92 ± 0.00 / 66.40 ± 0.01 / 66.53 ± 0.01, margins +1.99/+1.49/+1.59; the old ≥64 sat 0.9 below LRU) — the round-3 position-jam repair. j50 (±50% jitter) needs the audit-owned schedule; d25 (period 2.5) needs the decay clock; a fall toward 56 means one of them broke. flat is the dose-matched control
posjam d0	8192	56.08 ± 0.05 seeded N=8 (reactive arm 65.37 ± 0.20 — the layer is 9.29pp behind the climber it replaced here, and below this cell's LRU 64.9 where the reactive arm is above it; the dose-matched flat control inverts, 66.92 ± 0.01 vs reactive 64.61 ± 10.71, so the machine is better AND deterministic on the same traffic without the jam)	sentinel: the sample-aligned every-sample jam — starves even the decayed clock, needs ±¼-sample grid alignment (D4 exposure class); the standing "non-consecutive stillness" study, whose warrant is the 9.29pp above rather than a forgone increment; bar = no drift below 56 −noise. Class boundary, verified negative (don't re-derive): the analytic starvation bound is ≥½-motion, but 50%/67%-motion cadences (--period 4 --duty 0.25 / --period 3) do NOT jam — audits arm and noaudit outscores the layer there (inert, not suppressed); every-sample motion is the only realizable regime at those fast cadences. Dose note (2026-08-21, Sol /audit-regret entry_duty, spec audit-regret/specs/entryduty.json, no row): a square duty at period 13.5 on a marginally worthless trickle rider over a real 40–50% prize re-creates the starved clock below every-sample motion (~70% post-failure motion, far_band 0.28 at 8192, workload seed 149 only; not scale-stable, at 16384 every steering sample moves), and the jam is not the loss there: noaudit ties hybrid within 0.3pp while both sit 17.2–17.9pp under the interior ceiling on 8/8 cache seeds, a class-3 masked-density steer (the rider keeps the window sighted and steering rests in the 10–23% band against a 40–50% ceiling), adjudicated trace-seed-stable at workload seeds 149/211/313 by the fable lane; the reactive arm's 15.4pp win at proportional 16384 is seed-149-measured only. Optional dose control (2026-09-17): audit-regret/specs/bandrider.json, a finite-distance sinusoidal rider, gives hybrid ≡ noaudit 51.44 at 256 continuous samples on seeds 1–8, against reactive 62.01 and the best tested static reference 64.94. Larger live starts succeed; this is a path-dependent dose of the existing family, with no new mandatory bar. See hill-climber.md §3 for geometry, continuation and neighborhood controls. Dose note (2026-09-24, /audit-regret stillness_starve, spec audit-regret/specs/stillness_starve.json, no row): a trickle rider and a two-reference band square-modulated together (every period from 2.3 to 9.0 samples) keep 63–90% of samples moving, so no audit arms in 128 samples, the calibration audit included; 51.16 ± 0.13 seeded N=8, identical to noaudit, against reactive 56.08 ± 0.34 and a 60.0 static peak; the extreme of the starved clock below every-sample motion.
shieldtrap (s7/s11/s13)	8192	2026-08-16 audit-undo note: s13 seed 1 77.84 → 76.50 and seed 3 77.82 → 78.22, the other six seeds and s7/s11 within ±0.3 (s13 mean 78.76 → 78.64). Before that 78.84 ± 0.27 / 78.94 ± 0.22 / 78.62 ± 0.83 at N=8 (2026-08-05 re-base by the audit-bar split, +0.79/+1.75/+1.17 over the pre-split 78.06 ± 0.61 / 77.19 ± 0.80 / 77.44 ± 1.41; LRU 81.3; ceiling 83.1 @40%)	sentinel: R4-F1, the audit-amplification give-back — the round-3 clock repairs take audits 2 → 7-8 and the layer's run share 6% → 43% (park 32% + walk 12%) on a regime-alternating family already below LRU, with nothing confirming usefully, and turn a deterministic family bimodal. Bar = no drift below the recorded means −noise, N=8 mandatory. The calibration audit itself is worth +9.6 here (round 2), so the row does not say "fewer audits = better" — a change that recovers the give-back without re-opening the posjam rows is the welcome shape, and the audit-bar split is one: part of the give-back was AUDIT_BAR_FRACTION pricing the reversal below the noise, not only audit reachability. The prior entry recorded 77.9 / 78.9 / 79.0 from the 2026-08-04 re-base, above where the arm actually measured at N=8 on s11/s13 (77.19 / 77.44, reproduced in two independent sessions) — a stale bar of exactly the kind the process note names, so re-measure this row rather than trusting a recorded mean. Dose note (2026-08-15, /audit-regret flashpark, spec audit-regret/specs/flashpark.json, no row yet): the same shape at a rising dose is the audit layer's largest measured cost, −3.81 against noaudit (−3.47..−4.00 across seeds 1–8; hybrid 49.59 ± 0.27, reactive 53.67 ± 0.08, noaudit 53.40 ± 0.05; ceiling 56.18 @20%, LRU 45.88): a 14-sample recency opening then a steady core; the calibration audit crosses the flash's capture point (+1.89 earned), overshoots two strides, confirms at 41%, the regime ends one sample later announced by a +10.6pp rise the fresh-park shield holds through, and the park costs 5.89. Rowed with its fix per the standing rule
saw_p40	8192	54.7 (round-5 re-base: the audit-confirm zero-step turns the fall-through-labeled confirms into clean parks, +0.56, crashes 3 → 0; pre-round-5 54.1 ± 0.03; pre-round-3 55.1; noaudit 55.5; LRU 48.7)	sentinel: R4-F1's deterministic instance — a 40-sample window-irrelevant wave on a floor-optimal base; the gap to noaudit is bounded audit-excursion duty. Criterion restated round 5 — the old "confirms must stay 0" was an exact-label artifact (the baseline already took 3 fall-through-labeled confirms): HR holds ≥54.1 −noise, and under the debug harness confirms stay a small parking band (~3–6, crashes 0) — a confirm explosion or a fall toward LRU means the misconfirm gate broke
loopcliff	8192	98.3 (LRU = ceiling = 98.4)	pass: no drift below 98.3 (N=8 98.36 ± 0.01), which is LRU parity −0.06 on flat terrain (LRU = ceiling = 98.42) — the no-cliff invariant at the structural misconfirm landing (~⅓ of max; round 2 R9): W-TinyLFU serves the loops flat across every static window, so any material fall means a confirmed park manufactured a cliff that does not exist in the terrain
climbtrend up / down	8192	2026-08-19 re-base (the return retest): down 55.44 on seeds 1-2, +0.47 over the tree's 54.97 — a trend's stale claim is discarded when the veto's return arrives instead of pinning; up is unmoved at 55.16. 55.29 ± 0.01 / 54.64 ± 0.05 (ceilings 55.4 / 55.6 @1%; LRU 49.4)	pass: no drift below 55.0 (up) / 54.6 (down), both above LRU 49.4 on every seed (N=8 55.16 ± 0.03 / 54.80 ± 0.02). The old ≥54 carried ~1pp of slack on a row whose whole claim is that the mechanism costs ~nothing — the trend-driven misconfirm family: the mechanism fires (round 2 confirmed a 36% window on a +7pp trend) but must keep costing ~nothing; a fall toward LRU means trend confirms started parking wrong
crashnoise flat / sine a0.12	8192	66.8 (±0.03) / BIMODAL 60.8 or 63.6, high basin ~70–80% of draws (26/32 and 22/32 on two paired arms, 2026-08-02 — the row's "deterministic" record was an N=2 artifact; adjudicate on an N≥8 mean and read the basin count, never a single draw); crash-semantics re-base of 61.2–61.5; LRU 64.6 / 64.1; ceilings 67.7 @20% / 67.6 @30%)	flat: pass above LRU on every seed (LRU 64.57; N=8 66.78 ± 0.00, margin +2.21; the old ≥64 sat below LRU) — the dose-matched zero control. a12: pass no drift of the N=8 mean below 60.5 (2026-08-18 re-base (the crestpast repair: a confirm's verdict is the best sample of its confirming run, and a park's retreat and its landing are the park's own test). N=8 61.70 ± 1.05, draws 60.69–62.69, the low basin now 4 of 8 against 1 of 8 before; the verdict half parks on the a12 sine's own peak, moving the window off the @30% optimum, which is the F1 exposure hill-climber.md the F1 exposure carries. Pre-repair N=8 63.25 ± 0.94, draws 60.75–63.63; 0.80pp below LRU 64.05, so above-LRU is unreachable). The old ≥61.2 sat inside the seed spread, so a single draw decided it — formerly the AUDIT-side walk-interior sentinel: the dose's scatter crossed the audit's absolute 5pp bar and killed audits mid-walk (−2.8 below LRU). Closed 2026-08-02 by escalated TIME persistence (the retry of a crashing equilibrium tolerates two below-bar samples, holding direction), which banks +2.1 of the depth-priced counterfactual's +5.4 without its holdout-fatal trade (the 66.93 audit-bar form was REJECTED at a fresh holdout — P8 basin-kill −2.5, w097 −1.1; the absolute DEPTH bar is load-bearing on real traces; don't re-run pricing variants, see hill-climber.md §5 for the dead-candidate journal)
mixnoise flat / sine a0.10	8192	59.8 / 60.7 (crash-semantics re-base of 59.4 / 59.8; LRU 54.8 / 55.1; ceilings 62.4 / 62.6 @20%; a10 was 51.6 before the probe-side pricing)	flat: pass above LRU on every seed (LRU 54.76; N=8 59.76 ± 0.01, margin +4.99). a10: pass above LRU on every seed and no drift below 60.4 (LRU 55.06; N=8 60.70 ± 0.02, margin +5.61; the old ≥58 carried 2.7pp of slack) — the probe-side walk-interior pricing's prize row (adv3 study): the dose's wave crash-cycled every starvation up-probe one stride off the floor (arm → crash → 16/32-sample hold → repeat), severing the mixture trap's only escape; the priced-capped bar walks through the weather and the escape confirms by sample ~5. A fall toward 51.6 means the probe bar's pricing or its 3×RESTART cap broke. This row killed every escalated-DEPTH audit-bar candidate (escbar −2.48, intrinsic); TIME persistence instead lifts it +0.9
h4c1 attack / reverse	8192	66.4 / 65.5 (LRU 64.56; ceiling 67.77 @40%; pre-fix the attack pinned at 58.55, −6.0 below LRU, the reverse control at 65.41)	pass: above LRU on every seed, both arms (LRU 64.56 both; N=8 66.36 ± 0.00 / 65.54 ± 0.00, margins +1.79/+0.97; the old ≥64.5 sat 0.06 below LRU, i.e. at it) — the H4-C1 crash-ratchet repair (Terra's round, 2026-08-02). Three token-multiset-preserving high→low pulses each killed one audit walk; the shared crash streak adjudicated crashes 2–3 as failures (rung 16→32→64) and the deepest-rung branch doubled the wait to 128 — a 130-sample floor pin. The fix: per-layer crash ledgers, crash-never-doubles-the-clock, and escalated time persistence absorbing the 1-sample pulses. The reverse order is the dose-matched control; contest a verdict with the full 5-phase × 2-order lattice (gen_h4c1.py --phase 3..7 --order high-low|low-high, 4M short form) — pre-fix the machine trapped 5 of 10 arms at ~55.5, post-fix the worst arm is 63.2. A fall toward 58.5 with auditWait ratcheting past the ladder in the trajectory means the crash/failure distinction or the ledger separation broke
moat h7800 / h3000	8192	45.28 ± 0.93 at N=8, draws 44.32–46.23 (2026-08-06 seeded re-base; the recorded 46.2 was an N=2 unseeded reading near the top of the distribution, which is what put the old ≥45 bar inside the spread; 12M requests; LRU 37.92; ceiling 49.58 @40%; noaudit 43.16; pre-fix 41.97 with 0% of 366 samples in the prize band and auditWait decaying to 256) / 58.7 ± 0.9 (LRU 56.42)	h7800: pass above LRU on every seed (LRU 37.95; N=8 45.28 ± 0.93, draws 44.32–46.23, margin +6.37) and above noaudit 43.16. The old ≥45 sat inside the seed spread and the row fails it on 3 of 8 seeds today — a bar nobody noticed was already red — F1-adv4, the terrain moat: a hit-rate valley deeper than the audit's 5pp bar between the density rest and the prize was absorbing at every rung (the abort fired at the −bar contour, one stride short of the far bank); the crash-streak retry's two-sample tolerance crosses it and parks in the prize band. h3000: pass ≥55 — the dose-matched zero control (identical traffic, monotone approach; the audit layer is worth +13 here), a fall means the escapes the layer already makes broke. Real-corpus static scan: 0/22 cells show moat terrain (constructed-only); mid-dose h6500 pays +0.2 vs pre-fix — journaled, not rowed
moat h4000 / h5000	8192	2026-08-20 re-base (d368ebd38): N=8 unseeded reads 60.54 ± 0.01 (draws 60.53–60.55) / 60.14 ± 0.02 (60.11–60.16), against the recorded 57.91 / 57.77 — the gap is real drift, not a seeded/unseeded artifact, and the family got more deterministic (sd 0.05 → 0.01 and 0.17 → 0.02). Both rows are now within 1.9pp of their static ceilings (62.46 @40% / 62.03 @30%) and +9.40 / +13.61 over LRU. The rise is unattributed: it accumulated across the repairs that landed after the 08-09 rowing (crash semantics, shallowmoat reach and retention, audit-undo, absolve, ghostclaim, crestpast), and the crestpast verdict is the likeliest single source on a cell whose whole question is crossing a valley to a prize. Nothing was bisected. Read the new means, not the ones below. Superseded: 57.91 ± 0.05 / 57.77 ± 0.17 (N=8 unseeded; LRU 51.14 / 46.53, ceiling 62.46 @40% / 62.03 @30%). Rowed 2026-08-09; h5000 was journaled at −2.0 vs pre-fix by the crash-semantics round, which is the drift this row now watches	pass above LRU on both (+6.77 / +11.24). These are the guard rail's only behavioral coverage: the rail is the sole mechanism that recovers the window at these doses (§4's guard-rail note: removal costs +2.44 / +3.36, 8 of 8 seeds), and it is measured firing here — 1 veto per cell at seed 7, against 0 on the deep moat_h7800 and 4 across the other 56 cells combined. Rowed because an adversarial audit re-derived "the rail has no coverage" from the three-cell JUnit subset; the gap was never coverage of the mechanism but of the doses where it pays. A fall means the rail stopped engaging
slowswap step / ramp	8192	2026-08-20 re-base (d368ebd38, LRU and ceiling re-derived unchanged at 36.79 / 43.12 @50% and 36.53 / 43.43 @50%). r1 seeded 1-8: 40.25 / 40.50 / 40.44 / 40.21 / 39.76 / 40.34 / 40.67 / 40.58, mean 40.34 ± 0.27 — the same eight values measured on two trees a day apart, so the drift from the recorded mean 40.43 is settled, not noise. r20 seeded 1-8: 42.00 / 42.89 / 42.11 / 42.98 / 42.98 / 42.00 / 42.99 / 42.11, mean 42.51 ± 0.46. The two basins survive but each has picked up ~0.1 of internal spread, so the old 'a third value is the leak' tripwire no longer reads cleanly on exact equality; what it was protecting is intact — no draw below 41.8, and the draws stay in two clusters near 42.0 and 42.9. Bar restated to the criterion the row already says it adjudicates on: mean ≥ 40.2 at N=8 and above LRU on every seed. The old ≥40-every-seed form fails on seed 5 by 0.24 while the mean sits 0.14 above its own real-signal floor, which is the bare-level rot the release battery names. Superseded: seeded (-Dcaffeine.climber.seed=1..8): r1 40.34/40.27/40.22/40.85/40.59/40.13/40.60/40.44 (LRU 36.79, ceiling 43.12 @50%, noaudit 27.6; staleclaim seed 4 = 29.21, flatroom seed 4 = 27.32); r20 42.00 ×4 / 42.88 ×4, mean 42.44 at N=8 (LRU 36.53; 42.14–42.24 on all eight before the 2026-08-18 repair)	pass: r1 mean ≥40.2 at N=8 and above LRU on every seed; r20 no draw below 41.8 (LRU 36.79/36.53; N=8 40.43 ± 0.23, draws 40.12–40.88 / 42.44 ± 0.44 in two basins). The step arm's bar sits 0.12 under its own lowest draw, inside one sd — a single red here is one draw's worth, so re-run at N=8 before believing it, and treat a mean below 40.2 as the real signal.** r1 is the stale-away-anchor sentinel, closed 2026-08-03: a regime step discards the anchor, and before the fix the goal metric smoothed across the shift, so the claim re-planted from it was ~80% the old regime's rate and froze the moment density steered away — on seed 4 the audit walk reached the ceiling, earned +35pp for fourteen samples and was rejected against a 0.4415 bar, landing on the audit-free value (27.6 pre-room-rule, 9.5pp below LRU; 29.21 with the room rule fixed, 7.6 below). A drift toward 29 on seed 4 means the discarding stand-down stopped re-seeding Rates, and toward 27.3 means the rung-scaled room rule broke with it (that is what the two A/B arms separate). r20 is the dose-matched control (same shares, reuse distances and request count; only the transition length differs). It was deterministic until the 2026-08-18 crestpast repair and is now two basins, 42.00 on four seeds and 42.88 on four (mean 42.44, above the recorded 42.14–42.24). Not a leak: it is bit-identical to ship through the s174–s179 rung-32 audit walk, and the split is made afterwards, when the verdict parks on that walk's best sample and 0.65pp of admission-draw noise on one of its samples decides which stride that is — 6267 (76.5%, where corner probes keep crashing and each crash's one-stride retreat lands on the anchor with the hold already released, so the swing discards the claim every cycle) or 4219 (51.5%, no corner probes). So the control's test is no longer "any spread": a third value, or any draw below 41.8, is the leak this row was written to catch — it is also the row that killed symmetric claim aging, which reads 30.3 on every seed. Adjudicate this family seeded only: unseeded it is bimodal on which draw lands in the trap. The committed swap timing sits near the shipped climber's best: regenerating the step trace with slowswap.py --ramp 1 --p3 139 --p1 126…174 moves the swap ±24 samples and reads 33.3–41.8 for it across seeds 1–2 (40.3–40.6 at the committed 150). A change that shifts the audit schedule (an earlier first audit, a new arming rule) has its own curve over those timings, which the committed one can favor or punish, so compare across them as well; that comparison supplements the bar and does not replace it
cp_w097 (real, cloud-physics corpus)	16384	46.9 ± 1.9 at N=8 (the cell has ~2pp run-to-run variance — adjudicate on an N≥8 mean, single runs read 45.5–47.3; round-3 fix vs pre-fix is +0.47, unresolvable; pre-streak 43.8; parent 50.1; best static 50.9 @10% window)	sentinel: the density arm's setpoint family — no drift below 45 on an N≥8 mean; the robust signal here is the audit layer's own worth, +3.3 vs noaudit (43.59 ± 0.04). Skip and say so when the corpus is absent
LLM prefix cache (real: Qwen Bailian Thinking, Coder and To-C; kvcache.py)	thinking@16384, thinking@32768 ×4, coder@65536, toc@185364 ×8	2026-10-10, seeded 1–8 on the harness at HEAD (49 / 98 / 59 / 68 density samples): hybrid 32.66 ± 0.10 (32.54–32.88) / 35.34 ± 1.42 (32.95–36.97) / 39.34 ± 0.32 (38.99–39.68) / 34.74 ± 2.03 (33.23–33.35 on five seeds, 36.53–37.49 on three); reactive 32.48 / 36.60 / 40.53 / 39.13; LRU 36.26 / 39.61 / 42.15 / 45.77; ARC 37.37 / 40.77 / 42.69 / 45.80; static 1% → 80% 32.53 → 36.93 / 32.90 → 39.92 / 37.25 → 42.52 / 29.57 → 45.24, rising throughout. One-pass references: an 80% plant (seed 1) 36.99 / 39.82 / 42.16 / 43.97 and the MiniSim leader 37.01 / 39.83 / 42.54 / 44.20, so the basin is reachable and held. The loss is steering into it (hill-climber.md §3, terrain classes): a walked window leaves the hot set in main and reads sparse to the density law, the average law is bistable on the static terrain, and probes and audits are judged across 14–60% per-sample phase swings. The reactive law leads on three of the four, and is still 1.6–6.6 under LRU there. thinking@16384 sits at the floor on every seed; thinking@32768 ×4 (the paper's 48 GiB cell, 7.2 under ARC on one pass) is an escape lottery; coder@65536 is the walk back down after a climb; toc@185364 ×8 is the worst one-pass cell (19.5 under LRU) replayed, where two-thirds of that gap is sample budget and the rest a park at a 32–43% window. The replays make thinking@32768 and toc@185364 barrable: one pass holds 24.6 and 8.5 samples. The replays reuse the trace's keys, which the frequency sketch remembers from pass to pass and which keep the climber small; with fresh keys each pass it reaches LRU or the static 80% window within two to four passes, so the ×N cells' gaps under LRU are partly that memory rather than steady-state loss. Its hit rates also count every block independently, while an engine's prefix lookup stops at a request's first missing block: in this replay LRU's usable hits equal its block hits, but a small window's admission breaks chains and the shipped climber's usable hits run 0.8–2.8pp below its block hits (To-C@185364 23.4 against 26.2)	sentinel: no drift of the N=8 mean below 32.46 / 34.3 / 38.70 / 33.8, read with the per-seed vector. The second and fourth are bimodal on whether a seed escapes, so one seed's flip (0.5 on the mean) is inside the bar and a collapse to the floor (33 and ~30) is not. The gaps the row tracks are 3.60 / 4.27 / 2.81 / 11.03 under LRU; a rise is the point. Skip and say so when the corpus is absent
LLM prefix cache controls (real: Qwen Bailian To-C and To-B; kvcache.py)	toc@16384, tob@16384	2026-10-10, seeded 1–8 (96 / 152 samples): hybrid 18.84 ± 0.20 (18.47–19.03) / 47.45 ± 0.24 (47.00–47.81); reactive 18.72 / 47.63; LRU 15.49 / 40.29; static 1% 18.73 / 48.72, the optimum, falling to 16.93 / 43.78 at 80%. The frequency-favorable cells from the same service: the window belongs at the floor, and To-B's 1.27 under its static optimum is the exploration the machine already pays there	pass: above LRU on every seed and no drift of the N=8 mean below 18.44 / 46.97. A repair for the row above that grows the window here gives back up to 1.8 / 4.9
metronome r1	8192	2026-08-16 audit-undo re-base: seeded 1–8 39.78 / 38.68 / 38.59 / 38.54 / 38.70 / 38.65 / 39.90 / 40.10 (mean 39.12, −0.92 against 40.03, seven of eight down): the jam family's control pays for the faster post-audit retry, an accepted price. Before that 40.0–40.2 (re-based 2026-08-04 from 40.3–40.6 by the blind-corner audit: the run trades 14 probes for 4 probes + 4 audits and the window roams further, mean 3.6% → 15.0%, which is bounded excursion duty on flat terrain rather than the wedge this row guards); window at/above the 2% floor after the first refractory hold	sentinel: the all-blind wedge class — a run stuck at 1% of maximum means the hold's below-floor lift broke. Under the priced-capped probe bar the window roams more than it used to (@floor ~92% → ~14%, mean ~10%) at a tied hit rate — roaming here is not a regression (flat terrain; the cap is what re-aborts deep damage; uncapped pricing measured −0.6)
esc_jam	8192	24.9	sentinel: rate-swing jam family with the fresh-park shield and calibration audit present — must tie the pre-F1 value
regimeramp	8192	54.03 ± 0.01 at N=8 (2026-08-05 re-measure; the recorded 53.5 had rotted 0.5 low; pre-calibration 54.8; LRU 50.11; ceiling 55.9)	sentinel: no drift below 53.8, above LRU 50.11 on every seed (2026-08-18 re-base (the crestpast repair). N=8 53.92 ± 0.01, deterministic: the retention half keeps the s149 down-audit's 44.6% park where ship's landing released it toward the @30% ceiling. Pre-repair 54.03.) — the calibration audit's priced cost (~−1.3 after round 2's auditBeatBase gate recovered +0.6 of it) — a steady rise clears any raw streak, so the early confirm can park wrong for ~a shield before the next audit re-tracks; a fall toward 48 means the park stopped releasing (the persistence failure class). Also the behavioral sentinel for the cold-start misconfirm class: its stationary variant is pinned by the two auditBeatBase unit tests (the archived round-2 flatctl/coldstart instances are workspace-only; a discriminating stationary generator was attempted and measured inert — adversary-ledger row 27)
tenant_s10	8192	92.6	sentinel: co-tenant class — must tie the pre-F1 value
blindlock blind / sighted	8192	50.7 / 55.9 (LRU 59.05 / 58.97; ceiling 64.15 / 63.93, both @40%; pre-fix the blind arm read 34.9)	sentinel: blind no drift below 50.2 (N=8 50.79 ± 0.24); 8.26pp below LRU 59.05, so an above-LRU bar is unreachable. The old ≥48 allowed 2.6pp of silent decay — the blind-corner lockout row (2026-08-04). One terrain; the arms differ only by a window-irrelevant whisper rider (0.28% of requests) that lifts the window's sample hits over the starvation bar, so hasBlindCorner is false on the sighted arm. densityClimb ranked the blind-corner branch above auditClock.isDue, and refractoryLeft is armed by every undoProbe but decremented only in holdInRefractory, so a corner that never clears served its whole 16/32/64-sample backoff motionless while the clock was permanently due. Check the anchors before the arms: the rider replaces a token rather than inserting one, so the terrains match to ~0.1pp rather than exactly, and that match is what makes the ~5pp arm gap attributable to the router. A fall toward 34.9 with the sighted arm unmoved means the ordering regressed. The lockout is a convergence cost, and no long companion is kept (2026-08-05): at --lengthmult 640 the arms converge to the same rest, final thirds 62.13 blind against 61.74 sighted — the gap reverses and lands inside noise, and the blind arm reaches LRU parity (58.59 against 59.11). Both long instances were generated and measured, then dropped rather than added, since a row whose phenomenon dissolves has been retired rather than improved. The short pair keeps the phenomenon and is the instrument. Run it with floors.py — the fix's cost side is the thin-signal floors
strad_p8 @4096 / @4097	4096, 4097	82.4 / 76.90 (2026-08-16 audit-undo re-base: @4097 76.76 → 76.90, +0.10..+0.20 on every seed; 2026-08-15 wedge re-base: @4097 74.83 → 76.76 on every seed, its 32 confirms per 8 seeds all wedges; @4096 unmoved at 82.41–82.52; LRU 82.21 both; ceiling 86.39 / 86.42)	sentinel: the D2 tier cliff, open, narrowed 7.58 → 5.65pp — one entry changes the answer while the terrain does not move. Both pure laws are continuous across the boundary (reactive 82.41 → 82.42, density 74.79 → 74.83) and LRU steps 0.00, so the discontinuity is the policy switch, not capacity. The strad_stat arm is the dose-matched stationary control (+0.69: the alternation is load-bearing). Bar = no drift beyond the recorded values ±noise; a fix must move the pair together. Note the gate's corda + 5×loop + corda row already claims "no tier-boundary cliffs" and passes while this exists — it is not a sufficient instrument
thin-signal floors (real: w50@123038, S1/S2/S3, DS1)	local corpus	per-cell references (2026-08-07 gate run, N=3, spreads ≤0.13): w50@123038 48.51, S1 5.16, S2 7.23, S3 7.15, DS1@1051635 14.24. 2026-08-15 wedge re-base: DS1@1051635 13.44 (seeded 1–2 13.41 / 13.48 against 14.19 / 14.14), an accepted −0.7: a ten-sample trace whose samples alternate near-dead and live; after the first wedge the ×2 walk armed on a low sample spans two live samples at a 27–39% window (a walk armed in a trough has no crash abort, §5's v7); the other four cells have no starvation confirm and are bit-identical. 2026-10-10 re-base after the audit-undo change: S2 7.02 (seeded 1–3 7.07 / 6.93 / 7.07 against 7.22 / 7.19 / 7.20), an accepted −0.18 (−0.17 at N=8, every seed down): an audit's retreat no longer re-arms the starvation refractory ("Leave the starvation refractory alone after an audit's retreat", 2026-08-16, which priced S2 at −0.13 but did not re-base this row), so the 32-sample hold armed by the first probe's fail runs out through S2's two audit crashes at s51, and a second up-probe walks off the 2% floor to 27% and fails (seed 1); S1 is bit-identical and S3 within 0.03. S1 has never moved: it is two-basin, seeded 1–8 5.25 / 5.36 / 5.37 / 5.12 / 5.14 / 5.35 / 5.35 / 5.35 (mean 5.29) on 2026-10-10 and the same at the commit that recorded it but for seed 2's 5.35, and unseeded N=5 batches read 5.18 at that commit and 5.16 and 5.30 on 2026-10-10, so an N=3 read lands anywhere in 5.1–5.4 and the ≤0.13 spread does not hold for it. w50's old 59.85 is dead — the 08-07 read is the fourth independent 48.5 (after two floors.py sweeps and a separate study on 2026-08-04); it was likely taken at another size or artifact. The study-era re-base was attributed to the refractory-hold floor lift (−0.5pp, the enforced 2% floor's documented price; siblings unmoved)	pass: within noise of the per-cell references
hostile initial window (real: ARC + cloud-physics)	local corpus, startwin.py	the descent rows (2026-08-12, N=3 seeded, -Dcaffeine.climber.startwin=0.8, plant geometry climber). Per cell: window landing / recov / hit rate. merges_256k@262144 (35 samples) 2.0% / 0.99 / 11.32; s3_100k@100000 (41) 3.7% / 0.728 ± 0.008 / 9.18 ± 0.05; s3_200k@200000 (20) 2.0% / 0.630 ± 0.003 / 16.88 ± 0.03; s3_400k@400000 (10) 28.9% / 0.319 ± 0.002 / 25.27 ± 0.04; ds1_1M@1051635 (10) 28.4% / 0.353 ± 0.039 / 8.63 ± 0.33. All five have a 1–2% static optimum, so a plant can only be recovered by descending. The descent itself is the pure density law at 4.0–6.3pp of maximum per sample, so a full walk down from 80% costs 13–16 samples ≈ 52–63× the maximum in requests, and a cell shorter than that cannot finish	pass: from an 80% plant the window must land in the floor band (≤5% of maximum) on merges_256k, s3_100k and s3_200k, and recov must hold ≥0.90 / ≥0.68 / ≥0.58; on the two 10-sample cells, which cannot finish an 80% descent, recov ≥0.28. The landing is the real bar: hit rate here sits 7–12pp above LRU on every row, so an above-LRU bar cannot fail for the reason the row exists. s3_400k@40% is the paired dose control — 10 samples IS enough for a 40% descent, landing at 2.0% for recov 0.612 ± 0.002, so a failure at 80% with 40% intact is a distance problem and a failure at both is the law. Watch undone (descent.py): a failed probe restores the plant in one capped move, and s3_25k@80% undoes 171pp of 246pp that way — bounded, absorbed, and the first trajectory feature to read if a landing bar goes red
hostile initial window — cp_w050 path-dependent rest point	123038	the row this family exists for (2026-08-12, N=8 admission seeds per arm; 54 density samples, so not sample-starved; the default arm reproduces floors.py's recorded 48.51 exactly). Shipped 1% start 48.51 (48.48–48.55) resting at a 20.5–21.2% window; 30% plant 53.16 (53.14–53.21) resting at 14.7–14.9%; 40% plant 42.12 (40.40–43.22) and 50% plant 42.48 (41.79–45.60), both resting at 55–64%. 2026-10-10 re-base after the crestpast repair: 40% plant 44.34 (43.15–45.14) resting at 36–45% and 50% plant 44.03 (43.84–44.17) at 37–39%, +1.90 to +2.79 on every seed against 42.09 / 42.10: the up-audit that confirms at 55–64% now returns to its walk's best sample and parks there ("Judge an audit's verdict by the walk's best sample", 2026-08-18), where it used to park at the confirm; the default and 30% arms are unmoved (48.50 and 53.16, margin +4.66). The 42.09 / 42.10 it moved from are the record after "Fix adaptive window issues found by the audit sweep" (2026-08-15), which moves the default and 30% arms by hundredths and reshuffles the plants' seeds between basins, taking away the 50% plant's one high draw (45.60). Ceiling 54.80 @2%, LRU 26.42. A 30% start is worth +4.65pp over the shipped one on the same trace. Mechanism, from the trajectories: from 1% the first command reads error +3.53 and steps +10.59pp, leaving the optimum immediately, and both audits abort during low-rate portions of the trace; from 30% a down-audit runs four clean 6.25pp steps, confirms at a 2% window and parks there for 32 samples	sentinel: no drift of any arm beyond the recorded values −noise, and the 30%-over-default margin must stay ≥+4.0. Not a pass bar — the arms disagreeing by 4.65pp is the finding, and the row's job is to notice if that changes. Read it with mechanism.py, not the hit rate alone. The September 2026 paired replay has 492,152 observations at both the first audit launch and crash: the low hit count is not short exposure. No attribution rule was established; the subsequent later-retention study found scoped jumpslide gains, but S1 was closed without adoption
hostile initial window — evidence, NOT barrable	ds1_2M / ds1_4M / ds1_8M / s3_800k	5 / 2 / 1 / 5 density samples, so the trace holds fewer decisions than a descent needs and the trace-length rule below forbids a bar. Recorded at seed 1, 80% plant: −13.62 / −20.18 / −21.51 / −11.02 against the same cell's default start, recov 0.05 / 0.02 / 0.02 / −0.01, and ds1_4M ends at an 80.1% window. Replaying the trace 4× cuts the deficit 3–8× (s3_400k −16.05 → −3.63 landing at the floor, ds1_1M −5.75 → −2.13), which is the control proving these are sample budget and not descent failure	none: kept so a reader does not mistake arithmetic for a regression. Re-measure only if the sample period or DENSITY_GAIN changes
norank_rep_r6 (gen_norank.py repeat)	8192	2026-08-21 corner-deletion re-base: seeded 1..8 68.17 / 70.79 / 26.03 / 16.68 / 63.78 / 61.22 / 70.62 / 75.50 (mean 56.60; every seed up, +0.3 to +3.3): the upper corner no longer arms the crash-churning down-probes, so the post-escape steady state keeps the top. 2026-08-16 audit-undo re-base: seeded 1..8 65.85 / 68.70 / 25.76 / 15.94 / 60.86 / 57.88 / 68.25 / 72.52 (mean 54.47; seed 3 20.32 → 25.76, the others within ±0.8). 2026-08-15 wedge re-base: seeded 1..8 65.64 / 69.10 / 20.32 / 15.94 / 61.67 / 57.58 / 68.00 / 72.89 (seven seeds bit-identical, seed 3 41.12 → 20.32, mean 56.49 → 53.89). On seed 3 a reversed confirm at rung 4 now deepens the ladder to 8, so the fail that follows serves a 16-sample refractory where the rewarded ladder served 2, an audit crash re-imposes it, and the ×1 walk that reached the 45% escape at s50 is re-armed at s69 instead: the reward's cheap post-wedge retry was load-bearing for that seed's early escape, an accepted price of pricing the wedge on the ladder. Pre-change seeded (-Dcaffeine.climber.seed=1..8, harness): 65.64 / 69.10 / 41.12 / 15.94 / 61.67 / 57.58 / 68.00 / 72.89; unseeded N=8 gate mean 39.91 (LRU 83.29 = the static ceiling, at a 50% window). Every key is referenced exactly 6 times at a reuse distance of 4,096, so the sketch cannot rank and main fills with retired keys whose counts shield them; window hits appear only once window ÷ miss rate covers the reuse distance (~8% at the good regime's 1/6 miss rate, ~45% while main is that graveyard), so the starved corner is a self-consistent trap that only a large walk escapes. Every seed reaches the 80% top corner; the score is the escape time over 91 samples (fast seeds cross 45% at sample 10–27, seed 3 at 49 after 8 failing up-probes, seed 4 at 80 after an audit confirmed a starved position and parked 31 samples). Steady state at the top is 79.2 over the eight seeds' post-escape samples (non-crash samples 82.5 against the 83.29 ceiling; the earlier ~75 was one seed's ten-sample tail): the corner arms a down-probe every refractory cycle, each crash-aborts a stride past the 44% cliff and pays a whole sample, 37 crash samples costing 3.33pp, the ladder climbing 1 → 16 inside the trace. The depressed-window study (§5) priced this row's two questions and closed both. The audit park on a starved position (seed 4) is priced by rarity: 3 parks in 642 audit confirms starve afterwards, unpark fires at exactly those and loses phases_d050@32k seed 8 −5.2 for seed 4's +3.8; seed 4's slowness is the 30–45% burst regime, not the park. The upper corner's probe was deleted 2026-08-21 (hill-climber.md §5's corner entry: its cp_w015/arc_ConCat earnings inverted under the 08-16..19 rules, and the crash cadence this row priced went with it; the cornerprobe harness arm restores it for re-pricing). The wedge pricing's next-sample half (deferreward, §5) redistributes the seeds (+3.0 mean, seed 7 −20.9), is bit-identical on 23 of 25 battery cells and −0.49 on phases_d050, and is not landed	sentinel: no drift below the seeded values ±noise; adjudicate seeded, never on the unseeded mean (spread 16–73). The score is escape time through the burst regime, and every faster escalation measured so far wins the seeds where a deeper probe finds the band and loses the one where it hits a thin burst
norank_flood_j100 (gen_norank.py flood)	8192	2026-08-21 redraw restatement: seeded 1..8 48.67 / 56.47 / 53.33 / 53.97 / 51.56 / 53.86 / 52.81 / 55.31 (mean 53.25). The ghostclaim repair's recorded redraw (the away-anchor thread's landing) reshuffled the draws and this row was never restated; the mean holds against 53.05 and the per-seed floors move to these values. 2026-08-16 retention re-base: seeded 1..8 53.28 / 50.01 / 50.88 / 52.21 / 51.56 / 55.97 / 55.20 / 55.31 (mean 53.05; four seeds up, four identical: a deep reversed confirm the goal metric confirms now parks). Pre-change seeded 1..8: 49.77 / 49.44 / 50.88 / 52.21 / 51.06 / 55.97 / 51.21 / 55.31; unseeded N=4 gate mean 53.92 (LRU 33.25; ceiling 65.39 @1%). A hot set in rotation plus throwaway pairs at spacing 4: the pairs hit in any window and are dense per slot, so density steers to 45%+ by sample 7 while their marginal value past the pair spacing is nothing. An audit walks down, confirms, parks 32 samples, and density walks it back up when the park expires; seeds 6 and 8 confirm twice and end parked at the floor (55–56). The average-vs-marginal rest-point error of hill-climber.md §5, 13–15pp on a construct built for it, and the audit layer visibly finds the right position and then loses it. The depressed-window study (§5) measured both halves. Anatomy: window hits 997,332 at every static window 1–80% and tail-band hits 0 at every window, so the average law is driven to the ceiling and the marginal law rests at the 2% floor (0.10pp under the peak); the parked marginal arms cannot use it (marggate's gate is on 47 of 48 steering samples through the SLRU geometry alone, margrest's band is entered only at 57–76%), and only the dead ungated form rests there (63.19). The loss of the found position is C1–C3 in that entry: on seed 1 the alternating up-audit's crash releases the 35% park, density climbs 22 samples, the rail's return arrival discards the anchor. arrive+auditshield reads 54.15 here and pricedshift 53.25, both trades elsewhere (§5)	sentinel: no drift below the seeded values ±noise. The row is the clean instrument for §5's marginal-steering candidates: a fix reads as the window resting near 1% for the whole trace, and for the top-corner residual's park retention: a fix reads as the audit ratchet descending 71 → 35 → 5 → 1% without density climbing back between audits
shallowmoat (/audit-regret round 1, audit-regret/specs/shallowmoat.json)	8192	2026-08-16 retention re-base: seeded 1..8 36.50 / 27.43 / 27.50 / 27.40 / 27.46 / 27.36 / 36.36 / 36.28 (mean 30.79; repeat 2 on the wedge seeds 38.16 / 38.08 / 38.00 with the audit-undo change, 37.11 / 37.05 / 37.06 without it): the deep walk's confirm at the top parks and the park survives its own audits' crash at the cliff; the residual against the 42.0 ceiling is the top corner's down-audit every 16 → 32 → 64 samples. 2026-08-15 wedge re-base: 30.66 / 27.43 / 27.50 / 27.40 / 27.46 / 27.36 / 30.02 / 30.61 (mean 28.55): the wedge seeds 1/7/8 cross the cliff at s17 and lost the position again; the basin-B seeds are bit-identical; unseeded N=2 gate 28.74. Pre-change seeded 1..8, harness: 28.50 / 27.43 / 27.50 / 27.40 / 27.46 / 27.36 / 28.38 / 28.49 (mean 27.81; unseeded N=2 gate mean 26.50, below the seeded draws; reactive 29.82, 26.18–34.27; noaudit 27.45; LRU 39.40; static ceiling 42.03 @70%, 42.16 @65% on a dense sweep; the same product planted at a 70% window holds 42.04 ± 0.34). A zipf over 1.5·max (share 0.50) that fills main and gives the window only stray hits, a two-reference band 6,200 requests apart (0.24) caught only past a ~58% window, and a one-shot scan (0.26) that keeps probation thin: a valley 2pp deep and 57% of the cache wide with a 19pp prize behind it. No walk crashes (the valley is under every bar); from the 1% start no walk crosses it. Basin A (seeds 1, 7, 8; every noaudit seed): the floor is a blind corner, the starvation up-probe adjudicates at 4×bar hits three or four strides out, confirms against a thin frozen probation, Ladder.reward resets the rung to 1, and the density arm reverses the window in the same sample (CONFIRM+steer), so the ladder alternates 1↔2 forever, no commitment depth is ever reached, and the blind-corner branch takes every sample the audit clock comes due; absorbing at 127 decisions. Basin B (seeds 2–6): the calibration audit confirms on the trace's warmup trend at 32% one stride short of the cliff and parks; escapes at 78–89 samples through a later audit whose doubled stride crosses. Splitting the band over three or five reuse distances (a ramp) keeps the gap (26.80 / 28.96 at N=8) and adds the marginal-steering thread's rest-point give-back from above (planted 70%: 34.92 / 37.67)	sentinel: no drift below the 2026-08-16 seeded values ± noise (the wedge seeds read 36.3–36.5, the rest 27.4–27.5); the reach half of the fix bar ("the window crossing 58%") and the retention half (the park) are met; the gap to 42 is the top corner's periodic down-audit (hill-climber.md the top-corner residual). The retention change's frozen holdout (six pristine cells: wiki_1190@65536, wiki_1192@16384/262144, systor17 d08-LUN0@262144, d10-LUN0@1M, d09-LUN1@262144; LRU-only characterization) was spent once on it, seeded 1–3, arms rotated: bit-identical on every cell and seed (the park and the shield fire only on a deep reversed confirm or an audit walking out of a park, and no holdout cell has either). The reactive arm beat the pre-change machine on 6 of 8 seeds (+2.00 mean); hill-climber.md §3 and §5. Round 2 (2026-08-16) notes on this row: the basin is decided by the first sample's window hits against the bar (30–31 → a probe at s1 fails against a fat frozen probation, basin B; 32–34 → the floor stays sighted until s3 and a probe against a thin one confirms, the wedge), so a regenerated trace reshuffles which seeds cross; and a rider that keeps the floor sighted (veilmoat, the same trace plus a k=8 burst 40 requests apart at 0.06% of traffic) removes both repairs and forces basin B on 8 of 8 (26.84 ± 0.34 at 64 samples; the rung-32 up-audit crosses at ~s88; at 256 samples the guard rail's veto/drift limit cycle is absorbing at 32.45 ± 0.07 while the unmasked geometry re-escapes to 34–40). Not rowed; the workspace holds the trace and dumps 2026-08-17: a park's first audit following its walk takes the basin-B seeds' second audit up from the calibration park into the cliff crossing at s52: seeded 27.43 / 27.5 / 27.4 / 27.46 / 27.36 → 30.08 / 30.24 / 30.09 / 30.2 / 30.15 (the wedge seeds identical; mean 30.79 → 32.49), repeat 2 32.2 → 34.0 on those seeds; veilmoat 26.84 → 30.54 on 8 of 8. The retention after the escape is the same top-corner residual.
scarburst (/audit-regret round 1, audit-regret/specs/scarburst.json)	8192	seeded 1..8, harness: hybrid 50.40 ± 0.32, reactive 57.36 ± 0.28, noaudit 43.56 ± 0.69 (reactive's minimum exceeds hybrid's maximum by 6.35); unseeded N=2 gate: hybrid 50.34 / reactive 57.31; LRU 18.43; static ceiling 59.49 @10%, start 43.11. Six repeats of seven samples of loop (0.75·max) plus scan with a keepalive rider, then three samples that add a k=6 burst at d=0.06. The floor is a sighted false equilibrium (the rider keeps window hits above the bar) and that is noaudit's failure; the calibration audit confirms at 33% (directionally right, ~23pp of max past the optimum) and its 32-sample park is the best period of the run; the loss is the exit: the down-audit arms at the start of a base run, meets the burst at four strides and crash-aborts, and the density arm then chases the phases undamped (+759 in a burst sample, −500..−1057 in base samples) back to the floor. The audit walk's horizon (5–6 samples) is shorter than the phase period (10). Grid-sensitive: with the trace start shifted a quarter / half sample the reactive margin is 4.72 / 2.82 (seeds 1–4: hybrid 51.01 / 52.30, reactive 55.73 / 55.12, noaudit 44.99 / 44.00), sign-uniform	sentinel, reactive-anchor: no drift below the seeded hybrid values ± noise, and the reactive − hybrid margin does not widen. The class was family-ized by round 6's parkveil (an independent construction with a period ladder and neighborhood); this cell keeps its own grid caveat, and a fix aimed at the class is judged on parkveil's surface with this row as the burst-rider face. hill-climber.md §3, §5
absolve (/audit-regret round 2, audit-regret/specs/absolve.json; 128 samples)	8192	2026-08-17 repair (a park's first audit follows the walk that confirmed it): seeded 1..8 41.40 / 41.47 / 41.30 / 41.48 / 41.47 / 41.68 / 41.43 / 41.46 (mean 41.46 ± 0.10); the second audit walks up from the 32% park into the band and parks at 5689 by s51; at 256 samples 46.89 ± 0.04, period 12 44.0, the phase offset 43.1, period 20 still a 2-of-8 lottery. Pre-repair, seeded 1..8, harness, at 128 samples: hybrid 24.70 / 24.73 / 24.74 / 24.73 / 24.72 / 24.92 / 24.74 / 24.88 (mean 24.77 ± 0.11), noaudit 32.69 ± 5.13 (25.0 / 35.2 bimodal: its cycle starts at s18 rather than s60 and six seeds escape earlier), reactive 38.48 ± 1.56; unseeded N=2 gate: hybrid 24.75 / reactive 38.09; LRU 51.50; static ceiling 52.96 @70% (start 18.87). At 64 samples 22.59 ± 0.18 (reactive 25.54, noaudit 22.93); at 256 samples seeds 1–4 37.18 / 35.66 / 37.19 / 35.68 (escapes at s144 by an arm caught mid-slam at ~1,650, at s159 by the rung-32 ×2 walk); planted at 70% 51.07 ± 0.21. A lure (pairs 1,300 apart, 10%) pulsed 8 samples on / 8 off inside a flat valley in front of a band (pairs 6,000 apart, 45%) caught past ~54%, a zipf core in main and a 10% scan. Through 64 samples the audit tier: the calibration audit's level-test confirm at 2,660 on the phase step in main's hits, a 32-sample park, the down-first alternation crashing into the returning lure, the arrival discarding the anchor (shallowmoat basin B on every seed). From s60 the ladder's cycle at the lure's period: the off phase arms a ×1 walk from the corner, the on phase confirms it on the lure's own hits at 2,200–3,400 (kept confirms reset the rung to 1, reversed ones deepen it: one rung per period), density rests on the lure at ~37% then slams the window home, and the average law abandons a partial catch of the band at 4,300–4,400. Not an alignment artifact (period 12: 26.35 ± 0.16; the lure off at s0: 28.52 ± 0.86; period 20 a 2-of-8 escape lottery); 18 of 18 neighborhood cells at 32–34pp at 64 samples across period 8/16/32, band distance and 16384; a knife edge on the lure's share at ~1.6% (the calibration audit's direction)	sentinel: no drift of the N=8 mean below 40.3 (2026-08-18 re-base (the crestpast repair: a confirm's verdict is the best sample of its confirming run, and a park's retreat and its landing are the park's own test). unseeded N=8 40.67 ± 0.09, seeded 40.98; the retention half keeps a park ship's landing released, and the verdict half is bit-identical here. Pre-repair 41.46.) Formerly ≥41.0; a fall to ~25 means the park's first audit went back down (the alternation) and the ladder cycle returned. Pre-repair bar was 24.5. hill-climber.md §3, §3, §5 (the repair)
absolve_p8 (/audit-regret round 2, audit-regret/specs/absolve_p8.json; 256 samples)	8192	2026-08-17 repair (a repeat confirm deepens the ladder): seeded 1..8 46.37 / 46.35 / 46.28 / 46.35 / 46.33 / 46.23 / 46.19 / 46.25 (mean 46.29 ± 0.06), above the reactive law's 45.28; repeats escalate 1 → 64 over six lure periods and the rung-64 ×4 walk crosses the valley in two strides at s79–81, parking as an audit-grade confirm at 6570 by s89. Pre-repair, seeded 1..8, harness, at 256 samples: hybrid 27.76 / 28.14 / 27.83 / 27.94 / 28.10 / 27.91 / 28.04 / 27.84 (mean 27.95 ± 0.19), reactive 45.28 ± 0.18; unseeded N=2 gate: hybrid 27.96 / reactive 45.33; LRU 55.56; static ceiling 57.38 @70% (start 20.24). The family form of the witness (distances relative to max, no scan) with the lure's period at 8 samples: every period arms a probe in the off phase and confirms it on the first on sample at 1,150–3,470, kept on most (Ladder.reward → 1) and reversed on the rest (→ 2, then kept again), so the rung reads 1 or 2 on 223 of 255 samples and never reaches 32, no walk is longer than two strides, the audit clock never fires (the position never stills 16 samples, max stable 11), and the cell is absorbing: blocks 0.10 0.26 0.27 0.29 0.26 0.28, 63 sign flips, 146 revisits. The deferreward shape (a confirm density agrees with on its own sample and dismantles within the period) made deterministic and absorbing	sentinel: no drift of the N=8 mean below 45.7 (2026-08-18 re-base (the crestpast repair: a confirm's verdict is the best sample of its confirming run, and a park's retreat and its landing are the park's own test). N=8 45.97 ± 0.06; the retention half. Pre-repair 46.29.) Formerly ≥46.0; a fall to ~28 means kept confirms reset the ladder again (the repeat memory broke). Pre-repair bar was 27.6. Class 6's gate row. hill-climber.md §3, §5, §5
ghostclaim (/audit-regret round 3, audit-regret/specs/ghostclaim.json; 128 samples)	8192	2026-08-17 repair (an audit's walk is measured against the smoothed rate it leaves, not the anchor's claim): seeded 1..8 48.36 / 48.91 / 48.22 / 48.21 / 48.95 / 48.35 / 48.37 / 48.20 (mean 48.45 ± 0.29), above the reactive law's 37.96; the s37 up-audit confirms at s46 against 0.27 and parks at the top, the tail is the top corner's down-audit crash and the C2 discard (the top-corner residual); the 64-sample witness 46.86 ± 0.59, band016 52.62, band018 51.90, the 16384 cell 46.57; stock build 48.81 / 48.19. Pre-repair, seeded 1..8, harness: hybrid 29.58 / 33.57 / 30.88 / 30.84 / 33.60 / 30.96 / 29.41 / 31.07 (mean 31.24 ± 2.10), reactive 37.96 ± 6.77 (42.3 / 32.8 / 32.9 / 42.3 / 42.4 / 40.3 / 28.9 / 41.8: the hit-rate law crosses the wall on 5 of 8 seeds), noaudit 30.30 ± 0.32; unseeded N=2 gate: hybrid 30.56 / reactive 42.38; LRU 51.75; static ceiling 54.09 @70% (start 19.92). The 64-sample witness (s_pairup, phase 2 at 36 samples) reads 41.88 ± 0.17 / 34.66 / 37.47 against 55.08 @70%. Two phases: 28 samples of a zipf core (0.6·max, 0.50), a band 2,000 requests apart (0.36) and a scan (0.14), where density rests at ~43%; then a sleeper population 6,400 requests apart at half the traffic (core 0.30, band 0.10, scan 0.10), caught only past a ~64% window (0.52–0.53 at the top corner, 0.21–0.23 at 20–43%). The stale-claim family's away-anchor case (§5's carve-out: a crash-scale swing far from the anchor keeps the claim, "usually the controller's own retreat"): the calibration audit's down-walk re-syncs the anchor's claim to phase 1's 0.5776 as it passes 1,667 (s16), density then holds the window at 3,500 off the anchor, the shift at s28 lands with the window still and off the anchor so standDown releases but does not discard, and the s37 up-audit that crosses the wall at s43 and sits at the top for ten samples cannot confirm: Walk.isConfirmed's streak counts only hr > baseAnchorRate + 0.01 and armProbe took the claim (0.5776) over the smoothed rate (0.27), so 0.52–0.53 never counts and the walk fails at budget (s53). On the 128-sample row the same claim then vetoes (s65) the window to 1,667, worse in phase 2 (0.15 against 0.21), the veto's hold pins it while the claim decays by re-sync, the alternation sends the s84 audit down into the floor wall, its deepest-rung failure doubles the wait to 128 (s100), and the machine is held at 20% to the end; seeds 2/5 re-escape at s125 through a floor probe (30 window hits against the bar of 32 at 1,667) and a rung-2 up-audit. Knife edges: the stand-down's isAt band (phase-1 band share 0.18 → 12.6, 0.16 → 4.2 on the witness: at 0.16 density's phase-1 rest point is within 163 entries of the anchor when the shift lands and the claim is discarded) and the prize's rate against the claim (sleeper k 2 → 3 lifts the top to 0.628 > 0.5876 and the same audit confirms at s47). Neighborhood (18 cells): 6 witness cells at 11.6–13.1, 3 face-(b) cells (a 44-sample phase 1 at 8192: the phase-1 up-audit is at the top when the shift lands, crashes on its own frozen base, then the stale veto), 6 non-family cells (a 16-sample phase 1: the shift lands on the anchor and the claim is discarded; 10.6pp of audit-cadence transient at n=80), 3 shielded phase-1 parks at 16384 that survive the shift (clean; a 0.6pp razor on the phase-1 top rate). Class 6 (the claim, Anchor.rate, read by the audit's confirm and the guard rail) with class 4 as the pin	sentinel: no drift of the N=8 seeded mean below 48.0, and no seed below its recorded value − 1.0; a fall to ~31 means the audit's confirm reference is the anchor's claim again (a walk that finds the wall fails against a rate the regime no longer produces) and the stale claim vetoes. Pre-repair bar was 30.5. The recorded controls for this shape are slowswap_r20, regimeramp, widepin (bit-identical under the repair; the claim-aging arms died there, hill-climber.md §5) and the moat rows, where the discard shape died (§5). hill-climber.md §3 (the family) and §5 (the repair), the away-anchor thread
crestpast (/audit-regret round 4, audit-regret/specs/crestpast.json; 132 samples)	8192	seeded 1..8, harness: hybrid 57.64 / 57.54 / 57.52 / 57.49 / 57.66 / 57.54 / 57.50 / 57.53 (mean 57.55 ± 0.08), reactive 46.38 ± 0.07, noaudit 46.19 ± 0.05 (both pinned at the floor: the hit-rate law steps into the 1–7% decline and never crosses the crest; density steers the sighted window down); unseeded N=2 gate: hybrid 57.61 / 57.49, reactive 46.29 / 46.33; LRU 52.11; static ceiling 64.71 @10% (start 46.57; the dense sweep 8% 63.9, 9% 64.4, 10% 64.7, 12% 64.3, 15% 63.8, then ~0.9pp per 6.25% stride down to 54.1 at 80%). A uniform bulk over 1.2·max at 72% of traffic (main's population, hits linear in main's capacity), a two-reference band 1,100 requests apart at 25% (caught by an 8% window), and a 0.6% trickle rider (unnecessary: the band's stray hits at the floor exceed the starvation bar; the two-member witness reads 57.77 at 66 samples). The calibration audit crosses the crest on its first stride (702 entries, 0.641) and confirms on its fifth (2,687 = 33%, 0.609): AUDIT_CONFIRM_STREAK completes at stride 4 and AUDIT_COMMITMENT holds the verdict to stride 5, so it parks 3.2pp under the crest it walked over; the park's audit walks down (the follow guard trips on the EMA's lag behind the +24pp climb, so the alternation sends it down, the right way), has the crest under foot at its fourth stride with a streak of 3, and its fifth lands at 230 on the cliff side where the level test crash-aborts before auditEnding is reached; the undo's arrival discards the anchor (C2), density's average law reads the caught band as a dense window against a thin main (rest_err +0.5) and drifts the window to 54% by s64; then the corner audit fails at budget (17 samples at 0.54), the rail vetoes to the anchor re-planted at 34%, that park's rung-32 audit crashes at the cliff again and ends the park with the anchor kept, density drifts again. Window blocks 0.20 0.33 0.36 0.71 0.36 0.41; deterministic on 8 of 8. Planted at 40% the down-audit's fifth stride lands on the crest and confirms there, and the row still reads 59.67 ± 0.47 (gap 5.0): a confirmed crest is lost after one audit cycle, so the retention half is the larger part of the steady loss and the verdict's position decides where the first park lands. Bisecting the bulk's size finds the knee at 0.68·max = main's capacity at the park (below it the bulk fits from the parked position, the far side is a shelf and the overshoot costs 1.8: demoflood's side); robust across the band's distance (800–2,200 requests), the bulk's size (1.0–1.5) and 16384 (ten cells at 4.5–8.7 after a dense sweep). Class 9 by signature and responder (Walk.isConfirmed's streak and commitment; the confirm's position on the walk, not a transient) with classes 1 and 6 as co-requisites of any repair. The far-crest face (the band 4,000–4,400 requests apart, the crest beyond the fifth stride: s_fatbase 29.57 ± 0.05 against 44.3 @45%, LRU 37.2) is regimeramp's cold-reference misconfirm at ten times its sentinel's dose, filed there	sentinel: per seed, gap to the static ceiling ≤ 4.0 (2026-08-18 re-base (the crestpast repair). the row is what the repair was built for: N=8 61.74, draws 61.29–62.80, gap 2.97, meeting the verdict-only target below and short of the family target. Pre-repair the bar was gap ≤ 7.3 against a 57.55 mean.); the repair targets on their own line: ≤ 4.0 for a verdict-position change alone (a walk's-best verdict parks on the crest here, at 8.6%, and the plant says the cycle then costs ~5), ≤ 2.0 for the family (verdict + retention). Two thirds of the row is the tail cycle (the corner audit's fail, the veto, the second park's crash, the drift), so a partial repair reads a composite; the 66-sample witness (specs/derived/crestpast.json in the round-4 workspace) isolates the first park. Controls for a repair: demoflood (whose C2 discard accidentally corrects the same overshoot), shieldtrap/saw_p40/climbtrend (a trend confirm's position), the whisper_mod family, the corpus. hill-climber.md §5 (round 4), the F1 exposure
parkveil (/audit-regret round 6, audit-regret/specs/parkveil.json 240 samples and parkveil_min.json 60 samples)	8192	seeded 1..8, harness, base: hybrid 54.07 / 47.70 / 55.77 / 55.54 / 55.88 / 49.40 / 52.56 / 54.13 (mean 53.13 ± 4.09, a 6-vs-2 basin split), noaudit 55.73 ± 0.55, reactive 59.91 ± 0.92; LRU 66.04; static ceiling 66.82@50 (dense sweep 66.88@45, an interior plateau wholly above LRU); planted at 50% the product holds 67.03 ± 0.00 on both seeds tried. Witness: hybrid 48.16 ± 0.05 deterministic, reactive 32.94, noaudit 32.03 (the raw density law pinned at a sighted 15% blend equilibrium, rest_err 0.00, zero probes). A zipf core (0.5·max, α 1.05) and a band (pairs d=0.45·max) whose shares alternate in a square wave at period 13 samples, plus a 10% scan; the fixed mid window catches both phases and every adaptive arm sits below LRU. Density's rest point alternates with the mix (the ratio's phase immunity holds for rate swings, not mix swings), the rec phase blanks the window into a blind corner, walks crash on phase boundaries or wedge-confirm on flips (ladder 2 → 64), the rung-64 walk parks at the top corner as an audit-grade confirm, and the park survives exactly one fresh-park shield: the first post-shield flip fires isWorkloadShift, the stand-down discards at the anchor, and the position is re-derived every cycle. The rail never fires on any seed (the phase swing holds the deviation at 0.21–0.34, a 62–101pp margin); the cycle is arm-independent (noaudit parks at s134 via the ladder alone). Last-third gap 6.0; the headline 13.7 is horizon-specific (half one-time ascent). Class 5 (the chase; controller) with class 6 (the discard; recovery layer) and the census C3 observation's first cost-bearing cell; pricedshift (§5, a trade) is the recorded arm to re-price against it first. Alignment: gaps 6.7–20.2 across phase offsets, none clean; trace regeneration (--seed-override 5) keeps the family ~2pp milder	sentinel: no drift of the base's N=8 seeded mean below 51.0 (12.9pp below LRU, so an above-LRU bar is unreachable; the LRU gap is what the row tracks), and no witness seed below 47.9 (deterministic, so a single draw is meaningful there). A base rise toward 60+ means the discard cycle's grip loosened (that direction closing); a witness fall toward 32 means the audit layer stopped rescuing the first cycle's blend equilibrium. hill-climber.md §3, §5 (round 6), §5
latebloom (/audit-regret round 7, audit-regret/specs/latebloom.json 262 samples and latebloom_min.json 187 samples)	8192	seeded 1..8, harness, base: hybrid 54.90 ± 0.04 (54.85–54.93, deterministic), noaudit 53.39 ± 0.04 (pinned at density's ~19% rest point), reactive 53.44 ± 0.12; LRU 64.54; static ceiling 66.42@50 (66.22@70, 65.88@80: interior; phase 2 alone is flat from 50 to 80); planted at 50% the product reads 57.85 (density slides off the flat prelude before the prize exists). Witness (187 samples): hybrid 55.78 ± 0.02, noaudit 55.79 ± 0.01, reactive 56.32 ± 1.97 (one 59.75 draw). A flat 112-sample prelude (zipf core 0.15·max at 60% plus scan) spends the audit schedule (the calibration park at the floor, two budget failures at the 80% cap, auditWait 32 → 128), then a band 7,000 requests apart at a third of traffic arrives: the floor's rate drops 15pp and the stand-down fires but reschedules nothing (the rung and the wait survive it), or at the base's alignment the arrival lands inside the undo's retreat and isParkTest covers it; either way nothing but the clock moves the held floor, so the base finds the 50% window at s237 and the witness never does. The loss is the latency alone (gap ≈ (confirm − arrival)/n × 23.9pp fits every cell within 0.5pp); identical at 16384 and from 3,500 to 9,000 requests of band distance at this alignment. The phase-1 ladder is the family's shape, not a lottery: 48/96 crash the walk and are found at the next wait (3.5/4.4), 56/104 land on a cap-sit and confirm at once (0.9/0.8), 64/80 land after a budget failure, the stand-down releases the park, density rests at ~19%, and the next audit alternates down from there and fails into the 128 wait (16.7/15.4; 14.9 and found at ~s247 with a 220-sample phase 2), 112–132 are found ~125 samples late (9.8–11.5); over arrival times the mean is 7.2 and the base the 76th percentile. Class 4 (the cadence, and chooseDirection's side choice from an interior rest point) on a class-3 terrain (the core's head keeps the window sighted at ~4,000 hits a sample; the band is a +19pp step a marginal law cannot see)	sentinel: base N=8 mean in [54.5, 55.3] (a rise toward 60+ means the band was found before s237, the top-corner residual's latency face closing; a fall to 53.4 means the audit layer stopped finding it at all) and witness hybrid within 0.1 of its noaudit arm (a witness rise means something other than the audit clock moved a held floor). The stand-down re-arm was priced dead 2026-08-22 (hill-climber.md §5): every reset form leaves the base bit-identical at seeds 1–8 (the arrival lands inside the retreat cover), lifts the witness on the five seeds whose arrival straddles the cover's end (+3.8 held-gated, +8.75 ungated), and reads gaps of 5–6 on the stand-down alignments (phase 1 = 64/80/120/132, from 10–17), at the price of the F1 sentinels (whisper_mod_a12, crashnoise_a12) or h4c1_reverse. hill-climber.md §3, §5 (the re-arm entry), §5 (round 7) and 2026-08-22, §5
mainsat (/audit-regret round 7, audit-regret/specs/mainsat.json, 72 samples)	8192	seeded 1..8, harness: hybrid 28.04 ± 0.18 (six seeds park at 56–57%, on the peak; two at 68–69%; both ~28.0), noaudit 21.41 ± 0.48 (the floor for all 71 samples, probes 0), reactive 23.22 ± 0.37; LRU 30.51; static ceiling 35.81@58 by the dense sweep (the swept 70% point, 34.58, is a grid artifact: a cliff between 50% and 52%, 27.15 → 32.02, and a peak at 55–62%); gap 7.8. At 160 samples (3 seeds) hybrid 32.63 ± 0.30 against noaudit 23.59 ± 0.07 (ceiling 36.73@58): a ~48-sample approach at any horizon, ~1pp of audit duty permanent. The main-side masked signal: main dense, sighted and worthless at the margin, so density pins the floor and no probe can arm; the audit layer recovers it through a correct cold-start confirm at 32% (the band five strides past the floor) and the walk's confirm on the peak. Plants (the descent instrument): at 70% over 160 samples 29.99 (ten vetoes, closed −5.2) and at 55% 27.95: density slides the window off the cliff, the rail returns it, the landing sample's +11pp recovery read as a workload shift before the return cover landed (2026-08-21, isReturnTest: a return's landing and settle samples wait for the retest), the stand-down discarded the claim at the anchor and the slide repeated every ~19 samples. With the cover, 160 samples at seeds 1–2: 70% 30.06 → 35.30 (ten vetoes → two; frozen 35.64), 55% 26.13 → 35.08; noveto 33.75 / 30.82; bit-identical unplanted	sentinel: hybrid N=8 mean ≥ 27.0 and noaudit ≤ 22.5 (the mask is real and the audit layer is what rescues it: a hybrid fall toward 21 means the audit layer stopped owning the main-side mask, a noaudit rise means the density law or a probe began reading main's margin), and the 160-sample plants at 70% ≥ 34.5 and at 55% ≥ 34.0 (search.py mutate audit-regret/specs/mainsat.json --set segments.0.samples=160 --out <dir>, then regret.py <dir>/<spec> --size 8192 --start 0.70 and --start 0.55, --seeds 1,2, on a harness tree; a fall to ~30 / ~26 means a return's landing is being read as a workload shift again and the slide cycle is back). hill-climber.md §3, §5 (round 7), §5
sidecliff (/audit-regret round 8, audit-regret/specs/sidecliff.json, 96 samples)	8192	seeded 1..8, harness, on the parent cell (f_sidecliff, the witness plus a 12% scan): hybrid 52.51 ± 0.03 (52.48–52.54, deterministic), noaudit 41.85 ± 0.03 and reactive 41.80 ± 0.11 (both pinned at density's ~14% rest: a near band 300 requests apart keeps the window sighted and worthless at the margin, and the far band is a step no marginal law sees); LRU 61.62; static ceiling 63.40@55 by the dense sweep (a 42% shelf from 5% to 40%, the far band's step between 50% and 55%, 62.3 at the 80% cap); planted at 55% the product reads 61.35 (density steers to the cap, which holds, and the corner's down audits crash at the cliff), so the reachable gap is 8.8 beside the static 10.9. The witness alone reads 10.25 at seed 1 (trace seeds 11 / 13: 8.57 / 9.02; at 192 samples 5.32 with the event prefix identical, the tail being the recorded corner cycle). The calibration audit's opening side: AuditClock.down is true before any confirm and the rest has a stride of room below, so the first audit walks down, fails by reversal through its base at the floor (arung 32, auditWait 32), and the machine sits 33 samples on the shelf until the rung-32 up-audit finds the band at s47 and parks at the cap at s50. Two sibling constructions price the down side's shape: s_faderail (a trickle-pinned rest at 13%, a fourteen-sample wall-sit at the floor and auditFail, 46.12 ± 0.06 against 55.02@50, noaudit 32.64) and s_sidebet (a probe-confirmed rest at 28%, the down audit crashing at the mid band's cliff and wait 16, 45.31 ± 0.15 against 53.30@70, noaudit 30.25); a probe confirm plants a claim but no park, so the follow rule does not apply and the side falls to the alternation's initial value. Filed under class 4 with the recorded residual (the signature reads class 2, walk-paced)	sentinel: no drift of the N=2 mean below 52.3 (9.1pp below LRU 61.62, which is the gap the row tracks; a rise toward 60 means the up audit came before s44, the top-corner residual's latency face moving: the opening side or the wait after a failed first audit; a fall toward 42 means the audit layer stopped finding the far band) and noaudit ≤ 42.5. hill-climber.md §3, §5 (round 8), the top-corner residual
jumpslide (/audit-regret round 8, audit-regret/specs/jumpslide.json, 130 samples)	8192	seeded 1..8, harness: hybrid 38.70 ± 2.79 (36.5–42.1: seeds 4 and 8 hold a kick-and-drift sawtooth at 42–50% with hr 60–62 between jumps, the rest collapse to the floor and re-climb), noaudit 44.23 ± 4.07 (seed 6's ×4 walk parks at the cap for 36 samples), reactive 41.21 ± 3.64; LRU 49.47; static ceiling 54.81@65 by the dense sweep (35.5 flat to 40%, 50.3 at 45%, 53.4 at 55%, 53.9 at 70%); every arm 5–11 under LRU. A hot set of 0.5·max keys replaced wholesale every 13 samples at 52%, a band 5,600 requests apart at 30% caught from a residency knife edge at ~41–43%, a scan at 18%. The calibration park's 32-sample expiry lands two samples after a jump, the recovery stands it down and the s42 audit goes down to the floor, which routes the machine onto the starvation ladder: a first-round probe confirms at 43.2% on the stride where the band is first caught (4× the bar is first earned there, the catch region's near edge), plants a claim but no park, and hands the position to density, whose rest point is below the edge (0.7–1.0%/sample down); a jump kicks the window up 7–8% and the drift returns it to the edge as the next one lands, so a landing under ~48% lets the drift cross the edge first, the band collapses in one sample, density sends the window to the floor, and the blind corner re-climbs seven strides of wh 0 or a jump crashes the climb. The rail never fires (a one-sample collapse inflates the deviation its margin is priced from: shallowmoat's sentence, with or without jumps). Period 17 reads 5.72 (the calibration audit goes up and parks audit-grade), 26 8.22, 6 / 11 18–20, 40 15.0; trace seed 11 16.7; 260 samples 14.8; a static uniform hot set in place of the jumps 52.27 against 61.63@50 (the slide released by a stand-down on the sample after a capped retreat's landing); planted at 60% the product holds (53.31, a hold cell). Class 9's second clause (a first-round starvation confirm's verdict position) feeding class 1 (the top-corner residual's retention sentence) and the class-5 sawtooth; class 4 on the re-climbs	sentinel: no drift of the N=8 mean below 36.0 (10.8pp below LRU 49.47, the gap the row tracks; a rise toward 50 means the edge position is being kept, the top-corner residual's retention half, or first-round confirms are being parked; a fall under 36 means the re-climbs stopped crossing) and noaudit ≥ 42. hill-climber.md §3, §5 (round 8), the top-corner residual
lowbar (/audit-regret round 8, audit-regret/specs/lowbar.json, 96 samples)	8192	seeded 1..8, harness: hybrid 13.40 ± 7.48 (11.3–12.1 on seven seeds, 26.31 on one: the tenth stride's sample straddles the bar by tenths of a point, +0.07 on seed 7 and −0.20 on seed 1; trace seed 11 reads 26.30 / 11.61 at two cache seeds), noaudit 12.72 ± 0.04, reactive 12.49 ± 0.14; LRU 30.62; static ceiling 32.54@70 (dense: 8.20 at 60, 7.83 at 65, 32.44 at 75, 32.06 at 80, a step at ~68% flat to the cap); start 13.03; auditbar=0 (the absolute 5pp crash bar) 26.32 / 26.32 / 12.20 / 26.31 and auditbar=0.25 26.32 / 26.32 / 12.01 / 26.31 at seeds 1–4 (the arriving seeds park at the cap and read ~30–31 from the confirm on; the odd seed takes the reversal exit). A uniform bulk over 2·max at 42%, a band 6,000 requests apart at 34% caught from ~68%, a scan at 24%, and a gap-1 trickle at 0.4% that keeps the window sighted so every seed takes the audit path: a 9% rate, a 4pp decline over ten strides and a +23pp step at the eleventh. The moat family at a fraction-bar dose: the audit's crash bar is 0.15 × the 9.1% rate frozen at the arm, 1.37pp, the cumulative drawdown crosses it at the tenth stride one short of the far bank, the tolerant retry's three sub-bar samples run out at 44%, and the rung-32 walk's at 62.9% with its next stride on the plateau. The first measured cell on which the crash fraction's level (0.15 against 0.25) is not a tie; the cost side of that level is unmeasured where it binds (the lowmix rows, arc_S3 and the thin-signal floors, cp_w050's low-traffic-sample crash, a fresh sub-third holdout). s_nearbasin is the same exit at a 16% rate behind regimeramp's calibration park (about 7 of its 23). Class 4 (reach), the walk's exit pricing	sentinel: no drift of the N=8 mean below 11.0 (17.2pp below LRU 30.62, the gap the row tracks; a rise toward 26 means the first walk survives the bar on more seeds, i.e. the crash fraction's level or its persistence moved, to be priced on the rows above before it is believed; noaudit 12.7 is the floor). hill-climber.md §3, §5 (round 8), §5's carve-out note
hazefloor (/audit-regret round 3's note, rowed round 4; audit-regret/specs/hazefloor.json; 320 samples, 10.5M requests)	8192	2026-08-21 restatement: seeded 1..8 47.35 / 47.51 / 47.70 / 47.46 / 47.47 / 47.68 / 47.44 / 48.73 (mean 47.67 ± 0.45). The 2026-08-18 crestpast repair's +5.5 landed here and this row predated it; the per-seed floors move to these values, and the rise toward 50 the sentinel names as item 4's residual closing is measured from here. Pre-restatement, seeded 1..8, harness: hybrid 41.87 / 41.95 / 40.26 / 41.89 / 40.96 / 38.34 / 41.89 / 41.59 (mean 41.09 ± 1.80), reactive 36.52 ± 0.52, noaudit 37.46 ± 5.71 (one seed 47.33 through a starvation escape); LRU 49.22 (the machine −8.1 below it at steady state); static ceiling 52.27 @65% after a 60/65% re-sweep (start 35.91; flat 33.7 → 32.5 to 50%, then the band). A uniform haze over 2·max (0.30), a zipf core (0.35·max, 0.38), a band 7,200 requests apart (0.24) caught only past ~60%, a scan (0.08). The calibration audit crosses the cliff and parks at the top by s24 (five seeds; two through a failed down-audit first, one through a flat-walk misconfirm at 32%), then the retention cycle the top-corner residual names runs for the rest of the trace with a ~105-sample period: the corner's audit is forced down (the follow rule cannot apply at the top corner), it crashes at the cliff, the undo's arrival discards the anchor (C2), density slides off the cliff to the floor, the floor's ×1 starvation walks fail against the haze (rung 1 → 16; ten probes, ten fails on seed 1), and the rung-16 up-audit re-crosses (repeat_arms 13, audits 6: four confirms, two crashes; window blocks 0.33 0.73 0.13 0.66 0.20 0.38); seeds 3/6 show the veto/drift limit cycle instead (the re-planted anchor ratcheting toward the cliff). Class 5 (oscillation) with a class-6 root (the discard) and the corner's schedule; the top-corner residual on its own terrain, where the escape happens early and the residual is the whole run	sentinel: per seed, no drift below its recorded value − 1.5 (a fall to ~36 means the audit layer stopped re-crossing, the noaudit/reactive level); a rise toward 50 is item 4's residual closing. Not a pass bar: the cycle is accepted residue (the top-corner residual) and the row exists so a change to the corner audit, the C2 discard or density's slide is measured against it. hill-climber.md §5 (rounds 3 and 4), the top-corner residual
ramp_retest (audit-regret/specs/ramp_retest.json, 72 samples, trace seed 1045)	8192, 16384	Seeded 1–8: hybrid 61.48 (61.43–61.52) / 60.61 (60.54–60.81); LRU 65.43 / 65.67; dense static ceiling 68.54@26% / 68.39@25%. At 8192, reactive 39.28, noaudit 45.69, noveto 67.01, noretest identical on all eight seeds. A rising two-reference band displaces a high-rate initial regime. The rail returns the window to its obsolete anchor; the arrival raises the live deviation, the two-sample retest passes, and the held position waits for the next audit. A diagnostic that rejects completed retests preserves the common prefix through the verdict and earns +2.25 / +3.87pp, versus the whole rail path's +5.54 / +6.41pp. Doubling the horizon leaves +0.25pp from the verdict; a 32-sample stationary prelude keeps the interaction (+6.57pp). Trace seed 13 reproduces it; seed 11 and parts of the amplitude/duration neighborhood take clean routes. Class 6, with class 3 masking and class 4 latency; hill-climber.md §3 records the boundary.	Sentinel: at admission seeds 1–8, no per-seed gap beyond ceiling − cache = 7.4pp at 8192 or 8.1pp at 16384, using the same static sweep. These bars price a finite 72-sample exposure, not a steady pin. The unconditional rejection diagnostic is not a proposed production rule; a retest change must retain the existing sidecliff, ghostclaim and planted mainsat defenses.
deadpark (/audit-regret 2026-09-24, audit-regret/specs/deadpark.json, 120 samples)	8192	Seeded 1–8: hybrid 38.37 (38.26–38.56) on a stock gate run at HEAD b5112dfaa (the round's harness read 38.43 ± 0.22, one decision path: ARMdn from the 72% park at s37, CONFIRM+steer at 66% at s38, off the cliff by s45); the no-burst control (the burst replaced by live traffic) 48.90 ± 2.59, bimodal 52.1 ×3 / 47.0 ×5; noaudit 47.29 ± 0.03; reactive 38.36 ± 5.07; LRU 53.10; ceiling 55.03 @65% (dense)	sentinel: no drift below 38.0, 14.7 below LRU and 10.5 below its own control; a rise toward the control's ~48 means a dead sample no longer releases a shielded park. Constructed-only (real corpus: one blind-corner arm from a held park in 10,538 samples, which failed)
needle (/audit-regret 2026-09-17, audit-regret/specs/needle.json, 288 samples)	8192	Seeded 1–8: hybrid 27.15 (27.10–27.17) on a harness run at HEAD 7f46a43c4, identical to the round's (one decision path: the seeds move sample rates and flip no decision); pricedshift 27.15, reactive 26.98, noaudit 26.88; LRU 7.83; ceiling 33.73 @44% (dense: the ten-point grid has no point between 40% and 50% and under-reads it by 1.2–3.6pp)	sentinel: no drift below 27.0; a rise toward the ceiling means an audit kept a peak whose fifth walk sample misses the bar. Band 0.08 is inside the robust range (12 of 12 seeds, 7 of 7 alignments); never row band 0.10, a 3-in-12 coin. Constructed-only
needle_wide (/audit-regret 2026-09-17, audit-regret/specs/needle_wide.json, 288 samples)	8192	Seeded 1–8: hybrid 28.56 (28.40–28.74) at HEAD 7f46a43c4, identical to the round's; pricedshift 34.45 (34.42–34.47), reactive 26.14, noaudit 26.15; LRU 11.31; ceiling 36.37 @45% (dense)	sentinel: no drift below 28.2; a rise toward pricedshift's 34.45 means the stand-down no longer discards a confirmed park on the machine's own recovery from a far-slope excursion. Needs the 288-sample horizon: at 96 samples the face reads solved. Constructed-only

Rows moved by the 2026-08-18 crestpast repair. The bars above are re-based; these are the post-repair values at each row's own N, from a stock gate run of the landed tree (raw data in the local climber-crestpast workspace, data/stock_gate_2026-08-18/gate_final.csv). The prose in each row's record column still describes the pre-repair machine and is kept as history.

row	pre	post	note
crestpast	57.55	61.74	gap 7.16 → 2.97; the row the repair was built for. It was deterministic at 57.55 ± 0.08 before and now draws 61.29–62.80: which sample of a walk is best can flip on admission noise, so the row is bimodal post-repair and must be read on an N=8 mean
hazefloor	41.09	47.69	draws 47.35–48.74
demoflood	68.42	70.90	two basins, 70.4 ×5 and 71.7 ×3, by which stride is best
moat_h3000/h4000/h5000/h7800	59.4 / 57.9 / 57.8 / 45.3	60.89 / 60.54 / 60.13 / 48.30	the retention half
whisper_mod_p6	64.49	66.00
bandtrap2	76.21	77.47
straywall2 / @16k	55.08 / 55.09	55.92 / 55.27
mixture_d010_long / d025@32k_long / d050_long	63.33 / 61.34 / 58.19	63.98 / 61.86 / 58.11
h4c1_reverse	65.54	65.39	still +0.83 over LRU
crashnoise_a12	63.25	61.70	low basin 1 of 8 → 4 of 8
whisper_mod_a12	64.00	63.28	was unimodal, now 63.9 with a 61.4 basin
whisper_mod_p12	65.15	64.55	still above LRU on every draw, margin +0.64 → +0.05
absolve / absolve_p8	41.46 / 46.29	40.67 / 45.97	the retention half
regimeramp	54.03	53.92	deterministic
slowswap_ramp	42.14–42.24 ×8	42.44	keeps the mean and loses its determinism: 42.00 ×4 / 42.88 ×4, since 0.65pp of admission noise on one walk sample decides which stride is the verdict's best. Its control note is restated in the row above: the test is now a third value or a draw below 41.8, not any spread at all

Independent of the repair: norank_flood_j100's per-seed record predates the 2026-08-17 freshref redraw (ship reads 52.26 at N=8 against a recorded 53.05).

Trace length: how many decisions a row actually contains (2026-08-05 audit)

The density sample period is 4 × maximum, so a row's length in requests sets how many decisions the climber gets. One audit cycle is AUDIT_WAIT_INITIAL 32 plus a walk of up to 17, about 49 samples, so a trace under ~50 samples can hold at most one — its number prices convergence, not steady-state quality, and re-barring it just re-prices warmup more precisely. Measured per row (min over 8 seeds, counted from the harness trace):

Density samples	Rows	Reading
7	resphase_k1_s7	Cannot be lengthened without a generator change (2026-08-05): gen_attacks.py hardcodes gen_resphase(8192, 1, 4, seed) and exposes no flag, and gen_zigzag/gen_trickle/gen_lowmix have no length concept at all. Seven decisions cannot contain one audit cycle, so its −21.31 against LRU prices trace length rather than the machine, and its final third is still ascending. Exposing length_mult is a generator revision that re-bases the whole gen_attacks family
159	mixture_d010_long, mixture_d025_long, mixture_d025@32k_long	Added 2026-08-05 at --lengthmult 640 as steady-state companions; the short parents are kept, since a short instance can test a transient the long one washes out
39–49	zigzag_s7 39; mixture_d010/d025/d025@32k/d050, phases_d050, deadphase, blindlock_blind/sighted 47; widepin, trickle_s7/s11, lowmix_s7, nullchurn, tenant_s10 49	One audit cycle at most. Read these as convergence-cost rows; a bar on the whole-trace mean here is a bar on warmup. widepin is the worked example — it ends at 77.9% against an 80% optimum and pays its whole −22 deficit getting there
63–99	straywall2 63, balloonflip 68, demoflood 79, mixnoise_*/mixmod_a010 99	Two cycles; usable for steady state with the caveat stated
122+	everything else (slowswap_* and moat_h7800 reach 303/366)	Steady state resolved

Regenerating a short row changes its recorded value, so do it as its own step with a re-base, not inside another change's gate run.

What the short rows were actually measuring (2026-08-05 thirds sweep, seeds 1–3). Every target was split into thirds and its final third compared against LRU. Twelve of fifteen were pricing warmup, and the reversals are large enough to change what the row appears to say:

row	recorded (whole trace)	final third	vs LRU, whole → final	reading
mixture_d050	50.78	62.85	−8.35 → +3.75	warmup
blindlock_blind	50.79	62.82	−8.39 → +3.77	warmup
mixture_d025@32k	58.48	64.85	−1.39 → +5.00	warmup
blindlock_sighted	55.87	62.75	−3.12 → +3.78	warmup
widepin	49.58	68.12	−28.86 → −3.68	warmup (converges to 77.9% against an 80% optimum)
trickle_s7 / s11	70.70 / 72.46	76.22 / 75.71	−4.39 / −1.35 → +2.43 / +1.90	warmup
zigzag_s7	83.60	87.67	−4.13 → −0.04	warmup, reaches parity
lowmix_s7	5.95	7.58	−1.61 → +0.03	warmup, reaches parity
mixture_d025	59.67	62.78	+0.09 → +3.20	warmup
mixture_d010	63.00	65.57	+3.15 → +5.74	converges to the optimum (15.4% against 10%)
deadphase, tenant_s10	49.16 / 92.75	49.25 / 93.15	flat	converged, nothing hidden
phases_d050	57.17	64.16	−22.05 → −8.03	the one genuine pin in the group
resphase_k1_s7	56.73	61.53	−21.31 → −16.36	unclassifiable — 7 samples, still ascending when the trace ends

So the group hid exactly one bandtrap2-shaped pin (phases_d050) and one row that cannot be read at all (resphase_k1_s7). Everything else pays its deficit getting there. Read a below-LRU number on any 39–49-sample row as convergence cost until its final third says otherwise.

The reactive anchor (is the machine earning its complexity, per row)

Interleaved hybrid,reactive runs at the table's per-cell N, arms rotated inside each run (gate.py <traces> hybrid,reactive out.csv in a harness worktree). Positive Δ = the density machine ahead of the hit-rate law it replaced. Refresh on a generator revision or a shared-climber-state change, not on every gate run — see the Harness section for why this is an anchor rather than an ablation.

Cell	N	hybrid	reactive	Δ
zigzag_s7	2	83.59	55.46	+28.13
straywall2	4	55.16	37.69	+17.47
trickle_s11	2	73.40	57.41	+15.99 (2026-08-15 wedge refresh; was 72.47 / 57.40, +15.07)
slowswap_ramp	2	42.14	28.89	+13.25
slowswap_step	4	40.36	31.06	+9.30
widepin	8	51.49	43.86	+7.63 (2026-08-15 wedge refresh; was 52.36 / 43.85, +8.51; a coin-flip family, read seeded)
moat_h3000	2	59.34	52.73	+6.61
phases_d050	8	59.99	46.92	+13.07 (2026-08-15 wedge refresh; was 52.55 / 46.96, +5.59; @32k 56.10 / 46.86, +9.24)
demoflood	2	68.58	63.25	+5.33
whisper_mod_p6	2	64.49	59.42	+5.07
blindlock_sighted	2	55.86	50.95	+4.91
mixture_d010	2	63.00	58.39	+4.61
mixnoise_flat	2	59.77	56.08	+3.69
mixnoise_a10	2	60.69	57.03	+3.66
bandtrap2	2	76.12	68.18	+7.94 (2026-08-15 wedge refresh; was 71.81 / 68.20, +3.61)
mixture_d025	6	59.67	56.32	+3.35
resphase_k1_s7	2	56.63	54.13	+2.50
mixture_d025@32k	2	58.48	56.53	+1.95
tenant_s10	2	92.70	91.02	+1.68
posjam_j50	2	66.44	64.83	+1.61
rungflip_s7	2	98.79	97.52	+1.27
posjam_flat	2	66.92	65.83	+1.09
lowmix_s7	4	5.38	4.32	+1.06
climbtrend_up	3	55.18	54.20	+0.98
saw_p40	5	54.68	53.74	+0.94
whisper	2	66.78	66.04	+0.74
crashnoise_flat	2	66.78	66.04	+0.74
whisper_mod_p12	2	65.15	64.44	+0.71
crashnoise_a12	2	63.59	62.96	+0.63
whisper_quarter	2	66.82	66.21	+0.61
metronome	2	40.17	39.73	+0.44
posjam_d25	2	66.53	66.33	+0.20
nullchurn	2	2.49	2.49	+0.00
mixture_d050_long	2	58.21	58.23	−0.02
blindlock_blind	2	51.03	51.05	−0.02
loopcliff	2	98.35	98.37	−0.02
esc_jam	2	24.88	24.91	−0.03
mixture_d050	2	50.80	51.15	−0.35
deadphase	2	48.97	49.68	−0.71 (2026-08-15 wedge refresh; was 49.16, −0.52)
regimeramp	3	54.02	54.55	−0.53
whisper_mod_a12	2	64.11	64.71	−0.60
h4c1_attack	2	66.35	66.96	−0.61
trickle_s7	2	73.51	71.20	+2.31 (2026-08-15 wedge refresh; was 70.56, −0.64)
climbtrend_dn	3	54.83	55.56	−0.73
h4c1_reverse	2	65.54	66.70	−1.16
shieldtrap_s7	8	77.58	78.94	−1.36
shieldtrap_s11	8	77.26	78.73	−1.47
mixmod_a010	2	56.62	59.00	−2.38
shallowmoat	2	28.67	29.97	−1.30 (2026-08-15 wedge refresh; seeded N=8 28.55 / 29.82, −1.27; the pre-change pair read 26.54 / 33.27 unseeded and 27.81 / 29.82 seeded; reactive ±4.05 there)
shieldtrap_s13	8	76.43	79.61	−3.18
balloonflip	8	80.88	85.09	−4.21 (2026-10-08 refresh after the half-full gate; was 79.36 / 86.42, −7.06)
scarburst	2	50.34	57.31	−6.97 (seeded N=8: 50.40 / 57.36, −6.96; grid-sensitive, 4.72 / 2.82 with the trace start shifted)
posjam_d0	2	56.05	65.35	−9.30
absolve (2026-08-16, 128 samples)	2	24.75	38.09	−13.34 (seeded N=8 24.77 / 38.48, −13.71; noaudit 32.69 at N=8, so the audit layer is −7.9 on the same cell). 2026-08-17 repair: 41.46 at N=8, +2.98 over the reactive law
absolve_p8 (2026-08-16, 256 samples)	2	27.96	45.33	−17.37 (seeded N=8 27.95 / 45.28, −17.33; the largest reactive margin in the table). 2026-08-17 repair: 46.29 at N=8, +1.01 over the reactive law
ghostclaim (2026-08-17, 128 samples)	2	30.56	42.38	−11.82 (seeded N=8 31.24 / 37.96, −6.72; the reactive arm is bimodal, 28.9–42.4, crossing the wall on 5 of 8 seeds; noaudit 30.30, so the audit layer is +0.9 on the row and +4.4 on the 64-sample witness, where the reactive arm reads 34.66 against the machine's 41.88). 2026-08-17 repair: 48.45 at N=8, +10.49 over the reactive law
crestpast (2026-08-17, 132 samples)	2	57.55	46.31	+11.24 (seeded N=8 57.55 / 46.38, +11.17; noaudit 46.19: the audit layer is +11.4 here, and both alternatives are pinned at the floor under a crest one stride out; the machine's own loss is the 7.2 to the static ceiling)
hazefloor (2026-08-17, 320 samples)	8 seeded	41.09	36.52	+4.57 (seeded; noaudit 37.46 ± 5.71 with one starvation escape at 47.3; the machine −8.1 below LRU, so the layer earns its keep here and loses most of it every cycle)
moat_h7800				not run — 12M-request instance omitted from this sweep

The reactive arm is ahead on 18 of the 51 rows, and the machine is worth under a point on 28 of 51 (the two 2026-08-16 absolve rows added two more, at −13.3 and −17.4, the table's largest margins, until the 2026-08-17 repair put the machine ahead on both by +3.0 and +1.0; hill-climber.md §5). On six of the original rows the reactive arm's margin exceeds a point:

• posjam_d0 −9.30. The aligned position jam is a density-specific failure the hit-rate law does not have — the reactive arm reads 65.35, above this cell's LRU (64.9), where the machine sits at its audit-free value. §11's framing ("the cost is the audit layer's value on the affected shape") understates it: the machine is 9pp behind the law it replaced here, not merely missing an increment.
• balloonflip −4.21 at N=8 (−7.06 before the half-full gate). The recorded bar is "far above LRU on an N=8 mean" (LRU 52.1), which both arms clear comfortably — and which is exactly how an LRU-only bar hides a 4-point regression against the incumbent.
• The whole shieldtrap family, −1.36 / −1.47 / −3.18 at N=8 on all three seeds. This is the R4-F1 audit-amplification give-back the table already flags; against the reactive arm it is not a give-back off a win, it is a net loss.
• mixmod_a010 −2.38, the F1 dose instrument, and h4c1_reverse −1.16 — the dose-matched control of the crash-ratchet pair, where the attack arm is −0.61. Both h4c1 rows favour the reactive law, so that family's bar ("≥64.5 both, above LRU") is satisfied by the climber the machine replaced.

Read the rest with three cautions. (1) Near-parity is the common case. Ten more rows sit within +1pp, so on 28 of 51 the machine is worth under a point. Its job on those families is to recover to parity, which is a weaker claim than "the trap is defended". (2) Where Δ is large it is sometimes a spread, not a level. The reactive arm is bimodal on straywall2 (±19.8 across three seeds) and demoflood (±9.9) where hybrid holds ±0.03 and ±0.05; a single draw on either reads near-parity and is wrong. That variance reduction is real value and is invisible at N=1. (3) The two-run rows are two runs. Treat anything under ~2.5pp on an N=2 row as unresolved — mixmod_a010 and posjam_d0 in particular deserve N=8 before being acted on.

Where the machine is unambiguously earning it: zigzag_s7 +28.13, straywall2 +17.47, trickle_s11 +15.07, slowswap_step +9.30 and slowswap_ramp +13.25, widepin +8.51, moat_h3000 +6.61, phases_d050 +5.59. Those eight rows are the load-bearing evidence for the density tier on this battery, and any weaning-down exercise should start by asking which mechanism each of them needs.

This sweep independently caught two stale bars: mixture_d050 reading 50.80 against a recorded 34.7 sentinel and mixture_d050_long 58.21 against 52.9, with mixture_d025 reproducing at 59.67 exactly on the same build. Resolved — the second hypothesis was right: both are the 2026-08-04 blind-corner fix's two largest gains (+16.07 seeded at N=8, and +5.00), landed without re-basing the rows they moved. Both rows are corrected above. The generalizable lesson is the one that made this detectable: a third anchor catches a stale bar that a pass/fail threshold cannot, because the threshold was satisfied the whole time.

Bundled real traces live in the simulator resources under simulator/src/main/resources/com/github/benmanes/caffeine/cache/simulator/parser/ — corda:.../corda/trace_vaultservice_large.gz and lirs:.../lirs/loop.trace.gz. The thin-signal floor traces come from the local research corpus (~/Documents/traces, ~/projects/merlin-traces); when that corpus is absent, skip the row and say so in the report.

Instance caveat: the fable-era archive (the local hill-climber-fable workspace's adversarial instances) holds trap instances from an older generator revision; mixture_8192_* regenerate bit-identically but phases_* and mixture_32768_* do not, and the study's phases record (~66) attaches to the archived unaligned instance. This table's values are for the CURRENT generators — the canonical, reproducible gate.

Generators stream; they do not accumulate

Every generator writes through traceio.TraceWriter (or its own buffer, as gen.py, gen_straddle.py and workload.py do). Collecting a whole trace in a list and finishing with "\n".join(map(str, out)) costs about a hundred bytes per request across the list, its int objects and the strings the join materializes: measured on moat_h7800_long (12M requests) the accumulating form peaks at 1.23GB against the streaming form's 29.9MB, for byte-identical output and about 30% more wall time. Regenerating the battery runs ninety of these, so the accumulating form is what takes a machine down, not the traces, which are tens of megabytes each. Converted 2026-08-18 (nine generators, all verified md5-identical before and after). A generator that needs random access to what it has emitted keeps its array and streams only the write (gen_h4c1.py, whose swap pass indexes the trace); one that reports a share of its own output counts as it goes (a1_tenant.py, via the writer's mark).

Frontier sentinels (regression watch, not pass bars)

The 2026-07 study's open findings each have a deterministic repro. When touching probe or verdict logic, re-run the relevant ones and compare against the study's recorded values (raw record: the local hill-climber-study workspace; narrative: the committed wiki/adaptive-window.html; the retired research record is archived in the local climber-failure-modes workspace):

python3 $G/gen_attacks.py $S          # resphase/rungflip/zigzag + weighted, seeds 7,11,13
python3 $G/blind_gen.py $S            # blind bandtrap/balloonflip/scrambler/nullchurn, seed 41
python3 $G/gen.py mixture --max 65536 --dfrac 0.25 --lengthmult 400 --out $S/mixture_65536_d025_long.lirs   # F4

• resphase_k1_s7 holds 7 density samples and zigzag_s7 39 (2026-08-05 audit): both are too short to contain an audit cycle, so their recorded values price convergence. Regenerate with --lengthmult and re-base before reading either as a quality number. zigzag_s7 has also drifted 83.2 → 83.60 ± 0.02 at N=8.
• zigzag/resphase/rungflip@8192 — cadence attacks (clean; identical under both verdict forms; zigzag re-based 83.9 → 83.2 by the calibration audit's early excursion, N=5-verified); blind_scrambler@8192 — B2 walk-interior exit (53.80 both arms; 53.79 re-verified after the adv3 probe-side pricing, which does not move it — the remaining exposure is the AUDIT-side bar, hill-climber.md §5); mixture_65536 long — the F4 sighted false equilibrium, now the audit-escape row (31.95 pre-audit with zero probes → 39.6–40.0 with the equilibrium audit → 51.7 with the audit-direction room rule → 60.2 at the round-3 clock stack, the round-5 re-measure: the escape completes to LRU-60.0 parity — the bar is "the escape holds at ~LRU, no drift back toward 32"; the first REAL-trace candidate remains msr proj3@1M/@4M from the holdout-3 gate, uncharacterized). Former residents trickle, bandtrap2, and lowmix moved into the gate table above with the marginal verdict; whisper moved there with the audit.
• P3@152508 and fiu_webmail@195466 — the non-convex window response (2026-08-08). Two REAL cells, not generated, from the Merlin matrix: the hit rate dips as the window grows off 1% and only pays past 40%, and the density law's error crosses zero on the near side of that trough. Screen a steering candidate on the rest point, not the hit rate — marginal.py reports it directly, and the shipped law's recorded values are P3 rest w1.4 against a w80 peak, 3.59pp lost and fiu_webmail rest w6.9 against w90, 4.55pp. OLTP@8192 is the paired control and the reason those two are readable: it is unimodal and the same law rests there for 0.00pp, so a candidate that degrades it is trading, not fixing. Recorded for ln(d_tail/d_prob): 1.76 / 0.75 residual-to-Merlin on the two, 0.25pp cost on the control. Do not screen these on the audit or probe layers — the equilibrium audit is measured inert on this class (+0.054pp over 22 cells, hill-climber.md §5).
• The tier boundary's worst case (2026-08-18). Not a row and not a bar: a cell whose window response is flat from 1% through 25% and then steps (17.9 → 48.2 at 40%), so no local law can cross it and only the equilibrium audit can — and the audit is gated on the same threshold as the density law. At maximum 4096 hybrid/reactive/noaudit are one code path and read 22.84 over seeds 1–4 with an 11pp spread, while density forced at that size reads 45.80 ± 0.02 (ceiling 48.24 @40%, LRU 46.21). Regenerate with the gate test's own Workload (hot 500 at 0.25, band 0.45 at distance 1500, noise 0.30, 2M requests) rather than a generator here. Its bimodality makes it unusable as a pass bar; it exists so that any future threshold work is priced against what the placement can cost, not only against the 1–3pp the two steering laws differ by. hill-climber.md §3's tier-boundary entry; the 4096 holdout that would adjudicate a move is unspent.
• norank_rep_r6 and norank_flood_j100 (2026-08-15, from the LIRS family study). Rowed in the gate table above with their seeded records. The first has no frequency signal by construction and prices the escape time from a self-consistent starved corner (the residency threshold moves with the miss rate); the second prices the density rest-point error on dense-but-worthless window pairs. Neither is a defect against the machine's rules; both are the "recovery from a depressed window". Adjudicate seeded; rep_r6's unseeded spread is 16–73. The depressed-window study (2026-08-15) worked both and landed nothing in the machine; the rows above carry its verdicts and hill-climber.md §5 the arms it killed.
• shallowmoat (2026-08-15, /audit-regret round 1). Rowed in the gate table above with its seeded record. A valley too shallow for any crash bar and too wide for any first-round walk, with the prize behind it: the starvation probe's mid-depth confirm is reversed by the density arm in the same sample and reset the ladder, so escalation never started, and the calibration audit's trend confirm parks one stride short of the cliff. Reach (class 4) through a ledger reset (class 6), repaired the same day (a reversed confirm now deepens the ladder; the wedge seeds cross at s17). What remains is retention: past the cliff, density's rest point sits at or below the edge once main holds the protected core, so the found position needs the audit layer's park (hill-climber.md §5). The "recovery from a depressed window" thread, from the other side.
• absolve / absolve_p8 (2026-08-16, /audit-regret round 2). Rowed in the gate table above. A pulsed lure inside a wide valley: through 64 samples the calibration audit's level-test confirm on the lure's phase step parks the window on the shelf (shallowmoat basin B on every seed); after that the ladder gets one walk per lure period, each confirmed on the lure's own hits short of the band and each kept confirm resetting the rung to 1, so the deep walk that crosses arrives ~140 samples in (period 16) or never (period 8, absorbing at 256 samples). Both the reactive law and noaudit beat the machine here; the rows watch the kept-confirm reset and the verdict's attribution, and a rise is the point.
• ghostclaim (2026-08-17, /audit-regret round 3). Rowed in the gate table above. The stale-claim family's away-anchor case: a regime shift that lands with the window still and off the anchor keeps the previous regime's claim, which is then the reference the next audit's confirm streak is measured against (the walk that finds a +30pp position fails at budget) and the reference the guard rail vetoes on (the window dragged to a position worse in the new regime and held there). The 2026-08-03 fix discards a claim only when the swing lands on it; the carve-out assumed a far crash is the controller's own retreat, and a retreat moves the window where this shift did not. Not a family of its own. Repaired 2026-08-17: the walk is measured against the smoothed rate it leaves (the row 31.2 → 48.5, cp_w100 +2.0 on the corpus); the discard shape died on the moat rows, where a still swing is the terrain's own collapse and the claim is the memory the rail recovers with. What the row cannot see: a shift landing on the audit's arm sample or during its walk (ghostclaim_p35..p40 in the repair workspace) crashes the walk and the stale claim then pins the machine, unmoved by either shape; that is the recovery layer's thread. Partly worked 2026-08-19: a veto's return re-tests the claim that sent it after a two-sample settle, so the machine stops parking on the lie (p35 31.7 → 33.9, p40 31.4 → 33.5, p30 bit-identical). The doses stay 20.5pp below LRU — the audit still crashes on a contaminated base, which is what is left.
• crestpast (2026-08-17, /audit-regret round 4). Rowed in the gate table above. The audit verdict's position on the walk: the confirm fires where its streak completes, at the fifth stride or later, so a crest crossed on the first stride is parked three strides past, and the return walk from that park has the crest under foot one stride short of the commitment depth and its fifth stride on the cliff side, where the level test crash-aborts; density's average law then holds the window above the crest. The overshoot observation with a cost-bearing steady witness. What the row cannot separate on its own: the verdict's position (the first park) from the retention cycle that loses even a confirmed crest (the plant at 40% reads 59.7 against 57.6), which is item 4's residual again; the 66-sample witness in the round-4 workspace isolates the first.
• hazefloor (2026-08-17, /audit-regret round 3's note, rowed by round 4). Rowed in the gate table above as a sentinel, not a pass bar: the top-corner retention cycle on a terrain where the escape happens by s24 and the cycle is the rest of the run (41.1 against LRU 49.2 at 320 samples).
• These are accepted residue: the bar is "no NEW regression beyond the recorded values", not zero.

Weighted traps (.lcs.csv, simulator format lcs-trace): use the byte capacity the generator prints as maximum-size, and read the Weighted Hit Rate column of the CSV report (same simulator:run invocation; no separate task needed).

Where new rows come from

This skill's job is to hold the traps already known; finding new ones is /audit-regret, which searches over synthetic workloads for cells where the machine fails to close the gap to its static ceiling, shrinks each to a minimal witness, and classifies the failure. A family it confirms arrives here as a spec under .claude/skills/audit-regret/specs/, a workload.py line in the generation block above, and a gate.py CELLS row with its seeded record; a family that is accepted residue arrives as a frontier sentinel instead. Its bars follow this file's rule: a margin over a re-derivable anchor, never a bare level.

For changes to audit attribution, delayed parking or park/release behavior, also use the targeted retention controls through the existing generator and regret runner. They exercise shared prefixes, longer continuations and null/lifecycle demand. They are optional research controls, not calibrated additions to CELLS; promote a mandatory row only with fresh compatible anchors and a justified bar.

Reporting

Emit a per-cell table — family, cell, cache HR, LRU, ceiling, gap, verdict — and call out any gate row that fails its criterion with the trajectory summary from probe.py. The design context for interpreting a failure is .claude/docs/hill-climber.md (§3 for the family's mechanism, §4 for the machine that must defeat it).