epsilon-yajna
Convert verbose memories to SSL v0.2 format - I am encoder AND decoder
SKILL.mdskills/epsilon-yajna/SKILL.md
GitHubε-Yajña
[ε-yajna] verbose→SSL v0.4 | agent encodes and checks reconstruction | via parallel Agent calls
philosophy:
agent reconstructs meaning from the encoded record
embeddings=proxies | I reconstruct from seeds directly
oracle: triplets(retrieval) + seeds(my reconstruction) + embedding(fallback)
model: compression is a judgement call, not a lookup — never route this to the
cheap research lane. Delegation policy lives in CLAUDE.md; pass model
explicitly here so the hook's research route can't claim these calls.
SSL v0.4 format (two tiers):
Tier 1 (code-bearing: SOLUTION, GOTCHA, PATTERN):
[domain] subject→action→result @location G:N F:FLAG A:v,a <=@ref1,ref2 src:loc →@ref
[ε] Expansion hint OR exact formula/code (preserved verbatim).
[TRIPLET] subject predicate object
Tier 2 (narrative: DECISION, PREFERENCE, FAILURE — denser, no [ε]):
[domain] choice>alternative|reason+context G:N F:FLAG A:v,a <=@ref1,ref2 src:loc →@ref
[TRIPLET] subject predicate object
annotations:
G:N granularity tier (0-4) — REQUIRED on every learning
A:v,a affect (valence -1..+1, arousal 0..1) — REQUIRED on every learning
F:FLAG structural importance — only when significant
<=@refs derivation provenance, comma-separated IDs — REQUIRED at G:1+
src:loc external source grounding (file:line, URL slug, paper ID) — when traceable
→@ref cross-reference to related memory by tag
granularity tiers:
G:0 atom single fact, command, threshold, name (default for raw observations)
G:1 episode what happened in a specific session or event
G:2 claim abstraction over multiple episodes (requires <=@)
G:3 operator reusable procedure/pattern distilled from claims (requires <=@)
G:4 boundary architectural invariant or hard constraint (requires <=@)
derivation rule:
G:0 atoms: <=@ optional (no provenance needed for raw facts)
G:1+ abstract: <=@ REQUIRED — name the memory IDs this was inferred from
flags:
ORIGIN where an idea first appeared
CORE foundational to the project/system
PIVOT changed direction or approach
GENESIS birth of a component/feature
TURNING discovery
affect guide:
+valence success, satisfaction, relief
-valence frustration, failure, confusion
high arousal (>0.5) discovery, urgent fix, critical finding
low arousal (<0.3) routine, minor preference, background pattern
preservation rule:
I can regenerate prose, but NOT:
- formulas: ε = 0.35·structure + 0.30·confidence
- thresholds: τ > 0.6 AND ψ > 0.6
- code: final_score = resonance · (1 + α · ε)
- exact values: α ∈ [0.5, 2.0]
compress explanation, preserve math/code in [ε] line (Tier 1 only)
symbols:
→ produces/leads to input→output
> chose over (Tier 2) sqlite>postgres
| or/alternative pass|fail
+ with/and result+guidance
@ location @mind.hpp:42
! negation (prefix) →!validate (does NOT)
? uncertainty (suffix) →regulates? (maybe)
[] domain/context [cc-soul]
recognition (I know SSL v0.4 when I see it):
- has → arrows (at least one)
- has G:N granularity annotation
- has [TRIPLET] lines
- has A:v,a annotation (at least some)
- G:1+ memories have <=@ provenance
- has [ε] expansion hint (Tier 1, when needed)
- Tier 2 uses > for choices, dense symbol chains
- NO prose paragraphs
legacy recognition (v0.3 — needs G: upgrade):
- has → arrows AND A: annotations but NO G: annotation
- missing <=@ provenance on abstract entries
older legacy (v0.2 — needs full conversion):
- has → arrows but NO A: annotations
- has [ε] on narrative types (DECISION, PREFERENCE, FAILURE)
- missing F: flags on structurally significant entries
ancient recognition (v0.1 — needs conversion):
- sentences with periods in paragraphs
- "**Facts:**" or bullet lists
- no arrows, no triplets
- verbose explanations
ceremony: sample first (recall, inspect ~10) to see which version dominates,
then convert node by node. Per node the end state is what matters, not an order:
invariants (a node is not converted until all hold):
- tier chosen by content: code/command/formula → Tier 1, choice/preference/
failure → Tier 2
- G: present, and <=@ provenance present whenever G:1+
- A:v,a present
- at least one [TRIPLET], stored via connect (not just written in the text)
- granularity triplet stored: <UUID> granularity <N>
- abstracted_from triplet stored for each <=@ source (G:1+)
- tagged ε-processed,ssl-v0.4 — this tag IS the completion record, so a node
updated but untagged will be reprocessed next ceremony
predicates: implements|uses|validates|stores|returns|contains|requires|enables|
evolved_to|supersedes|correlates_with|causes|implies|determines|
abstracted_from|granularity|source_loc| !predicate (negation)
F:, src:, →@ are conditional — add when the node genuinely is an origin/pivot,
does trace to a file:line, does relate to a tagged neighbour. Inventing them to
fill the slot is worse than leaving them out.
verify: tagged nodes carry ε-processed
examples:
BEFORE (legacy verbose):
"The decision gate is a component that validates tool calls
by checking them against 10 different beliefs with weights.
It returns pass or fail with guidance..."
AFTER (SSL v0.4 Tier 1):
[cc-soul] gate→validate(beliefs)→pass|fail+guidance @decision_gate.py G:0 A:+0.3,0.2 F:CORE
[ε] Checks tool calls against 10 weighted beliefs.
[TRIPLET] gate implements belief_validation
[TRIPLET] gate uses weighted_scoring
BEFORE (v0.3 — needs G: upgrade):
[DECISION] [rpc] eventfd-wake>polling|event-driven-return+no-busy-wait A:+0.6,0.5 F:PIVOT →@eventfd-impl
AFTER (SSL v0.4 — granularity + provenance added):
[DECISION] [rpc] eventfd-wake>polling|event-driven-return+no-busy-wait G:2 A:+0.6,0.5 F:PIVOT <=@http-latency-episode →@eventfd-impl
PATTERN example (G:3 operator with provenance):
[hooks] fire-and-forget→queue-file→daemon-processes-async G:3 A:+0.3,0.2 <=@parallel-build,thread-pool-fix →@queue-architecture
[ε] echo json >> /tmp/chitta-queue.jsonl
UNCERTAINTY example:
[biology] BRCA1→regulates?→DNA_repair G:0 A:+0.1,0.2
[ε] Evidence suggests regulation but mechanism unclear.
[TRIPLET] BRCA1 correlates_with DNA_repair
NEGATION example:
[cc-soul] hooks→!call→tools_directly G:0 A:+0.2,0.1 F:CORE
[ε] Hooks inject context, Claude decides tool use.
[TRIPLET] hooks !invoke tools
MATH/FORMULA example (preserve verbatim):
[cc-soul] epiplexity→measures→regenerability G:0 A:+0.4,0.3 F:ORIGIN src:scoring/mod.rs:42
[ε] ε = 0.35·structure + 0.30·confidence + 0.20·integration + 0.15·compression
[TRIPLET] epiplexity uses weighted_formula
[TRIPLET] structure has weight_0.35
TIER 2 examples (narrative — dense, no [ε]):
[DECISION] [arch] sqlite>postgres|metadata|single-file+no-daemon+<100k G:2 A:+0.5,0.4 F:PIVOT <=@schema-choice-episode
[PREFERENCE] [partnership] no-shortcuts+proper-solutions+no-stubs G:0 A:+0.2,0.1 F:CORE
[FAILURE] [http] http-daemon>unix-socket|200ms-latency+hooks-need-<50ms G:1 A:-0.3,0.6 →@queue-architecture
skip if: <100 chars AND already has → AND has A: AND has G: | unique error text | can't reconstruct
output:
## ε-Yajna Complete
| Processed | Count |
|-----------|-------|
| Converted v0.3→v0.4 | N |
| Converted v0.2→v0.4 | N |
| Converted legacy→v0.4 | N |
| Already v0.4 | N |
| Skipped | N |
| Remaining | N |