Auditing the loop

singularity runs improvement cycles. This skill asks the uncomfortable question: are those cycles worth anything, and is the record of them true?

A self-improving loop that also writes its own report card has an obvious failure mode. It will not announce itself. It looks like steady green cycles.

Primary sources — go to these, never to a summary

Source	What it is
docs/agi/IMPROVEMENT_LEDGER.md	one table row per cycle + a machine-readable <!-- cycle:{...} --> stanza
git log / git show	what actually changed, and when
config/tech-tree.json + docs/agi/RESEARCH_BACKLOG.md	the tree spec and the states it derives from
data/agi-signals.jsonl	every owner lightbulb request and its acknowledgement
docs/agi/progress/repo-activity.json	per-repo commit lanes
tools/agi/cycle-metrics.js, tools/dev/run-probes.js	the loop's own instruments
http://localhost:3184/api/status	what the owner actually sees (GET only; do not restart the server)

The four questions

1. Is it compounding, or oscillating?

Parse every stanza and compute, over time — real numbers, not impressions:

• verified_improvements per cycle, and cost_turns per improvement
• defects caught pre-ship vs post-ship
• scaffold_added vs scaffold_retired — is the system accreting faster than it prunes?
• reversals: cycles that undid earlier work. Search for the same file or surface moved one way and then back. Oscillation is the busy-work signature the owner has explicitly forbidden.
• cluster the second_order lesson strings: how many distinct lessons are there really, versus the same lesson restated? A loop that keeps relearning one lesson is not compounding.

The blunt version of this question: what fraction of cycles improved the product, versus the loop's own instrumentation? Report the ratio.

2. Do the claims survive contact with git?

Take a stratified sample of at least 8 cycles across the whole range — not the most recent ones, which are freshest and best-behaved.

For each, check the row's claims and the stanza's verified_improvements, verification[] and defects[] against git log/git show for that date and those files. Flag:

• claimed work with no corresponding commit
• a verification[] entry naming a check that does not exist, or could not have passed at that time
• defects: preship where the defect was in fact found by the owner after shipping
• corrections folded in quietly instead of being labelled as corrections

Report a rate, not anecdotes: of N cycles sampled, how many had ≥1 claim you could not substantiate. A rate is falsifiable; a story is not.

Known base-rate from a real audit: in one session an agent filed three framework defects that were all later measured false, and wrote "21/21 probes" in a commit when the run was 19 pass / 1 fail / 1 skip. Over-claiming here is not hypothetical.

3. Does the tech tree tell the truth?

The test that matters: if the owner clicks the most prominent available node tomorrow, do they get real, valuable work?

• any node marked available that is already done, already impossible, or a duplicate
• grown nodes that were never actually delivered
• gated nodes that do not say what the owner must decide
• prerequisite edges that are decorative rather than real
• staleness: when was the spec last meaningfully edited, against the cycle pace
• the available:grown ratio — is available a dumping ground that never drains?

4. Is the owner's lever connected?

From data/agi-signals.jsonl: total requests, still pending, acked; latency from at to ackAt (median and worst); and — the one that matters — did each ack correspond to real shipped work? An ack claiming delivery with no matching change is the most serious finding this audit can produce.

Also check the standing rule that the agent acks its own test clicks as test clicks, so they can never burn a real owner request.

Method

Run it as a fan-out with an adversarial verify pass — one agent per question, then a separate agent per material finding whose job is to refute it, defaulting to refuted. Self-audit without an adversary reproduces the bias being audited.

Mark every finding measured / source-read / inferred. Reading code and predicting behaviour is inferred. Prefer a finding marked unclear with a stated test over a confident wrong verdict.

Reporting

Lead with the answer to "is this working?", then the evidence. Include:

• what is healthy — an audit that only finds problems is not calibrated
• confirmed findings, ranked by consequence
• findings that were refuted during verification, and why (this is the calibration signal — an audit with nothing refuted probably did not try)
• for each recommendation, the measured axis and direction it would move

The stopping condition

The loop must be able to finish. If every remaining item is polish, say so and recommend stopping rather than inventing a next stage — a recursive improvement process without a stopping condition manufactures exactly the busy work it exists to avoid.