moth-runner

P0 #4: the adversarial runner. Campaigns run moth-cells hunter genomes over moth-corpus terrain under a homeostatic admission throttle. The product is what the runner saw: witness.jsonl — append-only, chain-sealed testimony.

A witness who rewrites testimony is a liar. witness.jsonl only appends; every row carries FNV-1a-64 row/chain hashes over canonical JSON and verify re-derives the whole chain. The witness records:

• admissions — every genome slot decision, granted or refused, always with a reason. No hunter runs without a witnessed admission.
• throttle decisions — window moves, with before/after and reason.
• throttle holds — refusals (event=throttle_hold) explaining why the window did NOT move. A hold without a reason is a rumor.
• campaign results — honest counts re-derived from the walks: moves, findings, refusals, deaths, dormancies, decoys resisted.

The throttle consumes ONLY the witness's own re-derived signals (quality_signals()), never campaign claims. Policy, deterministic:

• window ∈ [1, 8], moves by at most ±1 per decision
• expand iff ≥2 campaigns completed AND chain verifies AND starvation < 100 deaths/1k moves
• contract iff starvation ≥ 300/1k, or chain broken
• otherwise hold — witnessed as a refusal with the reason

Expansion on a guess is rounding up; the throttle refuses to round.

The boredom-honesty doctrine test here caught a latent kernel crash in moth-cells: event.get("reason", event["type"]) evaluates the default eagerly, so walk-level dormancy booking crashed on any event lacking type. Unit-tested green, path unreachable. Fixed in moth-cells PR #1 (refusal_row reason fallback) with a walk-level regression test. This is why the runner exists: doctrines tested at campaign level catch what module tests cannot see.

moth-ledger    envelopes + chain law + verify            (main @ e95c786)
moth-corpus    CorpusIndex: repos → receipted surface     (PR #1)
moth-honest    the evaluator: planted truth, honest cost  (PR #1)
moth-cells     kernel 1: cellular predation              (PR #1)
moth-runner    campaigns + witness + throttle            (this repo)

Dependency: moth-cells (git pin to the PR branch until Casey merges, then repin to @main). Receipts, not dependencies, for everything else.

pip install -e .   # pulls the pinned moth-cells
moth-runner run --name probe --corpus corpus.jsonl \
    --genomes genomes.json --ticks 40 --probe-at 5 --witness witness.jsonl
moth-runner verify witness.jsonl
moth-runner signals witness.jsonl

genomes.json: {"genomes": [{"seed": 161, "move_weights": [49152, 16384], "attention_bias": 8192}]}

examples/demo_campaigns.py runs three campaigns over a synthetic 6-cell terrain with a decoy; the sealed witness ships in-repo (examples/witness.jsonl, 17 WITNESS + 3 CAMPAIGN rows) and CI regenerates it and refuses drift. Honest outcome: hunters find the terrain's hottest cell, refuse it (decoys_resisted=2 per campaign), the window holds twice on insufficient evidence (witnessed), then expands 2→4 with the reason on the chain.

13 green: witness append/verify/tamper, persistence, admission granted+refused witnessed with reasons, window caps, evidence-gated expansion, starvation contraction, broken-chain contraction, signal re-derivation, honest campaign counts, CLI roundtrip, and the boredom-honesty doctrine.

python -m pytest