[Dream Cycle 2026-10-16] ledger-signals: witness verify-report false-positives on a legitimate historical report + witness,verify scan #188

Rotation

Slot 1 (DAYINT % 5, DAYINT=20261016) → DEEP=ledger-signals, SCAN=witness,verify. No bonus modulus hit (% 25 = 16, % 75 = 66). Session commit (baseline, after git fetch --unshallow): b2abcc14a632d7e0948962065a7c6c8920cf306c.

Ledger Check

docs/dream-cycle/LEDGER.md last row 2026-10-12 (daysSinceLastRow=4). Fresh checkout was shallow (git rev-parse --is-shallow-repository → true, only 59 local commits) — the exact defect class already tracked by issue #149/PR #150 (open, draft, mergeable_state: dirty); unshallowed via git fetch --unshallow before any measurement.

dream-machine ledger signals (no flags): zeroMergeStreak: true — the tool's own documented worst-case default. Ground-truthed via pull_request_read on PR #177: merged: true, merged_at: 2026-10-12T15:45:01Z. zeroMergeStreak is actually false (last merge 4 days ago). reviewBacklogSize: 16 open, unmerged dream/* PRs, none overlapping tonight's candidate file.

Checked #149/#150 (shallow-clone git-show oracle) in full: correctly withheld by prior nights, not re-litigated. Checked the two most recent ledger-signals-surface PRs in full: #145 (2026-10-01, splitRow pipe-unescape, issue #144) and #154 (2026-10-06, blockedEvalStreak missing majority conjunct, issue #153) — both still open/draft/unmerged, their defects still present on current main (confirmed by direct code read); not re-proposed. duplicateDirections: [] (the 2026-10-11 dedupe-by-night fix, PR #167, is merged and present).

Deep Dive

Manually audited every file under docs/dream-cycle/ for ## Witness heading count. Exactly one file, 2026-09-10-compiler-parity-report.md, has two such lines (93, 118). Live repro:

$ node packages/cli/dist/bin.js witness verify-report docs/dream-cycle/2026-09-10-compiler-parity-report.md --commit 3edd426f6c9c4b1e80235f7447dc863e749345cc
✗ witness INVALID — report has 2 "## Witness" headings at lines 93, 118; expected exactly one

Root cause: the file predates splitAtWitnessSection/witness verify-report (added 2026-09-15, PR #111, closing issue #112). Its first occurrence (line 93) is a forward-reference placeholder — "See STEP 16 below..." — written mid-document before the real trailing section (line 118), not a genuine second section. Independently confirmed by hand that the originally-published report_sha256 (1c9ec5548ffab1370fadcaf24961a5779b0eaf52690b9314a6483d5e572bec2b) is correctly reproducible from the file's current bytes. This is a tool false-positive on an already-correct report, not a corrupted one.

Scoped across the whole corpus first: ran witness verify-report against every report with a ## Witness heading. Only this one file throws the "ambiguous headings" error. ~24 other files also fail self-verification, but for a different, already-partially-known reason (they predate, or don't follow, the file-embedded-triple convention PR #111 introduced). That is a separate, larger conceptual change, explicitly not built tonight (see Recommendation).

No existing corpus-wide witness verify-report regression gate exists anywhere in the repo — this exact false-positive could recur silently with nothing to catch it.

Hypothesis

Given the committed report corpus under docs/dream-cycle/, when a new governance test iterates every file and asserts any file with more than one literal ## Witness heading line is on an explicit, documented, hash-verified grandfather list (and that the list contains no stale/unreproducible entry), then the test should fail today pointing at 2026-09-10-compiler-parity-report.md's ambiguous heading, and pass once that one entry is added — without changing splitAtWitnessSection's existing fail-closed ambiguity behavior (a separate, already-passing test deliberately locks that in) and without editing any already-merged historical report's content.

Evaluation Receipt

Real evaluator: npm test (vitest run && npm run test:governance).

Non-tautology proven via 4 live negative controls (grandfather list emptied; stale entry added; fabricated entry with wrong hash added; real entry's hash corrupted) — each fails exactly as expected, then restored and diff-verified byte-identical. Full receipt in the PR body and committed report.

Darwin Results

DARWIN=not-applicable — a single new regression test over a fixed document corpus has no evolvable population or numeric objective (same judgment as every prior one-conceptual-change fix on this surface). Entrypoint probed live: npx --yes @metaharness/darwin --version responds (reachable, not blocked; unpinned-npx supply-chain note already tracked elsewhere, not this candidate's to fix).

Evidence

OBSERVATION (corpus audit finds exactly one ambiguous-heading file) → OBSERVATION (live repro) → OBSERVATION (byte-level hand reproduction proves the published hash correct — tool false-positive, not corruption) → HYPOTHESIS (frozen above) → MEASUREMENT (1004→1007 tests live before/after, 0 regressions) → MEASUREMENT (4 live negative controls) → INFERENCE (independent critic, separate context) → DECISION (critic: CLEAR + 1 disclosed gap, closed same session) → DECISION (ship; ACCEPT).

Reward-Hack Check

Independent adversarial critic (separate subagent, no authorship stake): CLEAR, one disclosed non-blocking gap found and closed same session. Confirmed diff purely additive (one new file; git status --porcelain/git diff --stat empty otherwise); no existing test/fixture/gold-answer/threshold touched (report-witness.test.ts's duplicate-heading-rejection tests re-read in full, unweakened); independently reproduced the grandfathered hash from scratch. Gap: first draft's grandfather list was a bare filename Set with no automatic proof of authenticity for a future entry. Closed: converted to a Map carrying each entry's own byte marker + expected hash, with the reproducibility test now generic/data-driven over every entry — mirroring @dream-machine/ledger's legacyPrefixDigest fail-closed design; re-verified the critic's exact attack (a fabricated entry with a wrong hash) now fails mechanically.

Security Review

No prompt-injection, MCP-authority, or credential surface touched. No new filesystem-write surface (read-only, fixed repo-relative path, no traversal surface). No network/exec surface. No regex (no ReDoS exposure). No production/runtime code path changed — new test infrastructure only.

Scan Findings

witness: witness verify-report false-positives on an already-legitimate, independently-reproducible historical report because of a pre-tooling forward-reference heading. Fixed with a narrow, explicit, hash-verified grandfather mechanism — splitAtWitnessSection's fail-closed ambiguity behavior is unchanged for any genuinely ambiguous future report.

verify: no corpus-wide verification gate existed for this tool at all. Tonight adds the first one, scoped to the one defect class discovered; the larger "older convention can't self-verify" gap (~24 files) is scanned, confirmed, and explicitly deferred.

Competitors

SLSA/in-toto and Sigstore/cosign (grade A, official docs): the standard pattern gates CI on verifier success rather than treating it as informational — the posture this candidate gives witness verify-report for the first time. Sakana AI Scientist / OpenHands / DSPy-GEPA / SWE-agent / AutoGPT lineage: none publish a comparable self-referential, file-embedded provenance triple to benchmark against (checked via live search, grade A sources only); no hits for a corpus-regression angle on any of the five.

Gist

GIST=LOCAL — no gh CLI gist capability this session (GH_TOKEN invalid) and no gist-creation MCP tool available. GitHub MCP access itself works fine — FALLBACK=false. Full report committed at docs/dream-cycle/2026-10-16-ledger-signals-report.md.

Witness

report_sha256 : 34e07d7b29706b44ece0b62e5843a32c7bf1b4ee8848b1cb7027ccbf20a13238
session_commit: b2abcc14a632d7e0948962065a7c6c8920cf306c
witness       : 3fc2fa96d63069b3d20d36e1ce3a519c0cee9c1b24bb99e5694eadbb40291bb3

Reproduce: dream-machine witness verify-report docs/dream-cycle/2026-10-16-ledger-signals-report.md --commit b2abcc14a632d7e0948962065a7c6c8920cf306c

Recommendation

evaluated: yes / verdict: ACCEPT. All promotion-gate conditions clear. Human review requested for merge — this session never self-merges. Next steps (not built tonight): (1) grandfather or re-stamp the remaining ~24 older-convention reports — a separate, larger conceptual change; (2) consider having STEP 26's self-review run witness verify-report on the night's own freshly-committed report; (3) wire ledger signals --merged/--open-count ground-truthing into the compiled prompt's STEP 1.1 instructions; (4) issue #144/PR #145 and issue #153/PR #154 remain open/draft, confirmed not duplicated.