Daily Signal Engine & Reviewer Pipeline

Scorecard

1 · Corpus consistency — August

The corpus holds two parcel layouts at once. That is the intended migration state, and the audit's first job was to prove the split is clean rather than assume it.

The 1,191 converted parcels are exactly the 1,191 reviewed parcels; the 1,361 old-shape are exactly the unreviewed. The layout boundary and the review boundary are the same boundary — there is no parcel that is reviewed but unconverted, or converted but unreviewed.

Reader equivalence

Every converted parcel was read through the independent reader classes and compared on findings, sources, review_run, contract_version and review_state:

This is the property that failed before the read-half fix landed — a converted parcel reported findings ABSENT, and absent is how OKF encodes "never reviewed". It now holds corpus-wide.

2 · The step function nobody should read as decay

Aggregate source-anchoring across the reviewed corpus is 27.3%. Read as a quality metric that looks alarming. It is not a quality metric — it is two contract regimes averaged together.

Every day before 2026-08-13 is 0%. Every day from 08-13 is 100%. There is no partial day and no decay: the anchor key was introduced by contract v2.3 and adopted completely on the first batch that ran under it. The two source shapes in the corpus are exactly (id, resource, title) ×9,845 and (anchor, id, resource, title) ×3,702 — no third shape, no half-populated records.

The same shape appears in findings density

Findings-per-document drops from ~26 to ~4.2 at the v2.4 → v3.5/3.6 boundary. This is the signal-or-nothing contract change working as designed — a no-signal document now earns the index layer only and is a complete record — not reviewers doing less work.

3 · Signal-graph engine correctness

The load-bearing question: does the parcel conversion change what the engine produces? Answered by re-emitting real August days from the converted corpus and comparing section by section against what shipped.

Compared as canonicalized JSON, section by section — nodes, edges, counts, axes, layout, quadrants, vendorable all hash equal. The conversion moves no payload.

A digest difference that is not a payload difference

The content_digest does differ between shipped and re-emitted payloads. Three causes, all provenance, none content:

• digest_recipe v1 → v2 — v1 hashed generated_from; v2 excludes it. A recipe change moves the digest by design.
• generated_from.parcels — shipped payloads carry the n8n runtime path /files/ViskaRes/…; current code emits repo-relative. Same corpus, two spellings.
• --since omission — re-emitting without it drops edges_from_history (28 → 0 on 08-23), because edges are additive across days. The engine is right; a bare --date re-emit is the wrong comparison.

4 · Findings

5 · Deliverable inventory

What actually exists on disk, per August day, against what the pipeline contract describes.

The DAILY-CLOSE absence was checked with a positive control — find resolves other names in the same tree, so the search could see. Zero files exist repo-wide under that name, and check_folder_complete.py names it as a leg-1 requirement. These are seat-only legs that read the local Dropbox corpus, so a gap here does not necessarily mean the day failed — but it does mean the artifact the contract names as the close is not in the repo for any August day.

Index and manifest agree with disk

The two excluded days are the insufficient_corpus holds, which the engine marks vendorable: false with a stated reason rather than shipping an unexplained empty graph. Correct behaviour. One nit: the manifest header still declares payload_schema: viska.signal-graph/1.0 while its contents are 1.1 and 1.4.

6 · Reviewer pipeline

The A1 anchor halt is substantially resolved: the class-B elision trims landed, and 08-23 now carries zero unanchored sources and zero missing quotes. The remaining gate failure is the single uncited thesis.

Two open A1 classes

• Class A — 5 anchors, RES, deliberately red. Body has docling page furniture spliced mid-sentence; the anchor is faithful to how the document reads. The gate is wrong here, not the reviewer, and the fix is a contract change to compare with furniture stripped — not a loosening applied at speed.
• Class C — 2 anchors, N8N-authorized strip, not yet applied. Needs the same measure-the-alternatives treatment class B received, because dropping the sources cascades into cites_unresolved.

Reviewer model spelling drift

August reviews name 16 distinct model strings for roughly 8 actual models — sonnet-5 beside claude-sonnet-5, mimo-v2.5 beside xiaomi-mimo-v2.5, four spellings of Haiku 4.5. The 33 OpenRouter provider-prefix records were repaired this cycle; the remaining variance is reviewer-side spelling, and it makes any per-model quality comparison unreliable until normalized.

7 · What was verified, and how

8 · Recommended order

• N8N — re-read the two unread 08-23 documents. Smallest gap, additive, closes the only coverage hole in the month.
• N8N — derive expected_at from date_precision. 160 records is the largest live-contract class and it is one rule stated once.
• RES — rule class A on the anchor contract. Extend the boilerplate-strip control to furniture spliced inside an anchored span; then apply class C.
• RES — reconcile CONTRACT-v2.md or retire it. Low severity by standing ruling, but it currently makes the operator's approval tool report a false NO.
• Unassigned — run the byte-parity comparison. The contract binding daily.yml's retirement says "Byte-parity with daily.yml is the acceptance test", and both seats are on record that it has never been run. RES's own condition for closing it — that the first unattended n8n run happen first — was met today (3a316f51, batch 08-23, automated, 11:45Z). The precondition is satisfied and the test is nobody's task.
• Deferred — the plot/daily/close deliverable gaps. Needs an owner ruling on whether those surfaces are still live before anyone backfills 16 days of payloads.