Generated from cat-harness/skills/authoring/authoring-core/compute-integration-watcher.md — do not edit here.
✎ Edit this page’s source

Compute Integration Watcher

Role

When a new claim lands on the default branch (or an open PR introduces one), the claim is not “done” until the compute chain that consumes it has been updated to reflect it. A proved theorem is only valuable to the project’s predictions if the production computation actually uses it rather than a fitted free parameter standing in for it.

This skill answers, on every new commit/PR event AND on demand against all existing blocks: for each claim, is the matching compute consumer wired in, using the correct primitives, at the correct precision, against the correct constants?

It is the cross-layer counterpart to:

local/compute-audit — patterns INSIDE one script
local/watch — event-stream router for upstream commits / PRs
local/proof-gap-audit — claims missing formal proofs
local/proof-integration-watcher — the proof-side analogue (chases gaps / naked conjectures across the default branch + sibling PRs). When both are armed: proof-integration-watcher owns proof-QA gaps; this skill owns compute-wiring gaps. Gaps bridging both layers are double-claimed — pick the one whose delivery cycle reaches the default branch first.

Where those look at one layer in isolation, this skill looks at the boundary between layers: narrative → compute, formal → compute, upstream witness → downstream consumer.

When to use

User phrase	Action
“audit compute integration”	full idle-sweep
“is X wired in?” / “where is X consumed?”	targeted wiring check
“what claims are not in compute?”	idle-sweep, filtered to NOT_WIRED
“watch + check wiring”	reactive mode
post-merge of a provable block	reactive (auto-triggered from `watch`)
idle / waiting on a long task	start an idle-sweep agent in background

Invocation modes

A. REACTIVE (called from `watch` — “Can we reuse?”)

When watch sees a new commit on the default branch (or PR event) that adds/changes a provable block (or its formal sibling) and reaches the “Computation” step, invoke this skill on the affected label. Output: WIRED / PROBE-ONLY / NOT-WIRED verdict, consumer scripts that use (or should use) the block, and a recommended fix.

B. IDLE-SWEEP (user-invoked or background)

When the foreground is blocked, walk every provable content block, apply the I-patterns below, and emit a ranked workplan to a dated audit doc. The idle-sweep MUST be safe to interrupt — checkpoint every 50 blocks to a .wip.json.

C. AUTHOR-QUESTION QUEUE (idle-driven)

Maintain a queue of author questions and pull one off when idle to ask via AskUserQuestion. Each entry carries trigger context, a ≤2-sentence stem, 3–4 mutually-exclusive options, a priority (P0 blocks production wiring; P2 is design), and whether answering requires touching content. Consume in priority order. Never block on the queue — if no pending P0/P1 question and the foreground is busy, work the next improvement. Add to the queue (don’t ask immediately) when a fix has architectural sub-decisions, a sibling PR needs owner sign-off, or a pattern hit is ambiguous.

The integration antipatterns (I-patterns)

Group A — Missing wiring

I1. PROVEN, NO PROBE. A provable block has a formal sibling with no open gaps but no probe script and no witness referencing the label via contentBlock. Fix: write a probe exercising the claim + emit a witness whose contentBlock points at it.
I2. PROBE-ONLY (NOT CONSUMED IN PRODUCTION). A probe witness exists, but the production consumers do not import or reference its witness file or derived constants. Fix: thread the result into the relevant production formula, or, if intentionally probe-only, add a # probe-only: <reason> line.
- I2-a. PROBE-ONLY that SUPERSEDES a production element. The new block is the derived, no-fit version of what an existing production element used as input. Wiring it in also requires a “fate of the predecessor” decision (keep as interpretation / demote / deprecate / delete; default keep + complementary), a calibration -count update if the predecessor was an input, a sweep for downstream consumers of the predecessor’s derived constants, and verification that the predecessor’s narrative content has a home before deprecating. Deliver in phases (foundation → predecessor deletion + repointing → production wiring), one commit each, keeping the validator clean.
I3. WIRED BUT GUARDED OFF. A consumer references the witness but the call is behind an if False / off-by-default env var / legacy fallback. Fix: promote to default-on after an equivalence cross-check.

Group B — Stale wiring

I4. WITNESS-HASH DRIFT. A consumer records upstream_witness_hashes; an upstream hash on disk no longer matches. Fix: rerun the producer and recommit both.
I5. CONTENTBLOCK REFERENCES MISSING LABEL. A witness’s contentBlock names a label with no .ts/.md file. Fix: create the block or point at the nearest neighbour.
I6. EMPTY data SECTION IN A PRODUCTION WITNESS. Run the script with fault handling; usually a silently-caught exception.
I7. PROBE WITNESS STALE vs THE BLOCK’S MTIME. The block was edited and the probe wasn’t rerun. Fix: rerun; if numbers shift, surface to the author before recommitting.

Group C — Wrong primitive

I8. PRODUCTION CHAIN USES LOW PRECISION WHERE THE WITNESS IS HIGH. The consumer erodes the witness’s precision via a lossy float cast. Trace the value through a precision-preserving conversion and confirm the final precision matches.
I9. HARDCODED EXTERNAL/DERIVED LITERAL OUTSIDE THE REGISTRY. A production literal that should be derived from a claim or imported from the registry. Surfaced here because it also breaks wiring.
I10. LEGACY FORMULA USED IN PLACE OF CANONICAL. A consumer inlines/imports a deprecated formula superseded by the current one. Fix: replace the call; keep the deprecated import in _deprecated/.

Group D — Cross-paper / cross-layer

I11. CROSS-PAPER uses[] NOT CONSUMED. A block’s uses[] cites a paper-dir:label upstream but no compute consumer references the upstream’s witness. Fix: drop it if narrative-only, or wire the dependent witness in.
I12. FORMAL-REFERENCE MISMATCH. A block’s formal-reference URI does not resolve to an actual declaration. Detected by the validator; surfaced here because it commonly catches “renamed, forgot to update”.

Group E — Discipline drift

I13. FITTED LITERAL LABELLED AS DERIVED. A production script defines a high-precision literal AND claims (docstring/witness) it is “derived”/”canonical”/”predicted”, but it was solved from external data rather than from the declared calibrations. The discrepancy inflates the project’s apparent accuracy. Fix: rename to *_FITTED_FROM_DATA, add a discipline-violation banner, emit the honest derivation status, and demote any downstream “canonical” label.

Workflow

Reactive (per-commit / per-PR-event)

For each new/changed provable label in the commit: diff to find affected labels; for each, walk I1–I13 in order, stop at the first hit, emit a concrete fix; if clean emit L: ✓ wired. Roll up into one comment with actionable items at the top. On a sibling-PR commit, act if the fix is small and local to your branch; otherwise queue a TODO on the PR.

Idle-sweep (full content scan)

Enumerate every provable .ts block. For each: read the formal sibling (skip conjecture-class blocks — they don’t need a compute consumer); grep witnesses for "contentBlock": "L"; grep production scripts for the producer witness or the label; walk I1–I12; emit hits. Emit a dated audit doc with a per-label table, aggregate counts by pattern, and next-actions ranked by leverage × tractability. Batch ≤30-minute fixes into one follow-up branch.

Verification checklist before declaring a claim “wired”

Formal side proved (or conjecture-class instance).
Probe witness exists with contentBlock: "<label>" and an assertion that fails CI if the claim is broken.
At least one production consumer imports the probe witness, the producer script, or the derived constant.
Equivalence cross-check recorded in the consumer’s witness.
Anti-staleness wiring — consumer’s witness has an upstream_witness_hashes entry for the probe witness.
The project validator passes (catches I5 / I12).

Anti-patterns this skill explicitly rejects

Auto-fixing wiring on someone else’s PR branch — I-pattern hits on sibling PRs become one batched comment, not a force-push.
Generating probes for conjecture-class claims — they don’t need compute consumers.
Declaring “wired” on a docstring-only mention — the consumer must actually load the witness or compute the derived constant.
Overclaiming convergence — “we wired it, look at the agreement!” when the routes compute different objects.

Tie-in with `local/watch`

watch’s “Can we reuse?” → “Computation” step is the entry point: when a commit adds a provable block, watch delegates here, captures the verdict, and acts-or-defers. When the commit adds only a witness (no new claim), this skill is unnecessary — compute-audit covers in-script patterns.

AGENTS.md interaction

Respects the same rules as compute-audit: never write to protected files; never close/resolve TodoItems (human-only); cite the matching I-pattern for every recommended fix; for dual paths (probe-only vs production), preserve the probe-only path as an opt-out — never delete it.