interlinked-observability

interlinked-observability — inspect what agents did

For internal claim-verification research, node --import tsx scripts/internal/jev.mjs claims <final-message> <transcript> --receipt-shadow <links-file> adds default-off, log-only deterministic comparisons to .interlinked/claim-receipt-shadow.jsonl. It does not run from public Stop hooks or change Jev's decision. Links explicitly bind a claim to receipt/source/ selection identities; missing or duplicate links remain not assessable. Only four narrow receipt predicates are supported; see docs/internal/jev.md. A missing Jev score is unmeasured, including short claims omitted by the legacy splitter. Separate replay statistics from live usage, and inspect receiptShadowFailures() before treating an absent row as no claim.

Ordinary hook delivery omits deep-audit style/design messages after one review notice, while raw check results remain intact. This applies to synchronous responses and the late-warning spool; their deduplication state is separate so preparing a fallback cannot consume the synchronous notice. Repeated guard-scope advice is deduplicated per session. Compare native delivered feedback separately from raw finding counts; fewer delivered warnings do not imply fewer findings or clean code. Interactive ChangeSet batches defer project-wide Gitleaks to explicit verification or background recovery, with an explicit non-verdict. Confirm actual tool timings and completion before claiming latency or security-coverage improvements.

For Claude compiler batching, distinguish tool calls, scheduled checks, compiler executions and delivered diagnostics. hook-observation.v1 rows record scheduled TypeScript work; batch results and timing are in check-results.jsonl. Pending work survives in .interlinked/compiler-batches/. Coverage observations are retained even when repeated model warnings are suppressed. Silence does not mean coverage is complete. Measure a repeated-edit problem by assistant-message/tool-use identity and timestamps: an edit emitted before an earlier hook's feedback cannot be attributed to that feedback.

check-executions.jsonl also carries hook-observation.v1: scheduled external overlays, normal complexity/token measurements, and shadow trajectory observations. These are telemetry, not check passes or confirmed model delivery. check-results.jsonl retains full findings, including novelty, findingCount and server-computed writeAttribution where available. Newly observed means report novelty, not proven causation. Repeated advisories may be hidden from model feedback while remaining in each call's records; hidden findings prevent all-clean.

Use evals/compare-builds.mjs for paired explicit-build agent pilots. It retains fixtures, native streams, usage categories, settings, daemon status, final diffs and evaluator-owned functional checks. Activity warning counts are ledger metrics. Missing native delivery/token data remains unknown; do not infer billed savings from shorter commands or compressed rows.

repeated_implementation check results retain each current AST-matched group with a group=<sha256> detail identity, even when unchanged agent-facing warnings are suppressed. Compare these identities rather than line numbers across formatting edits. Post-tool suppression is session-local and resets with the daemon; Stop uses its existing digest state. NOT CHECKED means missing analysis, not zero clones. Findings are review candidates, not proven defects or evidence that an extraction helped.

hook-reentry.jsonl records Stop/SubagentStop re-entry suppressed before daemon evaluation. Translation receipts now include advisory context representation as well as denials and rewrites. Operator-only Stop diagnostics are not model-context delivery. Audit native completion messages by unique ID; streaming chunks and daemon Stop counts are insufficient to count continuations. The SlopCodeBench exporter retains these receipts and registrations.

Shared hook-check receipts may include scopes with a batch ID, check name, configuration hash and captured request-file identities. kind: request-inputs is historical execution evidence, not a complete runtime/dependency closure or reusable cache key. Nonempty unavailable preserves partial evidence and leaves the obligation pending. Keep those states separate from clean results when comparing recovered check counts or latency.