deja-vu report verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

The current reports support a bounded implementation account. The previous semantic blockers are repaired; two local limits remain. I read the complete boundary, five reports, their types, correction answers, structural check, records contract and previous verification. The structural check reports none. Only epistemic supplies correction answers; they are arguments assessed below, not evidence of acceptance. I independently inspected source text through bounded reads and searches. SRC-1 HEAD matches the reviewed revision, short status is empty, and its top-level tree agrees with the boundary coverage. No target code, tests, hosts or providers were executed.

Records, anchors and evidence strength

Every declaration and annotation was checked for referent, supporting source anchor, consumer, evidence strength and consequential qualification. The dispositions below cover the declared records; they do not assert exhaustive implementation inspection.

  • Runtime delivery and policy. RT-RTE-auto-delivery, RT-RTE-requested-recall, RT-RTE-spawn-context, RT-OBJ-policy and RT-BAP-recall-reference support their route-scoped findings. Independent inspection covered the complete SRC-1 extensions/pi/index.ts and cmd/deja/hook_spawn.go, cache gates and stale service in cmd/deja/hook_context.go, MCP ranking/refiltering in cmd/deja/mcp.go, activation/egress checks in internal/policy/policy.go, and framing and delivery-token sites in cmd/deja/recall_frame.go, cmd/deja/hook_prompt.go and cmd/deja/hook_tool.go. The pi consumer turns stdout into model-facing messages or tool-result text; it neither schedules inference nor executes recalled repairs. The cache gate uses policy description and notes stat, not complete policy identity. Semantic fallback is filtered again. Spawn rewrites preserve other input fields and use a synthetic reader, not a shared parent window. Framing requests treatment as data; it does not enforce a permission boundary. No all-host parity or receipt guarantee follows.
  • Notes, lifecycle and extraction. RT-RTE-note-admission, MEM-OBJ-notes, EPI-OBJ-promoted-text, EPI-OBJ-lifecycle, EPI-RTE-promote-distill and EPI-RTE-conflict-check distinguish structural admission from semantic approval. SRC-1 internal/sources/notes.go implements append, reload, newest-first promoted messages, latest promoted-note selection and topic-overlap notification; internal/sources/lifecycle.go retains keyed caller state. cmd/deja/promote.go selects first ask/last assistant or detected decision and trims excerpts. The after-write overlap warning is not contradiction detection or rejection. Source identity and an accepted mark do not prove truth or human approval. The factual extraction classification is now appropriately indeterminate rather than a universal semantic-preservation claim.
  • Conventions and authority. MEM-RTE-convention-delivery and EPI-RTE-convention-use identify the same selector/renderer. Complete inspection of SRC-1 cmd/deja/hook_conventions.go confirms accepted local notes, imported metadata under auto policy, project matching, recency, rendered-line deduplication and first-sentence selection. Its call in cmd/deja/hook_context.go supplies the six-note/800-byte bound. The Use pgx. Only in service A. deduction is valid for the renderer but demonstrates changed directive scope, not a truth-apt transformation category. The corrected ledger separates pure directives from factual reminders. Imperative reminders crossing RT-BAP-recall-reference remain a model-interpretation tension, not an unresolved disagreement between reports.
  • Index and trace-derived parts. RT-OBJ-index-db, MEM-OBJ-records and MEM-OBJ-session-meta correctly distinguish derivative text from fixed-rule metadata and imported display fields. SRC-1 internal/index/ingest.go contains sanitized record writes and sidecar build/update calls; the convention reader establishes the imported lifecycle-text fallback. MEM-OBJ-cooccur, MEM-OBJ-commands, MEM-OBJ-commandfails and MEM-OBJ-sessionfacts have supported separate substrates and consumers in SRC-1 internal/index/cooccur.go, internal/index/commands.go, internal/index/commandfails.go and internal/index/sessionfacts.go. Co-occurrence is symbolic association, not model weights; recurrence counts are not endorsement. The session-fact exact-cap exception is retained in Limits. These positive mechanisms do not establish complete parser fidelity, incremental equivalence or adaptive use of usage history.
  • Fix checking, disposition, retention and serving. MEM-OBJ-fixraw, MEM-OBJ-fixes, EPI-OBJ-fixraw, EPI-OBJ-fixes, MEM-RTE-trace-mining, EPI-RTE-fix-check, EPI-RTE-fix-disposition, EPI-RTE-fix-retention and EPI-RTE-fix-selection preserve distinct phases. Independent reads of SRC-1 internal/index/fixpair.go covered build, judging, failure detectors, mining, edit fallback, replacement-key recomputation and final selection. Unknown outcomes can pass failure detectors; the corrected check target and entailed-result wording now say this. Judgment consumes feature/repetition evidence for historical selection, not repair causation. Raw evidence permits later recomputation; it is already mined material, not an original log. Candidate fallback is implemented despite the contrary struct comment. SRC-1 cmd/deja/fix.go labels it unconfirmed, while cmd/deja/hook_plan.go rejects it for plan remedies. Those are consumer-specific gates, not report conflicts.
  • Plan checks. EPI-RTE-plan-check and EPI-RTE-plan-delivery retain recurrence evidence and advisory delivery rather than approval. SRC-1 cmd/deja/hook_plan.go rechecks eligible carrier counts, filters command evidence, budgets findings and emits framed additional context without a deny decision. Its source passages support that scope; the reports withhold exact join completeness in internal/index/plan.go and do not interpret silence as plan safety.
  • Continuation and recovery. MEM-OBJ-compaction-data, MEM-OBJ-compaction-state and RT-RTE-compaction-recovery are supported by independent reads of SRC-1 internal/index/compaction.go, internal/digest/compaction_context.go and cmd/deja/hook_compaction.go, with summary/carry and parser-fallback searches in internal/digest/compaction_carry.go, internal/digest/compaction_rules.go and internal/sources/compaction_registry.go. Packet data are non-searchable retained content, distinct from freshness/identity metadata. Prior carry/rules can feed a new capture; storage has bounded retention and source/privacy gates. The emitter marks delivery after a successful complete write, not host receipt. The corrected epistemic ledger separately identifies the ampliative passed inference from no recognized friction. General extractive factual content remains indeterminate and pure directives remain non-truth-apt.
  • Withdrawal and alternative consumers. MEM-RTE-forget-cleanup and MEM-RTE-session-recovery retain their bounded local effects. SRC-1 internal/index/privacy.go persists suppression before rebuild and removes matching continuation state; internal/sources/notes.go distinguishes note deletion from clearing a source's borrowed title. cmd/deja/handoff.go supplies policy-aware bounded handoff; cmd/deja/restore.go checks destination safety. Direct reads of cmd/deja/resume_writeback.go and internal/index/writeback.go confirm conversation-only recreation, new-file/path guards and adoption of file metadata to preserve richer indexed evidence. None establishes original tool-state reconstruction, atomic global erasure or peer revocation.
  • Standing-rule workflow. RT-OBJ-rules, RT-RTE-rules-sync, RT-BAP-standing-rules and the workflow account and annotation are consistent. SRC-1 cmd/deja/rules_candidates.go, cmd/deja/rules.go, cmd/deja/marked_block.go and the complete skills/deja-search/SKILL.md support extraction, clipped/prefix-deduplicated evidence, prescribed semantic proposal and human selection, known/ready targets, partial distribution, unmatched-start refusal and source-driven removal. Model synthesis and consent are instructions to the host, not code-enforced acceptance. File-level delivery is not later loading or obedience. The lone-end countercase correctly bounds malformed-marker claims.
  • Models, machinery and evaluators. RT-CMP-embed-client, RT-RTE-embedding-records, RT-RTE-import-records, RT-RTE-binary-update and RT-RTE-wiring-refresh have supported source-native checks, not observed deployment. SRC-1 internal/embed/client.go, internal/embed/sidecar.go, cmd/deja/embed.go, internal/index/sync.go, cmd/deja/update.go and cmd/deja/wiring_state.go confirm configurable unpinned models, model-name/dimension reuse, record-origin egress filtering with unknown-key allowance, locked import setup/per-file rollback state, checksum agreement and retry-oriented wiring stamps. RT-RTE-retrieval-evaluation and EPI-RTE-offline-scores correctly separate gold-session membership, warning census and literal substring coverage in SRC-1 scripts/longmemeval/main.go, scripts/planbench/main.go and cmd/deja/bench_context.go. Reporting is not a production acceptance gate or causal benefit evidence.
  • Public/design claims and statuses. RT-CLM-search-before-edit, EPI-CLM-knows-held and EPI-CLM-history-judgement retain the source-native claims from SRC-1 skills/deja-search/SKILL.md, README.md and cmd/deja/mcp.go, with the necessary mechanism/warrant distinctions. No report declares an evidenced absence. wired and architectural implemented describe source-supported capacity; operation remains uninspected. No route is promoted into observed learning, obedience, reliable repair or causal improvement.

Relations, reconciliation and corrections

I checked every Part of: relation: index parts belong to the index container; fix retention explains the cross-kind maintenance relation; distillation and conflict checking are note-admission parts; convention delivery is a startup-delivery part. Containers remain valid alongside material parts. The three duplicate supersessions correctly identify the two fix pools and the convention selector. The combined memory trace chain is not identical to each epistemic phase. Local promoted prose and lifecycle state are distinct operative material within the broader notes account; imported lifecycle metadata is a different retained path. Workflow annotations add handoffs and delivery criteria without redeclaring runtime identity.

The current reconciliation dispositions no longer excuse semantic classification by calling the implementation extractive. They preserve the corrected indeterminate factual relations, separate directive example and referent identities. The previous reconciliation blocker is resolved. No missing conflict citation or unresolved inter-report disagreement remains. Its retained precision and alternative-delivery limits do not change analyst values or imply acceptance.

Both epistemic correction answers hold against the frozen source and all dependent ledger, route and bounded-conclusion passages. The former universal non-ampliative extraction classifications are removed; the continuation outcome inference is separately assessed. The fix-check target no longer assumes a recorded nonfailure. The remaining positive gates and consumer-specific dispositions remain intact. No blocker is carried forward merely because the previous verifier raised it.

Omission questions and stopping decision

The search basis was the frozen top-level tree, dispatch/consumer paths and producer-to-consumer dependencies, not only quoted anchors. Three consequential questions were selected:

  1. Does extraction change requested scope or factual content? Complete convention rendering, promotion distillation and continuation capture inspection establish scope-loss and inferred-outcome countercases. The corrected epistemic account preserves these rather than claiming universal semantic preservation. The bounded account remains supported.
  2. Does fix checking require successful completion? Failure detectors, mining, judging and downstream consumers in SRC-1 internal/index/fixpair.go settle this negatively for that premise: absent exit/output evidence can pass. The emitted findings now license parsed sequence/detector facts and historical selection, not actual successful repair.
  3. Can alternative shipped consumers invalidate immediate-delivery or recovery guarantees? The complete pi consumer was compared with SRC-1 cmd/deja/hook_deferred.go, including queue write, take and wrapper calls. Deferred paths retain text and later destructively consume it before final stdout receipt. The ordinary recovery emitter's full-write guarantee therefore cannot extend to the queue or host. The reports already withhold all-host immediacy and end-to-end retry safety; this remains a limit rather than a new broad guarantee.

Further optional omission expansion stopped because these challenges have supported dispositions, no identified material contradiction remains undisposed, and the central conclusions are positive route-scoped mechanisms. Parser, installer, export, specialized-hook, usage and evaluator alternatives remain incomplete inventories; no complete fidelity, global revocation, aggregate performance or system-complete epistemic negative depends on their omission. The frozen whole-system functional boundary remains supported; it does not assert exhaustive inspection.

Additional implementation support for the two carried limits is retained here:

// One command past the cap, kept for the one that checked the work. --- internal/index/sessionfacts.go

if err := os.WriteFile(tmp, []byte(text), 0o600); err != nil { return } _ = os.Rename(tmp, filepath.Join(d, name)) --- cmd/deja/hook_deferred.go

if err := os.Rename(src, taken); err != nil { return "" } defer os.RemoveAll(taken) --- cmd/deja/hook_deferred.go

These are implementation deductions, not executed probes. The verification preserves the separation between source wiring, activation and demonstrated benefit.

Blockers

none

Limits

  • MEM-OBJ-sessionfacts describes a four-command bound, but SRC-1 internal/index/sessionfacts.go, sessionFactOf, permits an additional verification command beyond the ordinary recent-command budget. This is a local inventory precision defect: retain the distinction between known exit and merely recorded command, but withhold exact four-entry capacity comparisons. It does not invalidate the trace-derived table or its later-consumer account.
  • The memory coverage table, epistemic coverage table and RT-RTE-auto-delivery retain incomplete alternative-hook coverage. SRC-1 cmd/deja/hook_deferred.go implements filesystem retention and later destructive queue take around ordinary hook delivery. The scoped pi account and index-to-hook-to-model opportunity remain supported; comparisons of all-host immediate delivery, queue withdrawal and retry safety must remain withheld. RT-RTE-compaction-recovery's full-write bookkeeping is not an end-to-end queue or host-receipt guarantee.