2026-07-31 ticket initialization

User instruction: kontynuuj, with an explicit correction that executable source
must not live under project/ticket-*.

Previous measured result:
cross-language expected=0/6
cross-language forbidden violations=0/6
raw E5 new platform candidates=2
manually accepted raw E5 candidates=0

Workflow state: PLAN
Implementation status: waiting for P-CORE-008 review

2026-07-31 owner approval and continuation

User approved work on subsequent todo2code tickets and requested an explicit
audit of:
user-* / ai-* -> Intent DSL -> divergence -> required respondent.

Workflow transition: PLAN -> TOOLS
Human participant file remains unchanged.

2026-07-31 communication fidelity validation

Focused regression: 25/25 PASS for communication, identity, pipeline and task
synthesis after the initial implementation.

External read-only migration (`wellmanifest/new-project`, historical
2b9e3c9):
- filename-only rename: 0 records; explicit owner-specific migration warnings,
- Opus, typed request/message: 9 human + 58 agent records, 0 issues,
- GPT56Luna, typed request/message: 9 human + 72 agent records, 3 unanswered
  prompt fragments, 0 false human-agent file conflict.

Full gates after implementation:
- npm run verify: PASS (247 total, 246 pass, 1 local JDK skip),
- evaluate:gold v2 and v1: PASS, 100% gated precision/recall,
- examples:check: PASS (227 records, 97 relations),
- npm audit --omit=dev: 0 vulnerabilities,
- CLI, MCP, A2A and Docker smoke: PASS.

2026-07-31 audited reranker evaluation

Offline contracts:
- candidate set bounded to 1..10 per declaration,
- retrieval creates no relation,
- accept/reject/abstain decisions require both record IDs and exact grounded
  quotes,
- accepted relations retain retrieval, decision, reranker and citation
  provenance,
- captured gold reranker: 6/6 expected, 0/6 forbidden violations, 1 abstention.

Live tracked repository:
- repository: subactor/platform,
- commit: 3e96573d587cb664741849ceba205bf303b9f418,
- graph fingerprint:
  250df4ff83f456fb371278d4a0c2cf17dd025582b6865a0cb2f01bd469fd1dd0,
- selected reciprocal E5 shortlist: 6 declarations; top-3=18 candidates,
  top-1=6 candidates,
- qwen/qwen3.7-plus attempt 1: missing decisions array,
- attempt 2: returned judgments instead of decisions,
- attempt 3: invalid non-numeric/out-of-range confidence,
- result: fail-closed, 0 materialized relations, no coverage claim.

Final gates:
- npm run verify: PASS (251 total, 250 pass, 1 local JDK skip),
- one earlier full-suite CLI-watch timing failure; isolated retry 3/3 PASS and
  repeated full verify PASS,
- evaluate:gold v2: deterministic linker 0/6; captured reranker 6/6 expected,
  0/6 forbidden, accepted 6, abstained 1,
- evaluate:gold v1: PASS,
- examples:check: PASS (227 records, 97 relations, five SDKs),
- npm audit --omit=dev: 0 vulnerabilities,
- CLI, MCP, A2A and Docker smoke: PASS.

Decision: reject production semantic reranking; do not export it and do not
change the deterministic linker. Workflow state: DONE.

Final communication re-analysis after closing documentation:
- participants: codex 51 records, tom-sapletta-com 4 records,
- 0 blocking, 8 warning, 8 review_required,
- 7 AGENT_CLAIM_WITHOUT_EVIDENCE -> codex (workspace remains uncommitted),
- 1 AGENT_HUMAN_DECISION_CLAIM_UNCONFIRMED -> tom-sapletta-com,
- 8 AGENT_WORK_OUTSIDE_REQUEST -> tom-sapletta-com because the detailed latest
  instruction is present in the conversation but not in the human-owned file.

No human-owned file was modified to suppress these findings.