Tasuki Keifu: Data Governance Agent Optimization History
Tasuki Keifu: Data Governance Agent Optimization History
Updated November 13, 2026. This retrospective preserves failed and unverified iterations rather than overwriting them with the current design. The PRD owns current requirements; the technical design owns implementation; this history owns reasons and evidence. Times below are UTC. The October batching history describes the older architecture and is not this baseline.
Scope and evidence policy
The target remains ordinary Nittaidai meet 333, men's 5000m, heat 1: 44 rows. Meet 332 is existing data. Ordinary meet and NCG semantics remain distinct. Operators fix the workflow and agents rather than manually deciding identities to manufacture better results.
Keep old tasks, attempts, responses and audit records. Comparisons must identify immutable input, database state, rule/policy versions and model. Previous outputs are not ground truth. Count DNS separately from starts, planned separately from applied changes, and lightweight objects separately from heavy-research objects.
This delivery performed development, isolated tests, deployment and documentation only. No task submission, retry, dry-run, live model call or result import was performed. Historical runs below were not executed again.
1. Whole-group lightweight judgment and bounded cross-check
Problem/case: simple organization variants do not fit safe general string replacements; person identity can reveal organization problems. “大東第1高校 / 大東大一高校” needs independent entity evidence. Seventeen uncertain identities among 44 rows do not mean seventeen conflicts or heavy calls.
Change: organization rules → one whole-group organization lightweight request → person rules → one whole-group person lightweight request → one cross-check → at most one organization adjudication and affected-person reassessment → bounded deeper research. The fixed DAG has no backward edge. Reject circular identity/organization evidence; do not add configurable sub-batches.
Evidence/status: implemented and deployed in Agent PR #6, with regression coverage and stage metrics. The latest real attempt stopped before cross-check; complete live success is not established.
2. Protocol diagnostics and timeout propagation
Problem/case: a retry-limit message does not establish API unavailability. Unresolved responses included stray target IDs; provider diagnostics were insufficient; a configured 120000 ms timeout was not reliably propagated through the runner's 25000 ms default.
Change: ignore and warn on stray unresolved targets without authorizing a match. Keep structural and candidate checks. Persist safe call/stage/error-phase details, duration and provider/response IDs, status and output types when actually returned. Fix timeout propagation. Production explicitly selects gpt-6.1-sol through the pro route with 120000 ms. Never treat example settings as effective configuration or promise a provider ID when none was returned.
Evidence/status: deployed; configuration and propagation checked. The 25-second default was application-controlled, not an intrinsic model limit. See PR #9 and the timeout fix. Neither health checks nor diagnostics prove identity quality.
3. Human-readable failure reporting
Problem: failed emails looked almost empty; titles lacked event/group context; raw JSON obscured progress.
Change: batch alert titles identify event and group; event summaries name the event. Recover completed work and actual calls from durable journals without double-counting saved executions. Show readable states, funnels, calls and issues first, with raw metrics expandable. Unknown values stay unknown; dry-run writes remain zero.
Evidence/status: deployed in Agent PR #13, site PR #16, and the title changes in Agent PR #6. Complete independent attempt-level navigation and all pricing coverage remain follow-up work. The next real run must establish practical reporting quality.
4. Deduplicated cross-check input
Problem: duplicated candidate histories exceeded the application's input-size guard before an API request; this was not a provider timeout.
Change: shared person/organization/evidence catalogs, related-evidence filtering, and full-group rows, coverage counts and truncation notices. No tools, hidden splitting, or removal of the 512000-character guard.
Evidence/status: deployed in Agent PR #15. On the same saved context, offline serialization changed:
| Measure | Before | After |
|---|---|---|
| Characters | 676807 | 119650 |
| UTF-8 bytes | 735743 | 128270 |
| Source rows | 44 | 44 |
| Source organizations | 20 | 20 |
The 82.32% character reduction is not measured token, cost or latency savings. The latest attempt stopped earlier, so successful live cross-check remains unverified.
5. Latest real failure baseline
Historical task 38e0ea4c-2d2a-405b-a286-c43848c2605c; attempt 3ecd5e94-5980-4980-befc-e6f2110c0c55; November 13, 2026, 01:16:38.198–01:17:14.181 UTC. Mode: dry_run, business writes disabled.
| Stage/measure | Observation |
|---|---|
| Input | 44 rows, 20 source organizations |
| Organization rules | 16 matches, 2 provisional creations, 2 lightweight items |
| Organization lightweight | 1 match, 1 unresolved |
| Person rules | 2 matches, 25 no-candidate provisional creations, 17 lightweight items |
| Person lightweight raw response | 16 unresolved, 1 match rejected for missing independent citations |
| Actual LLM calls | 2; organization 7990 ms, person 26001 ms |
| Usage | 23155 input tokens, 1652 output tokens, 0 web queries |
| Later stages | No cross-check or heavy research, final plan or business writes |
羽生祥成's match cited only source:1189, without citing the supplied independent identity history. Normalization person match has no independent identity evidence rejected the whole batch. This was evidence validation, not an API timeout and not seventeen heavy calls.
6. Historical-entry rules and per-item downgrade: this fix
Problem: absent Membership caused unnecessary uncertainty despite strong historical evidence; one poorly cited lightweight verdict blocked its valid peers.
Cases: 羽生祥成 and others had exact kanji/kana and same-canonical-organization history without Membership. 村中悠隼 had historical DNS registration, useful for identity but not participation. 伊藤亘輝 / 伊藤光輝 and 吉田大雅 / 吉田大河 exposed weak homophones. 杉野主宙, 竹原勇生 and 澤田聡太 still require independently resolved organization variants/IDs. 飯田翔真's prior 法政大学第二 affiliation differs from 小田原 NR; any remaining question concerns identity/affiliation evidence, not DNS itself.
Change: exact_name_kana_historical_organization requires a unique exact kanji/kana candidate, a sourced entry from another race no later than this race, identical existing canonical organization IDs, no contradictory cohort and no competing Membership-based candidate. DNS can support identity without creating a start, grade or Membership. Preserve weak candidates for cross-check.
A person lightweight match lacking target-owned independent evidence, unless an explicit birth date matches, becomes per-item unresolved with original target, reason and citations retained. Preserve valid peers and pass both domains' decisions, rule traces, candidates and warnings to cross-check. Unknown citations, foreign targets, malformed structure and explicit birth-date contradictions remain blocking. No tools or weaker identity safeguards were added.
Evidence/status: deployed via Agent PR #17. Rule 2026-11-13.1; policy whole-group-cross-check/4. All 36 unit and 15 disposable PostgreSQL integration tests passed, covering DNS and weak-homophone positives and negatives for missing kana, different organizations, future/current-race history, missing sources, duplicate exact candidates, cohort contradictions, invalid per-item citations and cross-check rejection. CI found three optional-field assertion typing issues; these were fixed and checks passed before merge.
Unmeasured: no rerun means no measured reductions in lightweight/heavy items, cost or total time, and no verified final success rate or new business facts. Tests and candidate estimates must not be reported as production gains.
7. Later V2 changes and round closure (November 17–18, 2026)
The previous section described the post-fix state before a rerun. The later V2 work was deployed and the production review was completed; implementation, tests, and production decisions are kept separate here.
7.1 Separate entry labels from person normalization
club, company, and unknown now represent unconfirmed race-entry labels. They do not create formal Organizations or Memberships and are not standalone affiliation evidence for person identity. Race results retain organizationNameSnapshot. Local athletics associations, university clubs, and temporary representative labels follow the same boundary. If an unconfirmed label maps to one athlete in a batch, the database has exactly one same-name candidate, and no birth-date conflict exists, the person may be normalized directly without creating an Organization or Membership.
This removed the false blocking case caused by 松戸市陸協: 松井俊介 was merged into the canonical person while results and source references were retained, with no Organization or Membership created. The Agent rule was tested, merged, and deployed; the business correction was recorded separately from the graph change.
7.2 Independent Membership governance and the name contract
When an event source supports a school or confirmed corporate team, Membership governance can be rerun after the race import. Missing academic years no longer prevent a source-supported relationship without dates; identity review is reserved for competing organizations in the same stage. Race affiliation is not treated as proof that two same-name people are different merely because a Membership row is absent.
The import contract now removes internal spaces from Japanese kanji names while preserving the source text in evidence. The contract is covered in the importer and tests, and older data is governed under the same display-name normalization.
7.3 Final production review of the 333rd event
The two remaining findings in the complete 333rd-event review received explicit decisions:
2189 野翔太郎: kana, Hosei history, and performance continuity matched; the source name was a missing-﨑/崎variant. Both temporary records and results were merged into野﨑翔太郎.2352 佐藤広人: only the reading matched佐藤大翔, whose affiliation conflicted;佐藤広知had a different reading. A separate new person was kept instead of forcing a merge.
Both findings were audited and marked resolved; the task summary ended with unresolved=0. This means every retained conflict received an explicit business decision. It does not imply that every decision was made automatically by the Graph; manual review and data corrections remain independently audited.
Round closure
The stable V2 boundary is now: write complete race results and source labels first, then route through rules, lightweight assessment, one bounded cross-check, and necessary heavy research. If research fails or evidence is insufficient, retain the finding and context rather than dropping data or promoting a temporary race label into a formal Organization. A new event such as the East Japan Corporate Ekiden starts a new run and optimization cycle; it should be recorded in a new document rather than rewriting this history.
Next observation
Append problem–case–change–evidence after each review and backport stable requirements/design into their documents. Separate offline tests, deployment verification and actual runs. Assess explainable general rules, entity-specific aliases and remaining semantic decisions without requiring a new rule every round.
The next user-led first-group run should report rule evidence, lightweight organization/person counts, cross-check rejections and corrections, affected reassessments, actual heavy objects/requests, final planned/applied changes, unresolved items, latency, tokens and known cost. This delivery stops after deployment and documentation.