[strickvl]

Requested by Alex · project thread

What changes

This follows #64, which added the fact-list updater as an eval-only prototype.

• Pipeline: update_profile(approach="facts") and HINBOX_PROFILE_UPDATE_APPROACH=facts make just process use src/engine/fact_list.py. The default stays edit.

• Saved rows: each profile saves its fact rows as a JSON string under profile["fact_rows_json"]. A row records:

  • each source's kind and quote, including for facts the model placed itself (one sentence can state several facts, and then those rows share it);
  • which articles only repeated it;
  • which row a conflict disputes.

  load_rows() reloads the rows on the next update. It drops a saved row whose sentence is no longer in the text, and builds rows from the text for any cited sentence that no saved row covers.

• Timelines: the compare step gets an update verdict for a later state of the same thing, such as a move to another camp, a decision after "still deciding", or a rescheduled trial date. It is written next to the old state, not as "Another report differs".

• Fewer false conflicts: a conflict must also pass same_subject(), which means the two sentences share at least 40% of their word stems, ignoring numbers.

• Details no longer dropped: a claim judged same that has a year, month or number the row lacks ("a 78-minute secret session") is written as a new fact instead of being dropped.

• Removed checks:

  • 5795add tried a looser conversion of same and update verdicts into conflicts. On test it brought back 6 false disputes per run and still caught 0 of 4 planted disputes, so it is reverted.
  • The earlier update-to-conflict check (different dates plus 60% shared stems) never fired on dev or test. Review showed it could turn a rescheduled trial date into a false dispute, so it is removed too.

Results

All runs use local Qwen through llama-server. Every run had 0 errors.

Eval, on the cc1158f code. That version still had the update-to-conflict check, but the check never fired, so the scores apply to this PR. Scores are cases passed:

	dev (81)	test (79)
Luna, facts	79	75
Luna, edit	75	74
Qwen-32k, facts	77	73
Qwen-32k, edit	69	67

Real just process run on the same 60 articles, edit against facts, counted by scripts/audit_entities.py. This run used 989d194, before the last fixes.

	edit	facts
Repeated sentences	320	23
"Sources disagree" sentences	312	4
Longest profile	55k chars	27k
profile.update seconds	10,256	9,212
Rows with a linked but uncited article	13	121

I sampled 11 of the 121 rows with linked but uncited articles:

• In 8, the article adds nothing new: it's a passing mention or a repeat.
• In 2, a small detail was lost. The same_with_detail change fixes that.
• 1 is unclear.

Known gap

Planted disputes on test are 0 of 4 for both models. The models call a different year for a past event "the same" or "an update". Fixing that inside the updater brought back false disputes, so it needs a separate check after the update, with its own eval.

Testing

uv run pytest -q tests -m "not asyncio": 590 passed after merging main. ruff is clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_017KuZSfznoSE6bnm9t8KWMf

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-04T05:26:27.093832Z	b52b25d	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 24d42adc5f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 52868be9cb

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".