feat(claims): extract stated event intervals with each bound's own precision (#717, slice 4) by cdeust · Pull Request #926 · cdeust/Cortex

cdeust · 2026-11-07T02:35:28Z

Summary

Tranche 4 de #717 : stated_event_time lisait toute phrase contenant une formulation d'intervalle comme unknown. Elle lit désormais les intervalles bornés par des dates complètes, chaque borne gardant sa précision et son fuseau écrits :

from D1 to|until|till|through D2 et between D1 and D2 → intervalle fermé ;
since D → début connu, fin inconnue ; until|till D → fin connue, début inconnu.

Cause racine : la tranche 3 avait volontairement laissé _RANGE_RE court-circuiter toute la lecture (« stays unknown until intervals are extracted »), alors que EventTime (tranche 1) et le stockage (tranche 2) portent déjà kind="interval" avec bornes optionnelles.

Ce qui reste unknown, pour ne pas inventer de date : les intervalles abrégés (« from 10 to 12 March 2026 », « 10-12 March 2026 ») dont la première borne ne donne ni mois ni année ; une fin avant le début ; une paire où une seule borne indique un fuseau (refusé par EventTime) ; toute autre formulation.

Refs #717

Type of change

Bug fix (non-breaking change which fixes an issue)
New feature (non-breaking change which adds functionality)
Breaking change (fix or feature that would change existing behavior)
Refactor (no functional change; rules/coding-standards.md compliance)
Documentation only
Audit-finding closure (cite the finding ID)

Test plan

pytest tests_py/core/test_claim_event_time.py : 75 passés. Quatre phrases qui étaient attendues unknown (« Since 2026-01… », « …until March 2026 », « From 2026-01 to 2026-03… », « Between March 2026 and 4 April 2026… ») sont maintenant des intervalles, comme voulu. Nouveaux cas négatifs : fin avant début, fuseau sur une seule borne, mauvais mot de liaison, trois dates, date impossible, date en tête de phrase.
Tests voisins (-k "claim or event_time or wiki_extract or mutation" sur tests_py/core et tests_py/handlers) : 140 passés, 1 ignoré.
Porte de mutation locale (scripts/mutation_check.sh sur claim_event_time.py) : 204 mutants, 203 tués, 1 survivant enregistré comme équivalent par construction (x__word_before__mutmut_8, valeur par défaut "" → "XXXX", aucune des deux n'est un mot de cadre).
ruff check / ruff format --check (0.17.3, version de uv.lock) sur les fichiers modifiés, porte craftsmanship OK.
Pas vérifié : la suite complète, pyright et les variantes PostgreSQL. torch ne se télécharge pas dans l'environnement de la routine. La CI fera foi.

Audit notes

Couches : changement limité à core (+ shared.wiki_ir déjà importé), pas de nouvel import.
Deferred : porter event_time dans la synthèse et la recherche (wiki_view ne filtre pas encore par event_time, et la sémantique de comparaison entre précisions et fuseaux reste à décider), plus les intervalles abrégés. Suivi dans feat: preserve event time separately from extraction time in ClaimEvent #717.

Coding-standards compliance

§2.2 Layer dependency direction preserved.
§3.2 No untyped dicts at boundaries.
§4.1 No file > 500 lines.
§4.2 No function > 50 lines.
§4.4 No function with > 4 parameters.
§7 Local reasoning preserved.
§8 Constants annotated with # source:.
§9 No dead code.

Breaking changes

Aucun pour le schéma : event_time accepte déjà interval depuis la tranche 1. Certaines claims qui sortaient unknown sortent maintenant interval.

Reviewer checklist

CHANGELOG.md updated under the appropriate section.
Documentation updated (README / SKILL.md / docs/). Module docstring mise à jour, pas d'autre doc concernée.
No secrets / credentials / PII in the diff.
CI passes on the latest commit.

Generated by Claude Code

…ecision (#717, slice 4) stated_event_time left every sentence with interval wording unknown. It now reads "from D1 to|until|till|through D2" and "between D1 and D2" as a closed interval, "since D" as a start with an unknown end, and "until|till D" as an end with an unknown start. Each bound keeps its written precision and zone. Abbreviated ranges ("from 10 to 12 March 2026", "10-12 March 2026") stay unknown: their first end names no month or year, and copying them from the second end would invent them. An end before its start, or a pair where only one end states a zone, is rejected by EventTime and stays unknown. The empty-default mutant of _word_before is registered as an equivalent. Refs #717 Co-Authored-By: Claude Opus 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_018kyZYs2Nms1G85LP4zC4Lk

chatgpt-codex-connector · 2026-11-07T02:35:39Z

Codex Review Summary: ✅ Completed

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

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-07T02:41:31.601737Z	`7751462`	New commits
🔒 Security Review	✅ Completed 2026-11-07T02:42:27.196930Z	`7751462`	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.

claude · 2026-11-07T02:35:59Z

Claude finished @cdeust's task in 2m 5s —— View job

PR #926 review (1 of 3)

Read the diff and the EventTime validators
Ran the five reviews in the foreground (code quality, docs accuracy, performance, security, test coverage)
Posted the review as a top-level comment: feat(claims): extract stated event intervals with each bound's own precision (#717, slice 4) #926 (comment)

I found no Blockers and posted three Notes. I couldn't run pytest or the craftsmanship script because Python needs approval in this sandbox, so CI will have to confirm them.

Verdict: approve

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8d1725205a

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 · 2026-11-07T02:37:07Z

+        if word in _OPEN_START_WORDS:
+            return EventTime(kind="interval", end=bound)


Reject abbreviated closed ranges before accepting an open bound

For We used Redis from 10 until 12 March 2026., only the second endpoint matches a full date, so this branch returns an interval with an unknown start and an end of March 12. The same happens with till. These are abbreviated closed ranges, which the documented contract explicitly leaves unknown, not genuinely open-start intervals. This silently drops the stated lower endpoint and persists a different temporal meaning. Before accepting a single-bound frame, check that it is not the surviving endpoint of an unsupported closed range, and add regression cases for both connectors.

Useful? React with 👍 / 👎.

Confirmé et corrigé dans 830568c. Sur 8d17252, « We used Redis from 10 until 12 March 2026. » (et la même phrase avec till) donnait un intervalle à fin seule au 2026-03-12, ce qui perdait la borne de début écrite. Une borne seule n'est désormais ouverte que si son mot-cadre est la seule formulation d'intervalle de la phrase (_RANGE_RE y trouve une seule correspondance). Ces deux phrases, plus « Since 10 until 12 March 2026… », restent unknown. « We worked from home until March 2026. » reste une fin seule, car « from home » n'est pas une formulation d'intervalle. Les 79 tests du module passent. Porte de mutation locale : aucun survivant non enregistré.

Generated by Claude Code

claude · 2026-11-07T02:38:04Z

Review 1 of 3 — PR #926 (#717, slice 4)

I found no Blockers. Five reviews ran (code quality, docs accuracy, performance, security, test coverage), all by reading the code:

Layering: the change stays inside core and imports only shared.wiki_ir.
Craftsmanship limits: the file is 196 lines and every function is under 40 lines. _CLOSED_ENDS has its # source: comment.
Rejections: when EventTime refuses an interval (end before start, or a zone on only one end), it raises a pydantic ValidationError, which is a ValueError. The try catches it and the result is unknown, as the docstring says.
Security: none of the regexes can backtrack badly, and no other exception gets out of stated_event_time.
Mutation note: x__word_before__mutmut_8 really is equivalent, and the test sentence it names exists.
Other tests: no other test expects unknown for interval wording.

I couldn't run pytest or the craftsmanship script here because Python needs approval in this sandbox, so CI will have to confirm them.

Notes, no action before merge

If the end bound is coarser than the start, wiki_ir.py:80 compares only the two values and rejects a valid interval. For example, "from 2026-03-12T10:00 to 2026-03-12" gives unknown; it fails safe but the docs don't mention it.
The slice-3 entry in CHANGELOG.md:62-63 still says "since 2026-01" stays unknown, which slice 4 changes.
No test covers from…till, or a lone through D (which sends the sentence to the interval path and then gives unknown).

View job run

Verdict: approve

… own precision (#717, slice 4) Codex review on 8d17252 (PR #926): in "from 10 until 12 March 2026" only the second end is a full date, so the single-bound branch returned an end-only interval and dropped the stated start. A single bound is now open only when its frame is the sentence's only interval wording; the abbreviated closed range stays unknown, as the module contract states. Co-Authored-By: Claude Opus 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_018kyZYs2Nms1G85LP4zC4Lk

claude · 2026-11-07T02:39:50Z

Claude finished @cdeust's task in 1m 47s —— View job

PR #926 review (2 of 3)

Checked the earlier Blockers: review 1 had none.
Reviewed only the commit pushed after 8d17252, which is 830568c (fixup). All five reviews (code quality, docs accuracy, performance, security, test coverage) ran in the foreground.
Posted the review: feat(claims): extract stated event intervals with each bound's own precision (#717, slice 4) #926 (comment)

The fixup fixes the Codex finding: "from 10 until 12 March 2026" now gives unknown instead of an interval with only an end date. I found no Blockers and posted two Notes. Some starts the regex doesn't recognise, such as "from the 10th", still give an end-only interval. The CHANGELOG wording for "since/until" doesn't say those forms now need to be the sentence's only interval wording.

I couldn't run pytest because Python needs approval in this sandbox, so CI will have to confirm the tests.

Verdict: approve

… own precision (#717, slice 4) Claude review 1 of 3 on PR #926: the slice-3 CHANGELOG entry still said "since 2026-01" stays unknown, which slice 4 changes; and no test covered from...till or a lone "through D" (interval wording with no open-bound frame, so unknown). Co-Authored-By: Claude Opus 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_018kyZYs2Nms1G85LP4zC4Lk

cdeust · 2026-11-07T02:40:34Z

Suite à la revue 1 of 3 (Claude), dans 7751462 :

CHANGELOG : l'entrée de la tranche 3 disait encore que « since 2026-01 » reste unknown. Corrigée : elle renvoie à la tranche 4 pour les intervalles à dates complètes.
Tests : ajout de from…till (intervalle fermé) et d'un « through D » seul (unknown, car « through » n'ouvre aucune borne seule). 81 tests du module passent.
Laissé tel quel : une fin plus grossière que le début (« from 2026-03-12T10:00 to 2026-03-12 ») est refusée par EventTime (wiki_ir.py:80 compare les valeurs, pas les périodes) et sort unknown. Le résultat est sûr, puisqu'aucune date n'est inventée. Comparer des bornes de précisions différentes est justement la décision ouverte notée sur feat: preserve event time separately from extraction time in ClaimEvent #717 pour la recherche, donc je ne la tranche pas dans cette PR.

Generated by Claude Code

claude · 2026-11-07T02:40:47Z

Claude finished @cdeust's task in 1m 34s —— View job

PR #926 review (2 of 3)

Checked the Codex finding against 830568c / 7751462. It is fixed: "from 10 until/till 12 March 2026" now gives unknown.
Ran the five reviews in the foreground (code quality, docs accuracy, performance, security, test coverage) on the commits after 8d17252 only
Posted the review as a top-level comment: feat(claims): extract stated event intervals with each bound's own precision (#717, slice 4) #926 (comment)

I found no Blockers and posted three Notes:

"From mid-February until 12 March 2026" still gives an interval with only an end, because a start that isn't a digit or a month name isn't counted.
Unrelated range wording elsewhere in a sentence now makes a "since"/"until" sentence unknown. This fails safe.
The module docstring doesn't mention the new rule; only the inline comment does.

I couldn't run pytest or the mutation gate because Python needs approval in this sandbox, so CI will have to confirm them.

Verdict: approve

claude · 2026-11-07T02:41:34Z

Review 2 of 3 — PR #926 (#717, slice 4)

Scope: only the commit pushed after 8d17252, which is 830568c (fixup). Review 1 found no Blockers, so there was nothing to re-check. The five reviews (code quality, docs accuracy, performance, security, test coverage) all ran by reading the code and tracing the regexes by hand. Python needs approval in this sandbox, so pytest was not run; CI will have to confirm the results.

The fix works for the Codex finding. For "from 10 until|till 12 March 2026", _RANGE_RE.findall returns two matches ("from 1" and "until 1"), so the new == 1 check fails and the sentence stays unknown (claim_event_time.py:165).

All three new negative tests kill a mutant that deletes the check or changes == 1 to >= 1.
The positive test "from home until March 2026" checks that the guard is not too strict.
The change adds no backtracking risk and no new exceptions, and the craftsmanship limits still hold: _interval is 21 lines and the file is 198 lines.

I found no Blockers.

Notes, no action before merge

Starts that _RANGE_RE does not recognise still give an end-only interval and drop the stated start: "from the 10th until 12 March 2026", "10 until 12 March 2026", "from Monday until…". The reverse also happens: "Since January 2026 to June" gives a start-only interval. These belong to the deferred abbreviated-interval work in feat: preserve event time separately from extraction time in ClaimEvent #717.
The slice-4 entry in CHANGELOG.md says "since D" / "until D" without conditions. Since this commit, those forms only work when they are the sentence's only interval wording, so "Since 2026-01, output rose from 3 to 5" now gives unknown.

View job run

Verdict: approve

cdeust · 2026-11-07T02:41:50Z

Revue 2 of 3 (Claude) : les deux notes sont prises en compte, sans nouveau push.

Débuts non reconnus (« from the 10th until 12 March 2026 », « 10 until 12 March 2026 », « from Monday until… ») : ils donnent encore une fin seule. Les distinguer de « from home until March 2026 », qui doit rester une fin seule, demande de reconnaître les débuts abrégés (ordinaux, jours de la semaine, nombres nus). Cela fait partie de la tranche « intervalles abrégés » listée sur feat: preserve event time separately from extraction time in ClaimEvent #717, je l'y reporte.
CHANGELOG : la condition « seule formulation d'intervalle de la phrase » sera précisée dans l'entrée de la tranche 4 au prochain push de code. Le cas cité (« Since 2026-01, output rose from 3 to 5 ») sort unknown, ce qui est sûr.

Generated by Claude Code

claude · 2026-11-07T02:42:23Z

Review 2 of 3 — PR #926 (#717, slice 4)

There were no earlier Blockers to check. This review covers only the commits pushed after 8d17252 (830568c, 7751462). I traced the code and ran five reviews: code quality, docs accuracy, performance, security and test coverage.

Codex P2 is fixed. "We used Redis from 10 until 12 March 2026." now gives two _RANGE_RE matches ("from 10" and "until 12"), so the result is unknown. The till sentence and "Since 10 until 12 March 2026…" also give unknown. "We worked from home until March 2026." still gives an interval with only an end.

Other checks are clean:

Size and layers: claim_event_time.py is 199 lines, _interval is 21 lines, there are no new constants and no new imports.
Security and performance: the findall runs only when the sentence has one date and _RANGE_RE already matched. _RANGE_RE has no nested quantifiers. No new exception can escape the try.
Tests: they pin the rule in both directions (two interval wordings give unknown, one gives an open bound), with a test for each of until and till.

I could not run pytest or the mutation gate because Python needs approval in this sandbox, so the results above come from reading the code and CI will have to confirm them.

Notes, no action before merge

The undercount is still open when the start is not a digit or a month name. "From mid-February until 12 March 2026 we ran trials." gives one match, so it still becomes an interval with only an end and loses the stated start.
Unrelated range wording elsewhere in a sentence now blocks an open bound. "Since March 2026 we grew from 3 to 5 sites." gives unknown where 8d17252 gave a start-only interval. This fails safe.
The single-wording rule is only in the inline comment at claim_event_time.py:163, not in the module docstring at lines 28–35.

Verdict: approve

cdeust · 2026-11-07T02:42:37Z

Deuxième passage de la revue 2 of 3 (sur 7751462) : ses trois notes recoupent celles auxquelles j'ai déjà répondu.

Débuts non reconnus (« From mid-February until 12 March 2026… ») : je les reporte à la tranche « intervalles abrégés » de feat: preserve event time separately from extraction time in ClaimEvent #717, comme « from the 10th ».
Phrase avec une autre formulation d'intervalle (« Since March 2026 we grew from 3 to 5 sites. ») : elle sort unknown, ce qui est sûr et voulu.
Docstring : la règle « seule formulation d'intervalle » sera ajoutée à la docstring du module en même temps que la précision du CHANGELOG, au prochain push de code. Ce n'est pas bloquant, donc je ne pousse pas que pour ça.

Generated by Claude Code

cdeust mentioned this pull request Nov 7, 2026

feat: preserve event time separately from extraction time in ClaimEvent #717

Open

3 tasks

chatgpt-codex-connector Bot reviewed Nov 7, 2026

View reviewed changes

github-actions Bot merged commit 16ab975 into main Nov 7, 2026
18 checks passed

		if word in _OPEN_START_WORDS:
		return EventTime(kind="interval", end=bound)

Conversation

cdeust commented Nov 7, 2026

Summary

Type of change

Test plan

Audit notes

Coding-standards compliance

Breaking changes

Reviewer checklist

Uh oh!

chatgpt-codex-connector Bot commented Nov 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

claude Bot commented Nov 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR #926 review (1 of 3)

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Nov 7, 2026

Choose a reason for hiding this comment

Uh oh!

cdeust Nov 7, 2026

Choose a reason for hiding this comment

Uh oh!

claude Bot commented Nov 7, 2026

Review 1 of 3 — PR #926 (#717, slice 4)

Uh oh!

claude Bot commented Nov 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR #926 review (2 of 3)

Uh oh!

cdeust commented Nov 7, 2026

Uh oh!

claude Bot commented Nov 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR #926 review (2 of 3)

Uh oh!

claude Bot commented Nov 7, 2026

Review 2 of 3 — PR #926 (#717, slice 4)

Uh oh!

cdeust commented Nov 7, 2026

Uh oh!

claude Bot commented Nov 7, 2026

Review 2 of 3 — PR #926 (#717, slice 4)

Uh oh!

cdeust commented Nov 7, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

chatgpt-codex-connector Bot commented Nov 7, 2026 •

edited

Loading

claude Bot commented Nov 7, 2026 •

edited

Loading

claude Bot commented Nov 7, 2026 •

edited

Loading

claude Bot commented Nov 7, 2026 •

edited

Loading