Repository navigation
要約の原文照合と修復の安定性を改善 - #25
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b5fe1b12ad
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".
|
@codex review 前回のレビュー対象
意味検証、必須判定、修復最大1回、最終上限20件を維持しています。候補v2・確定v4・事実ID・要約キャッシュv34の意味契約は変更していません。新しい例外や属性の無条件補充、旧診断へのID補完は追加していません。 検証結果:
特に、容量調整後の必須保持と修復範囲、smartの選択境界、現在の失敗と保存復元の診断対応に漏れがないか再確認をお願いします。判断・反例・未達条件は 開発記録 と 評価記録 に追記しています。 |
|
Codex Review: Didn't find any major issues. Delightful! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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". |
|
@codex review 前回レビュー対象
検証は462テスト、型チェック・lint・ビルド成功。最終版Gの同PDF・同モデル・full・各3回固定の実LLM生成は初回9/9、47事実の原PDF照合で観測した誤受理・必須欠落0、保存9結果の厳密再判定も通過。BlueMemeは最終ビルドで実TDnet・実API1回/26秒、表示・診断出力・保存復元まで成功しました。設定変更中の応答を保留する固定API実画面も、新要求だけの表示/保存と再操作を確認しました。 途中版Fの成功9回は、後から括弧内の境界の誤受理を発見して修正したため、Gの結果へ混ぜていません。複数期間の本文分解・未知形式・過去日の実TDnet経路は未達のままです。Gで実モデル修復は発生せず、smart実生成・後続分析は再実行していません。詳細な反証と証拠は |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8f56cccaf2
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".
|
@codex review 再レビュー対象
475テスト/25ファイル・型チェック・lint・ビルド成功。修正版Hの同PDF・OpenRouter/deepseek-v4.1-flash・full・各3回固定の新規生成は、必須を満たす初回成功9/9。48確定事実の原文照合で観測した誤受理・必須欠落0、保存Hの9結果/48事実とGの9結果/47事実の厳密再判定も通過しました。最終ビルドのBlueMemeは実TDnet/実API1回・29秒で表示・診断出力・保存復元まで成功。固定APIの復元診断照合と、次の生成前失敗で旧traceを拒否する経路も確認しています。 一部補足候補の状態や根拠は証明できず未確認に残っており、全候補の受理率とは区別しています。Hでは実モデル修復が発生せず、smart実生成・後続採点/分析も今回は再実行していません。未知形式・多期間本文の分解・過去日実TDnet・利用者の当時のログ等の限界は従来どおり明記しています。 |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 93cf139c60
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".
|
@codex review 前回のレビュー対象
指摘への返信、正本の検証記録、PR本文を更新しました。差分の範囲・基準は前回の成功回数を今回の新規生成実績へ混ぜずに記載しています。 |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e826cda3c4
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".
|
@codex review レビュー対象
追加表記の実PDF/新規実モデル/実APIと、Qだけの期間軸・累計/単独の解釈は未確認です。正本とPR本文に検証範囲・証拠を記録しています。 |
|
Codex Review: Didn't find any major issues. Already looking forward to the next diff. Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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". |
|
@codex review 追加セルフレビューの2件を修正しました。基準
検証は574テスト・type-check・lint・build成功、最終保存反例の強化後も対象33件・型・対象lint成功です。同条件の既存実LLM証拠は最終ソースで9結果/48事実を再判定し、全件受理しました。最終ビルドでは実PDF抽出+固定APIで拒否・修復・局所属性表示・保存復元・診断対応・後続利用を確認し、実TDnet/実APIでも1回の生成で11事実の表示・保存復元が成功しています。条件と未確認事項は 特に、後続文の未検査部分を確定額扱いしていないこと、局所属性を既定値で上書きしていないこと、候補・必須判定・保存・後続利用の間で意味判定が一致することをご確認ください。 |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 352943b5f6
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".
|
@codex review 最新レビューの4件を修正しました。基準
恒久回帰は追加10件、分析指紋はv39→v40です。旧v38/v39の誤受理・未確認欠落キャッシュを表示しません。最終585テスト/26ファイル、type-check・lint・buildが成功しました。 既存集合Hの保存9結果/48事実は最終コードで厳密再判定し、全受理・拒否0。必須slotの出典変更でBlueMemeの通常生成入力が変わったため、新規集合Jを重点3資料各3回へ事前固定し、最終は各3/3、API10回で成功しました。BlueMemeの1回では不要な配当属性を初回で拒否し、確定10件のIDを保持して差分1回で11件へ修復しています。 最終ビルドでは実Offscreen/PDF.js+固定APIの拒否・括弧否定event修復・未確認表示/保存・番号付き欄の受理・局所属性表示・保存復元・診断・後続利用を確認しました。実TDnet/PDF+実APIでも1回/22秒で11事実の表示・保存復元・診断保存が成功しています。条件・hash・利用量・結果は 特に、必須値と修復出典が同じ報告単位へ対応すること、局所属性を表紙で上書きしないこと、未解消の警告を消さず解消済み診断だけを除去すること、生成・保存・表示・後続利用の判定が一致することをご確認ください。 |
|
@codex review 最新4件をすべて採用し、66928dbへ反映しました。各スレッドに対応内容を返信しています。 指摘の傾向は、形・限定・対象述語を別々の原文から組み合わせることと、共有関数でも接頭辞一致を完全な意味証明とみなすことです。今回は以下を共通の責務へまとめました。
検証: 基準89ba53eで最終恒久90件中12失敗、修正後873件/27ファイル・type-check・lint・build成功。14ケース+旧v54非表示は既存matrixへ組み込み、新しいテストファイルなし。6原PDF/29固定事実を保持。分析指紋v55。 BlueMemeだけ区間両端13組の日付役割で生成入力が変わるため、新規3回に限定しすべて初回成功・33事実/修復0。自己株取得/月次O6は既存応答を再利用し、O6+Q3を最終コードへ完全再生。旧38保存結果の68拒否は旧BlueMeme17結果×4予想/EPSの旧日付役割だけで、旧P3生応答の再生ではその他の全属性・初回判定を保持。新Q3はstrict保存33事実全受理です。 最終ビルドの固定17事実と実TDnet/実API11事実で拒否/修復/表示/保存復元/診断保存を確認しました。追加APIはCLI3+画面1に限定、再表示0。条件・結果・digest・意図した旧保存拒否・未確認事項はdocs/development.mdのT1〜T4とevaluation/README.mdへ記録しています。未知形式/OCR/多主張分解等は未確認のままです。 |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 66928dbd2f
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".
|
@codex review 最新HEADは926115dです(製品修正0c07142、原文説明の記録訂正926115d)。 最新5件を全て採用し、66928db → 0c07142 で修正しました。各スレッドへ対応内容と検証を返信しています。 今回の傾向は、同じ段落/ページの情報を対応証明なしに共有していたことです。予定状態は述語/明示日程、期間は局所FYと形を所有する原文、文書発行者は表紙領域、同段落のevent/statusは事実の種類に責務を揃えました。予想修正見出しは全文一致を保って完全な末尾を組み合わせます。 追加探索で、予定の取消接続形/継続形、可能性だけの正常予定、数値状態と日付役割、局所会社3形式、event/statusの順序とdelta修復・別種類で診断を消す経路まで対応しました。安定ID算法/候補v2/確定v4/最大1修復/20件上限は維持。分析指紋v56で旧v55キャッシュを除外します。 恒久回帰111件は基準で19失敗、最終895テスト/27ファイル・type-check・lint・build成功。最終固定19ページ/20事実と実API11事実の表示/保存復元/診断確認成功。入力変更したBlueMemeのみ新規3回/全初回成功、他資料の6生応答は再利用し、診断差1件を明記しました。詳細は正本T1〜T4とevaluation/README.md。未知形式/OCR/多主張分解等は未確認として保持しています。 |
There was a problem hiding this comment.
💡 Codex Review
tdnet-digest/src/lib/fact-coverage.ts
Lines 471 to 474 in 926115d
When an M&A disclosure says 株式を取得し子会社化することを決定いたしました, the outer M&A gate matches but this obligation is created only for 決議しました forms. Although assertion validation already recognizes 決定しました as decided, coverage can pass without the core acquisition event whenever another key fact is present, allowing the summary to omit the document's principal decision. Include the supported 決定 predicate forms in the acquisition-decision obligation.
tdnet-digest/src/lib/fact-coverage.ts
Lines 402 to 403 in 926115d
A validated revision-table fact may use the schema-supported period 2028年3月期(予想), but after normalization this regex accepts only a bare 予想 suffix and not (予想). Such correct forecastBefore and forecastAfter facts are excluded from candidates, so every revision obligation remains missing and the one repair attempt cannot produce an acceptable summary without changing the source-backed period spelling. Compare the fiscal period semantically or accept the same parenthesized form allowed by fact validation.
tdnet-digest/src/lib/metric-semantics.ts
Lines 30 to 33 in 926115d
The closed unit grammar accepts currencies such as JPY, GBP, and CNY, including scaled forms such as 千JPY, but this classifier recognizes only yen text, dollars, USD, and EUR. Consequently a validated 売上高 or profit value expressed in one of the other supported currencies receives metricKind=other; standardMetric then excludes it from earnings coverage, so required results cannot be satisfied. Share the currency set with the accepted unit grammar instead of maintaining a narrower list here.
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".
|
@codex review 最新8指摘をすべて採用し、926115d → 2ba1477で対応しました。各スレッドへ原因・変更・検証を返信しています。 共通する傾向は、語の存在と主張/原文単位への対応を別々に決めることでした。局所宣言の切替、正本発行者、完全な予想見出し、予定の日付/主張、区間、極性、対象義務、修復診断を既存の意味責務へ揃えました。同原因の見出しなし会社切替、局所scope/basis切替、取消済み実行日、否定/別範囲の未定statusも先回りして修正しています。 今回の差分:
指紋v57。基準恒久145件中31失敗→現行930件/27ファイル・type-check・lint・build成功。既存matrixへ34例+旧キャッシュ1例、新規恒久ファイルなし。最終固定19ページ/20事実と実TDnet/実API11事実で表示/保存復元/診断対応成功。3資料の生成入力は不変、既存9応答の最終再生は前回結果と完全一致し、追加CLI生成0回。保存44結果の受理/拒否も不変です。 正本docs/development.mdのT1〜T4とevaluation/README.mdに条件・結果・中間失敗・未確認範囲を記録しました。今回の責務を跨ぐ同原因の誤受理/誤拒否、正常例の維持を中心に再レビューをお願いします。 |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 2ba1477f18
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".
|
@codex review 最新5指摘をすべて採用し、2ba1477 → 4b190adで対応しました。各スレッドに原因・変更・証拠を返信しています。 今回の傾向は、継承する属性の所有する節、取消が効く節、必須項目の報告単位、修復メタデータの境界を一律に扱うことでした。原文の階層/節/報告単位へ判定を揃え、隣接経路も調べています。
同原因の子会社利益率によるissuer義務の作成/代替、契約という語への言及だけで日付を締結済みとする経路も修正しました。 指紋v58。基準恒久162件中16失敗→現行948件/27ファイル・type-check・lint・build成功。既存matrixへ17例+旧キャッシュ1例、新規恒久ファイルなし。最終固定19ページ/20事実と実TDnet/実API11事実で表示/保存復元/診断対応成功。入力が変わった自己株取得/BlueMemeだけ各3回を事前固定して全て初回成功、月次は既存3応答を完全再生して再生成0回。新6保存42事実はstrict全受理、旧44保存の判定も不変です。 正本docs/development.mdのT1〜T4とevaluation/README.mdに条件・結果・中間失敗・未確認範囲を記録しました。今回の責務を跨ぐ同原因の誤受理/誤拒否と、正常例の維持を中心に再レビューをお願いします。 |
There was a problem hiding this comment.
💡 Codex Review
tdnet-digest/src/lib/fact-coverage.ts
Lines 480 to 484 in 4b190ad
When a buyback disclosure for issuer A also contains a locally scoped company B, an A share-count fact and a B acquisition-amount fact satisfy both required buyback slots: this predicate checks only that subject is nonempty. The summary can therefore omit A's acquisition ceiling while coverage reports success; compare the subject and source ownership with documentSubject(context) for each required buyback fact.
tdnet-digest/src/lib/fact-coverage.ts
Lines 559 to 560 in 4b190ad
When a target-company table continues onto another page but discloses only revenue and operating profit (for example, its second continuation row is employee count), tableContinuations still creates a link and this unconditional loop requires netProfit as well. No candidate or repair can satisfy that absent source row, so the entire M&A summary fails; create each metric obligation only when the linked table actually contains a provable row for that metric.
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".
|
4b190adのレビューの7件を、bfd83e8で対応しました。5スレッドへの個別返信に加え、レビュー本文の2件も修正しています。
傾向は、局所的に正しい事実と必須の報告単位の混同、義務・充足・修復の別判定、非同期処理の所有者の欠落でした。表紙/予想宣言の報告対象を共有し、局所属性を上書きせずに充足を照合します。関連経路の予想修正・月次他社値/速報/KPI・選択外の先行予想宣言にも対応しました。取消の辞書形と数量文の見出し誤認は共有文法へ、診断保存は要求開始順の所有者と直列キューへ揃えています。 既存matrixへ正誤20件、診断競合3件、旧v58非表示1件を追加し、新しい恒久テストファイルは増やしていません。表の候補fixtureを共通化し、古いheadline正常例の表紙属性を対応する属性へ整理しました。候補v2/確定v4/安定ID/最大1修復/20件上限を維持、分析指紋はv59です。 検証: 972テスト/27ファイル・type-check・lint・build成功。基準の23ケースは22失敗/正常1成功。新規BlueMeme3回+月次3回は全初回成功・修復/最終失敗0。自己株取得3応答を再利用し、最終9応答の入力/事実/警告/成否/期待値が完全一致。旧50保存結果の285受理/68既知拒否は不変、新6結果40事実は全受理。最終buildの拒否→修復/局所表示/保存復元/診断/後続利用、設定変更での完了逆転、実TDnet/PDF/API画面も成功。途中失敗・旧L停止・未確認範囲は正本へ残しています。 記録: docs/development.mdの既存T1〜T4、evaluation/README.mdの4b190ad条件と最終結果。 @codex review 特に、必須単位と修復先で対象期・属性・選択範囲・原文役割が一致すること、取消と後続予定の分離、M&Aの開示行だけの義務、最新要求の診断保持をご確認ください。未知形式/OCR/多主張分解などの既存未確認範囲を、今回の修正で保証済みとは扱っていません。 |
There was a problem hiding this comment.
💡 Codex Review
tdnet-digest/src/lib/fact-coverage.ts
Lines 719 to 722 in bfd83e8
If the issuer's acquisition-decision fact is omitted but a later target-company section contains its own validated 株式取得...決議しました event, this predicate accepts that target-company fact because it checks only kind, state, and quote text. Coverage can therefore pass without the issuer's decision; require the fact's subject and evidence owner to match documentSubject(context) as the share-repurchase checks do.
tdnet-digest/src/lib/fact-coverage.ts
Line 632 in bfd83e8
The trigger is evaluated against all selected text, so an 上限 appearing only in a subsidiary or other-company section creates both issuer count and amount obligations. Since satisfaction and repair sources are subsequently issuer-filtered, such a document can be forced into an unrepairable missing obligation that the issuer never disclosed; derive this trigger from issuer-owned source blocks instead of the page-wide string.
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".
| const latest = link.periodIds | ||
| .map((id) => normalized(text([id]))) | ||
| .sort() | ||
| .slice(-1)[0]; |
There was a problem hiding this comment.
Compare M&A periods chronologically
When a target company changes its fiscal year-end, periods in the same calendar year are ordered incorrectly as strings: for example, sorting 2027年3月期 and 2027年12月期 makes the March period the last element because 3 sorts after 1. The resulting obligations and repair sources target the older column instead of the latest disclosed period, so parse the year/month and compare them numerically.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
妥当と判断し、5ba3d3fで修正しました。M&A継続表の最新期は年・月を数値比較し、義務・充足・修復先で同じ最新列を使います。同年3月/12月の反例を既存matrixへ追加し、生成修復と保存再照合で一致を確認しました。
| (f) => | ||
| f.kind === 'number' && |
There was a problem hiding this comment.
Match monthly KPI facts to the linked issuer source
For an issuer KPI obligation created by noteLinks, a validated fact from another company with the same metric label and month satisfies this test because neither its subject nor its evidence source is checked. In a document containing issuer and subsidiary monthly tables, the issuer's linked KPI may consequently be omitted; constrain the fact to the issuer and the source unit represented by the link.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
妥当と判断し、5ba3d3fで修正しました。主要KPIは事実と原文の発行者に加え、noteLinksが結ぶ同じ根拠数量IDで充足させます。修復も同じ結合から根拠を列挙します。他社の同月同名値と、発行者自身の別原文の値を拒否する対を恒久化し、正しいリンク元の候補・保存は維持しました。
| .flatMap((p) => p.blocks) | ||
| .filter((b) => isLossRecordingPlan(b.text)); | ||
| const plannedAttributes = attributesFor(plannedLossBlocks); | ||
| if ( | ||
| /特別損失に計上.*予定/.test(source) && | ||
| plannedLossBlocks.length > 0 && |
There was a problem hiding this comment.
Keep subsidiary loss plans from creating issuer obligations
When an earnings release contains a subsidiary-owned paragraph stating that the subsidiary plans to record a special loss, this unscoped collection creates the loss-plan obligation, but the later satisfaction predicate requires the fact's subject to be the document issuer. A truthful subsidiary fact therefore cannot satisfy the requirement, while changing it to the issuer fails source ownership validation, so generation always fails for that document; either restrict obligation sources to issuer-owned blocks or preserve each source block's actual owner in the requirement.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
妥当と判断し、5ba3d3fで修正しました。計上予定の義務と修復先を発行者所有の原文に限定しました。同原因の概算純損失の背景も揃えています。他社だけなら発行者義務を作らず、同じ原文を発行者へ移すと義務が戻る対を候補・保存で確認しました。局所会社の真の主張の意味照合は維持します。
| const scalar = '-?\\d+(?:\\.\\d+)?'; | ||
| const amount = claim.range | ||
| ? `${scalar}[~〜~]${scalar}${escape(claim.unit)}` | ||
| : `${scalar}(?:${claim.unit === '円' ? '円\\d{2}銭|' : ''}${escape(claim.unit)})`; |
There was a problem hiding this comment.
Accept grouped digits in prose quantities
The prose quantity grammar accepts only uninterrupted digits, so a common assertion such as 売上高は1,000百万円です cannot match even though proseQuantities and parseExactQuantity both recognize the grouped value. If a required metric is disclosed only in that prose form, every candidate and repair is rejected; use the same \d{1,3}(?:,\d{3})+|\d+ numeric grammar already used by the quantity parser, including for range endpoints.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
この指摘は基準bfd83e8で再現しなかったため、数量文法の製品変更は採用しません。verifyProseEvidenceのcompactは照合前に桁区切りを除去し、proseQuantities/parseExactQuantityは原文の区切りを保持しています。1,000・全角・負の小数・区間は受理、不正な1,00は拒否を確認しました。3実績の生成→保存を通る代表恒久テストを5ba3d3fに追加し、基準でも成功しています。条件と証拠はevaluation/README.mdのbfd83e8レビュー節に記録しました。
There was a problem hiding this comment.
💡 Codex Review
tdnet-digest/src/lib/fact-coverage.ts
Lines 631 to 633 in bfd83e8
If a selected local section for another company contains 上限 while the issuer does not disclose upper-limit repurchase quantities, this page-wide check still creates both issuer count and amount obligations. The new satisfaction and repair filtering accepts only issuer-owned sources, so those obligations have no truthful repair target and the summary necessarily fails; derive the trigger from issuer-owned blocks just as the later checks do.
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".
| const cover = context.bindings.find((b) => b.page === 1); | ||
| const actual: ReportingTarget | null = | ||
| report && cover ? { ...report, attributes: reportingAttributesAt(cover, true) } : null; |
There was a problem hiding this comment.
Include separate cover fields in the actual reporting target
When a cover title omits its scope or basis but supplies them in following supported fields such as 範囲 連結 or 会計基準 IFRS, the later actual facts validate with those local declarations, while this target reads only document-origin declarations from the first page binding. The target therefore retains null for those attributes, so sameReportingAttributes rejects every required actual and the corresponding repair sources even though they belong to the explicitly declared reporting unit.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
妥当と判断し、5ba3d3fで修正しました。document-contextの連続した表紙領域判定を共有し、値または節の開始前の明示欄だけを必須報告対象へ参照します。表題の明示属性は保持し、その欠落部分を同じ表紙の別欄で証明します。直後に値がある形式と経営成績の節を挟む形式で生成・保存が成功し、後の局所節が表紙対象を上書きしない既存反例も維持しました。
| @@ -260,7 +697,18 @@ export function verifyCoverage( | |||
| ) | |||
There was a problem hiding this comment.
Require the KPI fact to belong to the issuer
When an issuer-owned monthly KPI link exists but only a locally declared subsidiary or comparison company has a validated fact with the same metric and month, this predicate treats that other company's fact as satisfying the issuer obligation because it checks neither the fact's subject nor its source ownership. This can silently omit the issuer's required KPI even though repair-source discovery is restricted to issuer sources.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
妥当と判断し、5ba3d3fで修正しました。主要KPIは事実と原文の発行者に加え、noteLinksが結ぶ同じ根拠数量IDで充足させます。修復も同じ結合から根拠を列挙します。他社の同月同名値と、発行者自身の別原文の値を拒否する対を恒久化し、正しいリンク元の候補・保存は維持しました。
| await setupOffscreenDocument(); | ||
| const extraction = await extractTextFromPDF(data, documentType, mode); |
There was a problem hiding this comment.
Create the trace before fallible request setup
Fresh evidence beyond the covered generation failure is a fetch or extraction failure: currentSummaryRunId is assigned as soon as the request arrives, but the trace is not created or saved until after settings, PDF fetch, offscreen setup, extraction, hashing, and serialization all succeed. If the newest request fails in any of those steps, it prevents an older in-flight request from updating the trace yet writes no trace for its own returned run ID, so the diagnostic button for that failure can only report a run mismatch.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
妥当と判断し、5ba3d3fで修正しました。traceを設定読込前に開始し、設定/PDF取得/offscreen/抽出/hash/入力直列化から最終保存まで同じ失敗処理を通します。未取得属性・hashはnull、API呼出し前のattempts/usageは空です。既存のrunId所有者と直列保存を維持し、新要求の生成前失敗を旧完了が上書きしない競合を追加しました。最終ビルドで復元診断の一致と次の設定失敗自身のJSONダウンロードも成功、追加APIは0でした。
| const latest = link.periodIds | ||
| .map((id) => normalized(text([id]))) | ||
| .sort() | ||
| .slice(-1)[0]; |
There was a problem hiding this comment.
Order M&A fiscal periods chronologically
When a target-company table contains periods with different month numbers in the same year, such as 2027年3月期 and 2027年12月期 after a fiscal-year change, lexical sorting selects March because 3 sorts after the leading 1 of 12. The resulting obligations and repair source IDs are therefore bound to the older column rather than the latest disclosed period; parse and compare the year and month numerically.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
妥当と判断し、5ba3d3fで修正しました。M&A継続表の最新期は年・月を数値比較し、義務・充足・修復先で同じ最新列を使います。同年3月/12月の反例を既存matrixへ追加し、生成修復と保存再照合で一致を確認しました。
| if (/単独/.test(qualified)) return `standaloneQ${q}` as VerifiedFact['semantics']['periodKind']; | ||
| if (/累計|中間期/.test(qualified) || q === '1') | ||
| return `cumulativeQ${q}` as VerifiedFact['semantics']['periodKind']; |
There was a problem hiding this comment.
Map cumulative fourth quarters to a supported period kind
For an explicit 第4四半期累計 source, this branch returns cumulativeQ4, but FactPeriodKind, checkSemantics, and the candidate prompt do not permit that value. Using fullYear also fails the later equality check, so no candidate or restored fact can represent an otherwise valid fourth-quarter cumulative disclosure. Align this branch and matchesReportingPeriod with a supported full-year or cumulative-Q4 representation.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
累計Q4が現行スキーマで表現できない点は確認しましたが、今回の形式拡張は採用しません。docs/development.mdおよびevaluation/README.mdは以前から累計Q4を未対応/未確認の境界として記録し、Q表紙属性の対応と期間スキーマの拡張を分けています。fullYearへ黙って変換したり、未規定のenumを追加したりせず拒否を維持します。対応形式を広げる製品判断が別途必要な残タスクとして記録しています。
M&A年月・月次KPI根拠・損失と取得の発行者義務・表紙別欄を照合し、指紋v60へ更新。テストの重複を整理して責務別戦略を固定し、最終870件と再生・保存・画面検証を記録。
|
レビュー本文のM&A決議・自己株取得上限の2件も妥当と判断し、5ba3d3fで修正しました(対象レビュー: #25 (review) / https://github.com/ivgtr/tdnet-digest/pull/25#pullrequestreview-5400078845)。各インライン指摘には個別に返信しています。 @codex review bfd83e8から5ba3d3fの差分を再レビューしてください。重複を除く9件のうち7件を採用しました。
桁区切り数量は基準でも生成・保存が正常で、文法の製品変更を採用せず正常テストを保持しました。累計Q4は既存の未対応境界として拒否を維持し、fullYearへの変換やenum拡張は行っていません。 検証: 基準の重点22件は18失敗/正常4成功、最終870件・type-check/lint/build成功。重点3入力不変、S/U/V9生応答の再生一致、56保存結果325受理/68既知拒否で不変。新規CLI実生成0回。最終ビルドの固定修復/表示/保存復元/後続利用と、次の設定失敗自身の診断保存が成功。実TDnet/PDF/API通常要約は1回/21秒、初回11事実・修復0です。 最終source/buildと画面traceのdigest一致を確認し、条件・途中失敗・証拠・未確認範囲をdocs/development.mdの既存T1〜T4表とevaluation/README.mdへ記録しました。未知形式/OCR/多主張分解・累計Q4・今回反例自体の実モデル修復等は未確認のままです。 |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5ba3d3f027
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".
| const match = period.match(/^(20\d{2})年(\d{1,2})月期$/); | ||
| return match && +match[2] >= 1 && +match[2] <= 12 | ||
| ? { period, key: +match[1] * 12 + +match[2] } | ||
| : null; | ||
| }); | ||
| if (dated.some((p) => p === null)) return []; |
There was a problem hiding this comment.
Preserve supported suffixes when ordering M&A periods
When a continued target-company table labels a column with a supported qualified period such as 2027年12月期通期, tableContinuations accepts the axis and the fact schema accepts that period, but this anchored match returns null. Because any null entry causes the entire link to be discarded, maMetricSources creates no obligations for any metrics in that table, allowing the latest target-company results to be omitted. Extract the year and month for comparison without requiring the complete axis text to be only YYYY年M月期, while retaining the original period string for evidence matching.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
妥当と判断し、cfc552aで修正しました。既存sourceFiscalPeriodで単一FYを解決して年月だけを比較キーにし、限定を含む原軸を保持します。同原因の修復expected.periodの通期欠落も、同じ最新根拠から取得するよう揃えました。充足は通常の期間照合済みの同じ最新セルIDで判定し、FY/通期の同義表記を別の文字列比較で拒否しません。既存matrixへ通期限定1例を追加し、欠落拒否・1回修復・保存一致を確認しています。期間スキーマや累計Q4は拡張していません。
| try { | ||
| trace.pdfUrl = fullUrl(request.pdfUrl); | ||
| await saveTrace(); |
There was a problem hiding this comment.
Allow exporting traces for rejected PDF URLs
Fresh evidence beyond the covered settings and fetch failures is a malformed or non-TDnet PDF URL: fullUrl throws here, the catch path saves a failure trace containing the original unnormalized URL, and the UI later calls matchingSummaryTrace, which invokes the same URL normalizer before it can match the run ID. Consequently the diagnostic button only repeats TDnetのPDF URLではありません and cannot download the trace created for this failure. Run-specific failure traces should be matchable without requiring normalization that has already failed.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
妥当と判断し、cfc552aで修正しました。失敗要求は同じrunId・原URL・outcome=failure/resultId=nullの場合だけ、再び失敗するURL正規化を要求せず診断を出力します。成功/保存復元は引き続き正規化URLと実行/結果IDを照合し、URL取得許可も変更していません。既存拒否例を外部URL・構文不正・非PDFへ拡張し、取得/生成0、別run/別URL・成功/復元への転用拒否を確認。最終ビルドの実画面で外部URL拒否→その要求の診断JSONダウンロードが成功し、PDF/API取得0でした。
年月比較キーと原軸を分離し、義務・充足・修復・保存を同じ最新根拠へ揃える。失敗要求IDと原URLで診断を照合し、取得条件と成功/保存復元の照合は維持。指紋v61、最終873件・再生/保存/固定画面の検証を記録。
|
5ba3d3fの最新2指摘を採用し、cfc552aで対応しました。各スレッドへ原因・修正・検証を返信しています。 @codex review 5ba3d3f → cfc552aの差分を再レビューしてください。 今回の傾向は、比較キーと原文表記、失敗要求の同一性と取得許可の混同でした。判定する責務を分け、同じ意味を後続経路で文字列だけ再判定しないようにしています。
指紋v61、候補v2/確定v4/安定ID/最大1修復/20件上限は維持。既存matrixへ通期限定1例、既存URL拒否を3境界へ拡張し、恒久ケースは870→873件です。基準の重点7件は4失敗/正常3成功、最終873件・type-check/lint/build成功。重点3入力不変、9生応答の再生一致、56保存結果325受理/68既知拒否で不変。 最終固定画面で修復/表示/保存復元/診断/後続利用と、拒否URLの診断保存(PDF/API取得0)が成功。最終source/buildと画面traceのdigest一致を確認しました。入力・モデル/設定・LLM通信が不変のため追加実API0回。条件・途中失敗・証拠は既存T1〜T4表とevaluation/README.mdへ記録しました。今回反例自体の実モデル修復、累計Q4、未知形式/OCR/多主張等の既存未確認範囲は保持しています。 |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cfc552ab0b
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".
| const fiscal = sourceFiscalPeriod(period, ''); | ||
| const match = fiscal?.match(/^(20\d{2})年(\d{1,2})月期$/); |
There was a problem hiding this comment.
Exclude unsupported cumulative-Q4 axes from M&A obligations
The fresh evidence since the earlier cumulative-Q4 discussion is this broadened sourceFiscalPeriod check: a continued M&A table headed 2027年3月期第4四半期累計 now passes the chronological filter and creates required revenue/profit slots. provedMappedNumber accepts that source when constructing the obligation, but FactPeriodKind has no cumulativeQ4, so no generated or restored fact can satisfy it and both generation attempts must fail. Restrict this path to representable qualifiers (such as 通期) or reject unsupported period kinds before creating obligations.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
ご指摘は妥当です。f43443fで修正しました。
原因は、maMetricSourcesが利用する数量の原文一致検証と、現行契約で表せる期間区分の証明が分離していたことです。M&Aだけに除外条件を足すのではなく、共通のverifyPeriodAndKindからperiodKindを通すようにし、periodKindが未対応のcumulativeQ4を型castで返す箇所を修正しました。義務生成・修復先・候補確定・保存再照合が同じ表現可能性の境界を使います。累計Q4を通期へ変換したり、スキーマを拡張したりしていません。
同原因の限定未証明Q2も義務/修復先なし・候補/保存で拒否し、証明済みの取得予定だけで通常生成が不要な修復なしに完了することを確認しました。対になる累計Q3・単独Q4は、義務発生・欠落拒否・最大1回修復・保存一致を維持します。既存M&A matrixへ4境界だけ追加し、新しいテストファイルや全経路の総当たりは増やしていません。分析指紋はv62です。
877テスト/28ファイル・type-check・lint・build成功。3入力/9応答/56保存結果の最終再判定は不変、追加実API0。最終ビルドの固定画面で旧v61非表示・修復/表示/保存復元/診断/後続利用を確認。条件とhash対応、未確認範囲はdocs/development.mdとevaluation/README.mdへ記録しました。
|
@codex review f43443fの差分を再レビューしてください。最新のP2(M&Aの累計Q4から充足不能な義務を作る問題)を修正しました。 傾向は、比較キーや原文の文字一致を、そのまま事実の意味証明として扱う接続漏れでした。今回は共通の数量意味検証verifyPeriodAndKindから期間区分判定periodKindを通し、未対応区分を返す型castを除去しています。義務生成・修復先・候補・保存の同じ意味境界を確認してください。局所属性や原期間を既定値で上書きせず、累計Q4のスキーマ追加も行っていません。 先行対応として限定未証明Q2も拒否し、累計Q3/単独Q4の正常生成・欠落拒否・1回修復・保存を維持しました。恒久テストは既存M&A matrixへ4境界追加のみ。分析指紋v62、877テスト/28ファイル・type-check/lint/build成功。重点3入力・9生応答の再生・56 strict保存結果は不変。追加実API0、最終固定画面で旧v61非表示・修復/表示/保存復元/診断/後続利用が成功しました。 正本のdocs/development.mdの既存タスク表とevaluation/README.mdに、基準HEAD cfc552a、最終ソース/ビルドhash、検証・再利用条件、未確認範囲を記録済みです。未知形式/OCR/多主張分解や累計Q4の対応拡張は今回の範囲外です。 |
|
Codex Review: Didn't find any major issues. Breezy! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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". |
要約の数値・期間・会社・報告範囲を原文と照合し、誤った数量の表示と、修復しても要約が完了しない問題を改善しました。
分析条件の更新により、旧条件の要約キャッシュは再利用せず再生成します。テストを責務別に整理し、今後の追加・再実行・実API評価の基準も固定しました。
検証: 877テスト/28ファイル、型チェック・lint・build成功。Codexの最終再レビューはf43443fで新しい指摘なし。6資料29固定事実を維持し、重点3資料の生成入力・既存9応答の再生・56保存結果の再判定も不変です。最終ビルドの固定画面で修復・表示・保存復元・診断・後続利用を確認しました。今回の追加実API呼出しは0回です。
未知形式・OCR・複数主張の分解・累計Q4の対応拡張、今回反例の実モデルによる修復は未確認です。詳細な条件と既存の停止履歴は評価記録に記録しています。