@mcptoolshop/research-os 0.4.0 → 0.6.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +253 -0
- package/README.es.md +33 -2
- package/README.fr.md +32 -1
- package/README.hi.md +52 -1
- package/README.it.md +33 -2
- package/README.ja.md +32 -1
- package/README.md +53 -1
- package/README.pt-BR.md +32 -1
- package/README.zh.md +33 -2
- package/dist/calibration/aggregate-receipt-schema.d.ts +547 -0
- package/dist/calibration/aggregate-receipt-schema.js +160 -0
- package/dist/calibration/aggregate-receipt-schema.js.map +1 -0
- package/dist/calibration/aggregate.d.ts +37 -0
- package/dist/calibration/aggregate.js +493 -0
- package/dist/calibration/aggregate.js.map +1 -0
- package/dist/calibration/receipt-schema.d.ts +356 -0
- package/dist/calibration/receipt-schema.js +83 -0
- package/dist/calibration/receipt-schema.js.map +1 -0
- package/dist/calibration/receipt.d.ts +32 -0
- package/dist/calibration/receipt.js +170 -0
- package/dist/calibration/receipt.js.map +1 -0
- package/dist/cli.js +1041 -851
- package/dist/cli.js.map +1 -1
- package/dist/index.d.ts +154 -49
- package/dist/index.js +881 -818
- package/dist/index.js.map +1 -1
- package/dist/reviewer-options-schema-PZacF_MO.d.ts +27 -0
- package/package.json +1 -1
package/README.ja.md
CHANGED
|
@@ -7,7 +7,7 @@
|
|
|
7
7
|
</p>
|
|
8
8
|
|
|
9
9
|
<p align="center">
|
|
10
|
-
<a href="https://github.com/mcp-tool-shop-org/research-os/releases/tag/v0.
|
|
10
|
+
<a href="https://github.com/mcp-tool-shop-org/research-os/releases/tag/v0.6.0"><img src="https://img.shields.io/badge/version-0.6.0-blue" alt="version 0.6.0"></a>
|
|
11
11
|
<a href="https://github.com/mcp-tool-shop-org/research-os/actions/workflows/ci.yml"><img src="https://github.com/mcp-tool-shop-org/research-os/actions/workflows/ci.yml/badge.svg" alt="CI"></a>
|
|
12
12
|
<a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-green" alt="MIT License"></a>
|
|
13
13
|
<img src="https://img.shields.io/badge/node-%E2%89%A520-brightgreen" alt="Node ≥20">
|
|
@@ -149,8 +149,39 @@ discover
|
|
|
149
149
|
|
|
150
150
|
`research-os` は、ローカル環境で動作するCLIです。このツールは、指定された研究パッケージのディレクトリ内のファイルを読み書きし、`gather` コマンドを使用する場合、提供されたソースコードのURLから情報を取得するために、HTTPリクエストを送信します。このツールは、サーバーを起動したり、外部からの接続を受け付けたり、認証情報を保存したり、テレメトリデータを送信したりすることはありません。また、機密情報はパッケージのファイルに書き込まれません。脆弱性に関する報告については、[SECURITY.md](SECURITY.md) を参照してください。
|
|
151
151
|
|
|
152
|
+
## レビュー担当者のキャリブレーション
|
|
153
|
+
|
|
154
|
+
v0.5.0では、レビュー担当者のキャリブレーションがより堅牢になりました。レビュー担当者のプロファイルは、単に一度実行されたというだけで信頼されるわけではありません。構造化された意図的なエラーの記録と、複数回の実行による集計によって、信頼度を獲得します。v0.6.0では、本番環境のレビュープロセスとキャリブレーション環境に、再現性のあるレビュー担当者オプションが追加されました。
|
|
155
|
+
|
|
156
|
+
**現在、どのプロファイルも`trusted_baseline`として認められていません。** リポジトリ内の標準的な記録には、`hermes-two-pass=failed`、`mistral-nemo-two-pass=conditional_pass`、`hermes-single-pass=comparison_only`、`hermes-two-pass-deterministic=failed`と記載されています。これは意図的なものです。信頼は、仮定ではなく、繰り返しの検証による証拠によって獲得されます。`hermes-two-pass-deterministic`の記録には、構造的なモデルの能力ギャップ(6種類の判断のうち2種類しか生成できない。3種類が必要)があり、これはばらつきの問題ではありません。
|
|
157
|
+
|
|
158
|
+
キャリブレーションの記録は、`calibration/reviewer-profiles/<profile>/seeded-v1.{json,md}`に保存されています。各記録は、7つの項目に対するPASS/FAILの結果、4つのステータスラベル(`trusted_baseline`、`conditional_pass`、`failed`、`comparison_only`)、および、テストできない内容を正直に開示しています(`needs_contradiction_mapping`は`seeded-v1`からはアクセスできません)。詳細は[CHANGELOG.md](CHANGELOG.md)を参照してください。
|
|
159
|
+
|
|
160
|
+
```bash
|
|
161
|
+
# Single-run calibration (quick local check)
|
|
162
|
+
node scripts/reviewer-calibration.mjs --model hermes3:8b --two-pass --profile hermes-two-pass
|
|
163
|
+
|
|
164
|
+
# Multi-run aggregate calibration (canonical evidence — 3 runs, median-based PASS/FAIL)
|
|
165
|
+
node scripts/reviewer-calibration.mjs --model hermes3:8b --two-pass --profile hermes-two-pass --runs 3
|
|
166
|
+
|
|
167
|
+
# Deterministic multi-run calibration (temperature + seed explicit in receipt)
|
|
168
|
+
node scripts/reviewer-calibration.mjs --model hermes3:8b --two-pass \
|
|
169
|
+
--temperature 0 --seed 7 --runs 3 --profile hermes-two-pass-deterministic
|
|
170
|
+
|
|
171
|
+
# Promote a section's review — auto-populates calibration_summary from pack-relative receipt
|
|
172
|
+
research-os review-promote 01-section --pack <pack> --profile hermes-two-pass
|
|
173
|
+
```
|
|
174
|
+
|
|
175
|
+
`--runs <n>`オプションを使用すると、各実行の記録が`<profile>/runs/run-NNN.json`に書き込まれ、集計された記録(中央値に基づいた項目と、再発するエラーの検出を含む)が`<profile>/seeded-v1.{json,md}`に書き込まれます。集計された記録には、`receipt_kind: 'aggregate'`という情報が含まれており、これにより単一実行の記録と区別できます。単一実行モード(`--runs 1`または省略)では、既存の直接書き込みの動作が維持されます。
|
|
176
|
+
|
|
177
|
+
**再現性のあるレビュー担当者プロファイル** — `research.yaml`の`review_profiles.<name>.reviewer_options`を使用して、`temperature`、`seed`、およびその他のOllamaのサンプリングパラメータを、本番環境のレビュープロセスにおけるすべての`OllamaInternReviewer`の構築に適用します。`hermes-two-pass-deterministic`プロファイルは、組み込みのサンプルとして提供されています。詳細は[`docs/experiment-6-proof.md`](docs/experiment-6-proof.md)と、[レビュー担当者キャリブレーションハンドブック](https://mcp-tool-shop-org.github.io/research-os/handbook/reviewer-calibration/)を参照してください。
|
|
178
|
+
|
|
152
179
|
## ステータス
|
|
153
180
|
|
|
181
|
+
**v0.6.0** — npmに`@mcptoolshop/research-os@0.6.0`として公開されました。2026年5月10日。v0.6.0では、実験6が、レビュー担当者の信頼性に関する証拠とともに完了しました。これにより、research-osは、再現可能で、帰属可能な、標準的なモデルのベースラインを生成できるようになりました。変更点:本番環境のレビュープロセスにおける再現性のあるレビュー担当者オプション(`review_profiles.<name>.reviewer_options`を`research.yaml`に追加)、既存のv0.3.3以前のフローズンアーティファクトに対するゲートスキーマの互換性(F-53)、レビュー出力にサンプリング条件が直接`review.json`と`review.md`に表示されるように変更(F-54)、標準的な再現性のある集計記録がコミットされました(`hermes-two-pass-deterministic`、`temperature:0, seed:7`)。**どのプロファイルも`trusted_baseline`として認められていません。** `hermes-two-pass-deterministic=failed`(判断の語彙における構造的なモデルの能力ギャップ。ばらつきの問題ではない)。**Hermesは`trusted_baseline`として昇格しません。** 重要なのは、メカニズムであり、単に合格する記録ではありません。ゲート、フリーズ、または合成法の変更はありません。すべてのフローズンパックが、バイト単位で同一であることを確認しました。713/713のvitestが合格しました。詳細は[CHANGELOG.md](CHANGELOG.md)と[`docs/experiment-6-proof.md`](docs/experiment-6-proof.md)を参照してください。
|
|
182
|
+
|
|
183
|
+
**v0.5.0** — npmに `@mcptoolshop/research-os@0.5.0` として公開。2026年5月10日。v0.5.0では、レビュー担当者の評価の信頼性を高めるための機能が導入されました。レビュー担当者のプロファイルは、単に一度実行されたというだけで信頼されるわけではありません。構造化されたテストケースと複数回の実行結果を組み合わせることで、信頼度を評価します。同梱内容:構造化された評価結果スキーマ (`seeded-v1.{json,md}`、Zodによる検証、4つのステータスラベル)、複数回の実行をサポートする機能 (`--runs <n>`、各実行の分離、中央値に基づいた合否判定、繰り返し発生するエラーに対する評価の引き下げ)、アーキテクチャを考慮した意思決定のための語彙セット、`review-promote` 内でのパッケージ相対的な評価結果の参照機能。**信頼できる基準値は認められません:** `hermes-two-pass=failed` (集計、3回の実行)、`mistral-nemo-two-pass=conditional_pass`、`hermes-single-pass=comparison_only`。research-osは、繰り返し発生するテストの失敗が信頼を裏付けることができない場合、レビュー担当者のプロファイルを信頼しないようにすることができます。**ゲート、フリーズ、または合成規則に関する変更はありません。すべての4つのパッケージが、バイト単位で完全に同一であることを検証済みです。** 671/671のvitestテストが合格。詳細は [CHANGELOG.md](CHANGELOG.md) を参照してください。
|
|
184
|
+
|
|
154
185
|
**v0.4.0** — npmに`@mcptoolshop/research-os@0.4.0`として公開。2026年5月10日。v0.4.0では、ソースの同一性を維持できるようになりました。決定論的なソースタイプルールにより、再現可能な多数が処理され、オーバーライドされたレジャーにより、再収集時のオペレーターによる修正が保持され、`source-card audit`コマンドが、従来のスクリプトのずれチェックを置き換え、より使いやすいCLIインターフェースを提供します。同梱内容:集中型のソースタイプ分類器(コンポーネントB — `classifySourceType`、11種類のベンダー、`source-type-rules.json`)、ソースカードのオーバーライドレジャー(コンポーネントA — `source-card-overrides.jsonl`、`validate`および`list`サブコマンド)、およびソースカード監査CLI(コンポーネントD — `research-os source-card audit --pack <dir>`、7種類の検出結果、JSONおよびMarkdown形式のレポート、`--apply --from`による適用パス)。F-46:見た目の修正。パッケージのマニフェストには、`research.yaml`に固定されたバージョンではなく、実行中のバイナリのバージョンが記録されるようになりました。**ゲート、フリーズ、または合成に関する変更はありません。既存のすべてのパッケージは、バイト単位で同一であることを検証済みです。** 620/620のvitestテストが合格しました。詳細は[CHANGELOG.md](CHANGELOG.md)および[ソースカード監査に関するハンドブック](https://mcp-tool-shop-org.github.io/research-os/handbook/source-card-audit/)を参照してください。
|
|
155
186
|
|
|
156
187
|
**v0.3.3** — npmに`@mcptoolshop/research-os@0.3.3`として公開。2026年5月10日。Pack-3(Godotのエクスポート/ランタイムの安定性、実験3のパッケージ#3)によって得られた、ゲートのセマンティクスに関する明確化が含まれています。ゲートの出力には、セクションごとのパブリッシャー数と主要なカウントに加えて、パッケージ全体のカウントも表示されます(F-43)。`no_source_cluster_monopoly`は、警告から情報診断に変更されました(F-41)。**合格/不合格の動作は変更されていません。既存のパッケージは、バイト単位で同一であることを検証済みです。** 570/570のvitestテストが合格しました。詳細は[CHANGELOG.md](CHANGELOG.md)および[`docs/section-scoped-waivers.md`](docs/section-scoped-waivers.md)を参照してください。
|
package/README.md
CHANGED
|
@@ -7,7 +7,7 @@
|
|
|
7
7
|
</p>
|
|
8
8
|
|
|
9
9
|
<p align="center">
|
|
10
|
-
<a href="https://github.com/mcp-tool-shop-org/research-os/releases/tag/v0.
|
|
10
|
+
<a href="https://github.com/mcp-tool-shop-org/research-os/releases/tag/v0.6.0"><img src="https://img.shields.io/badge/version-0.6.0-blue" alt="version 0.6.0"></a>
|
|
11
11
|
<a href="https://github.com/mcp-tool-shop-org/research-os/actions/workflows/ci.yml"><img src="https://github.com/mcp-tool-shop-org/research-os/actions/workflows/ci.yml/badge.svg" alt="CI"></a>
|
|
12
12
|
<a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-green" alt="MIT License"></a>
|
|
13
13
|
<img src="https://img.shields.io/badge/node-%E2%89%A520-brightgreen" alt="Node ≥20">
|
|
@@ -149,8 +149,60 @@ This is the structural alternative to *search → summarize → pretty report*.
|
|
|
149
149
|
|
|
150
150
|
`research-os` is a local-first CLI. It reads and writes files within the research-pack directory you point it at, and (when using `gather`) issues outbound HTTP requests to fetch source URLs you provide. It does not: run a server, accept inbound connections, store credentials, or send telemetry. No secrets are written to pack artifacts. See [SECURITY.md](SECURITY.md) for the vulnerability reporting policy.
|
|
151
151
|
|
|
152
|
+
## Reviewer calibration
|
|
153
|
+
|
|
154
|
+
v0.5.0 makes reviewer calibration durable. A reviewer profile is not trusted because
|
|
155
|
+
it ran once; it earns a status through structured seeded-failure receipts and
|
|
156
|
+
multi-run aggregation. v0.6.0 adds deterministic reviewer options to the production
|
|
157
|
+
review path and calibration harness.
|
|
158
|
+
|
|
159
|
+
**No profile is currently admitted as `trusted_baseline`.** The canonical receipts
|
|
160
|
+
in the repo show `hermes-two-pass=failed`, `mistral-nemo-two-pass=conditional_pass`,
|
|
161
|
+
`hermes-single-pass=comparison_only`, `hermes-two-pass-deterministic=failed`. This is
|
|
162
|
+
intentional: trust is earned through repeated seeded-failure evidence, not assumed.
|
|
163
|
+
The `hermes-two-pass-deterministic` receipt has a structural model-capability gap
|
|
164
|
+
(2/6 decision types produced; requires 3/6) that is not a variance problem.
|
|
165
|
+
|
|
166
|
+
Calibration receipts live at `calibration/reviewer-profiles/<profile>/seeded-v1.{json,md}`.
|
|
167
|
+
Each receipt records PASS/FAIL against seven bars, four status labels
|
|
168
|
+
(`trusted_baseline`, `conditional_pass`, `failed`, `comparison_only`), and
|
|
169
|
+
honestly discloses what the fixture cannot test (`needs_contradiction_mapping`
|
|
170
|
+
is unreachable from `seeded-v1`). See [CHANGELOG.md](CHANGELOG.md).
|
|
171
|
+
|
|
172
|
+
```bash
|
|
173
|
+
# Single-run calibration (quick local check)
|
|
174
|
+
node scripts/reviewer-calibration.mjs --model hermes3:8b --two-pass --profile hermes-two-pass
|
|
175
|
+
|
|
176
|
+
# Multi-run aggregate calibration (canonical evidence — 3 runs, median-based PASS/FAIL)
|
|
177
|
+
node scripts/reviewer-calibration.mjs --model hermes3:8b --two-pass --profile hermes-two-pass --runs 3
|
|
178
|
+
|
|
179
|
+
# Deterministic multi-run calibration (temperature + seed explicit in receipt)
|
|
180
|
+
node scripts/reviewer-calibration.mjs --model hermes3:8b --two-pass \
|
|
181
|
+
--temperature 0 --seed 7 --runs 3 --profile hermes-two-pass-deterministic
|
|
182
|
+
|
|
183
|
+
# Promote a section's review — auto-populates calibration_summary from pack-relative receipt
|
|
184
|
+
research-os review-promote 01-section --pack <pack> --profile hermes-two-pass
|
|
185
|
+
```
|
|
186
|
+
|
|
187
|
+
When `--runs <n>` is used, per-run receipts are written to `<profile>/runs/run-NNN.json`
|
|
188
|
+
and an aggregate receipt (with median-based bars and recurring-failure detection) is written
|
|
189
|
+
to `<profile>/seeded-v1.{json,md}`. The aggregate receipt carries `receipt_kind: 'aggregate'`
|
|
190
|
+
to discriminate from single-run receipts. Single-run mode (`--runs 1` or omitted) preserves
|
|
191
|
+
the existing direct-write behavior.
|
|
192
|
+
|
|
193
|
+
**Deterministic reviewer profiles** — use `review_profiles.<name>.reviewer_options` in
|
|
194
|
+
`research.yaml` to carry `temperature`, `seed`, and other Ollama sampling parameters
|
|
195
|
+
into every `OllamaInternReviewer` construction in the production review path. The
|
|
196
|
+
`hermes-two-pass-deterministic` profile ships as a built-in example. See
|
|
197
|
+
[`docs/experiment-6-proof.md`](docs/experiment-6-proof.md) and the
|
|
198
|
+
[reviewer calibration handbook page](https://mcp-tool-shop-org.github.io/research-os/handbook/reviewer-calibration/).
|
|
199
|
+
|
|
152
200
|
## Status
|
|
153
201
|
|
|
202
|
+
**v0.6.0** — published to npm as `@mcptoolshop/research-os@0.6.0`, 2026-05-10. v0.6.0 closes Experiment 6 with reviewer-trust evidence: research-os can now produce a reproducible, attributable canonical-model baseline. Ships: deterministic reviewer options on the production review path (`review_profiles.<name>.reviewer_options` in `research.yaml`); gate schema backward compatibility for pre-v0.3.3 frozen artifacts (F-53); review output discloses sampling conditions directly on `review.json` and `review.md` (F-54); canonical deterministic aggregate receipt committed (`hermes-two-pass-deterministic`, `temperature:0, seed:7`). **No trusted baseline admitted.** `hermes-two-pass-deterministic=failed` (structural model-capability gap in decision vocabulary, not variance). **Hermes is not promoted to `trusted_baseline`.** The win is the mechanism, not a passing receipt. No gate, freeze, or synthesis-law changes. All four frozen packs verify-pack byte-identically. 713/713 vitest passing. See [CHANGELOG.md](CHANGELOG.md) and [`docs/experiment-6-proof.md`](docs/experiment-6-proof.md).
|
|
203
|
+
|
|
204
|
+
**v0.5.0** — published to npm as `@mcptoolshop/research-os@0.5.0`, 2026-05-10. v0.5.0 makes reviewer calibration durable. A reviewer profile is not trusted because it ran once; it earns a status through structured seeded-failure receipts and multi-run aggregation. Ships: structured calibration receipt schema (`seeded-v1.{json,md}`, Zod-validated, four status labels); multi-run harness (`--runs <n>`, per-run isolation, median-based PASS/FAIL bars, recurring-failure demotion); architecture-aware decision-vocab bar; pack-relative receipt lookup in `review-promote`. **No trusted baseline admitted:** `hermes-two-pass=failed` (aggregate, 3 runs), `mistral-nemo-two-pass=conditional_pass`, `hermes-single-pass=comparison_only`. research-os can now refuse to trust a reviewer profile when repeated seeded failures do not support trust. **No gate, freeze, or synthesis-law changes. All four frozen packs verify-pack byte-identically.** 671/671 vitest passing. See [CHANGELOG.md](CHANGELOG.md).
|
|
205
|
+
|
|
154
206
|
**v0.4.0** — published to npm as `@mcptoolshop/research-os@0.4.0`, 2026-05-10. v0.4.0 makes source identity durable. Deterministic source-type rules handle the repeatable majority, override ledgers preserve operator corrections across re-gather, and `source-card audit` replaces scratch-script drift checks with a first-class CLI surface. Ships: centralized source-type classifier (Component B — `classifySourceType`, 11 canonical vendors, `source-type-rules.json`); source-card override ledger (Component A — `source-card-overrides.jsonl`, `validate` + `list` subcommands); and source-card audit CLI (Component D — `research-os source-card audit --pack <dir>`, 7 finding kinds, JSON + Markdown artifacts, `--apply --from` apply path). F-46 cosmetic fix: pack manifests now stamp the live binary version rather than the version frozen into `research.yaml` at pack-init. **No gate, freeze, or synthesis-law changes. All four existing frozen packs verify-pack byte-identically.** 620/620 vitest passing. See [CHANGELOG.md](CHANGELOG.md) and the [source-card audit handbook page](https://mcp-tool-shop-org.github.io/research-os/handbook/source-card-audit/).
|
|
155
207
|
|
|
156
208
|
**v0.3.3** — published to npm as `@mcptoolshop/research-os@0.3.3`, 2026-05-10. Ships gate-semantics clarity earned by Pack-3 (Godot export/runtime durability, Experiment 3 pack #3 of 3). Gate output now carries section-scoped publisher + primary counts alongside pack-wide counts (F-43); `no_source_cluster_monopoly` reworded from WARN to informational diagnostic (F-41). **Pass/fail behavior unchanged; existing frozen packs verify-pack byte-identically.** 570/570 vitest passing. See [CHANGELOG.md](CHANGELOG.md) and [`docs/section-scoped-waivers.md`](docs/section-scoped-waivers.md).
|
package/README.pt-BR.md
CHANGED
|
@@ -7,7 +7,7 @@
|
|
|
7
7
|
</p>
|
|
8
8
|
|
|
9
9
|
<p align="center">
|
|
10
|
-
<a href="https://github.com/mcp-tool-shop-org/research-os/releases/tag/v0.
|
|
10
|
+
<a href="https://github.com/mcp-tool-shop-org/research-os/releases/tag/v0.6.0"><img src="https://img.shields.io/badge/version-0.6.0-blue" alt="version 0.6.0"></a>
|
|
11
11
|
<a href="https://github.com/mcp-tool-shop-org/research-os/actions/workflows/ci.yml"><img src="https://github.com/mcp-tool-shop-org/research-os/actions/workflows/ci.yml/badge.svg" alt="CI"></a>
|
|
12
12
|
<a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-green" alt="MIT License"></a>
|
|
13
13
|
<img src="https://img.shields.io/badge/node-%E2%89%A520-brightgreen" alt="Node ≥20">
|
|
@@ -149,8 +149,39 @@ Esta é a alternativa estrutural para *pesquisar → resumir → gerar relatóri
|
|
|
149
149
|
|
|
150
150
|
`research-os` é uma ferramenta de linha de comando que opera localmente. Ela lê e grava arquivos dentro do diretório do pacote de pesquisa que você especificar e, quando usa o comando `gather`, faz solicitações HTTP para buscar URLs de origem que você fornecer. Ela não: executa um servidor, aceita conexões de entrada, armazena credenciais ou envia dados de telemetria. Nenhum segredo é gravado nos arquivos do pacote. Consulte [SECURITY.md](SECURITY.md) para a política de relatório de vulnerabilidades.
|
|
151
151
|
|
|
152
|
+
## Calibração de revisores
|
|
153
|
+
|
|
154
|
+
v0.5.0 torna a calibração dos revisores mais robusta. Um perfil de revisor não é considerado confiável apenas porque foi executado uma vez; ele adquire um status através de relatórios estruturados de falhas simuladas e agregação de múltiplas execuções. v0.6.0 adiciona opções de revisor determinísticas ao fluxo de revisão de produção e ao sistema de calibração.
|
|
155
|
+
|
|
156
|
+
**Nenhum perfil é atualmente aceito como `baseline confiável`.** Os relatórios canônicos no repositório mostram `hermes-two-pass=falha`, `mistral-nemo-two-pass=aprovação condicional`, `hermes-single-pass=comparação apenas`, `hermes-two-pass-deterministic=falha`. Isso é intencional: a confiança é conquistada através de evidências repetidas de falhas simuladas, e não é presumida. O relatório `hermes-two-pass-deterministic` apresenta uma lacuna estrutural na capacidade do modelo (2/6 tipos de decisão produzidos; requer 3/6) que não é um problema de variância.
|
|
157
|
+
|
|
158
|
+
Os relatórios de calibração estão localizados em `calibration/reviewer-profiles/<perfil>/seeded-v1.{json,md}`. Cada relatório registra PASS/FAIL em relação a sete critérios, quatro rótulos de status (`trusted_baseline`, `conditional_pass`, `failed`, `comparison_only`), e revela honestamente o que o teste não consegue verificar (`needs_contradiction_mapping` é inacessível a partir de `seeded-v1`). Consulte [CHANGELOG.md](CHANGELOG.md).
|
|
159
|
+
|
|
160
|
+
```bash
|
|
161
|
+
# Single-run calibration (quick local check)
|
|
162
|
+
node scripts/reviewer-calibration.mjs --model hermes3:8b --two-pass --profile hermes-two-pass
|
|
163
|
+
|
|
164
|
+
# Multi-run aggregate calibration (canonical evidence — 3 runs, median-based PASS/FAIL)
|
|
165
|
+
node scripts/reviewer-calibration.mjs --model hermes3:8b --two-pass --profile hermes-two-pass --runs 3
|
|
166
|
+
|
|
167
|
+
# Deterministic multi-run calibration (temperature + seed explicit in receipt)
|
|
168
|
+
node scripts/reviewer-calibration.mjs --model hermes3:8b --two-pass \
|
|
169
|
+
--temperature 0 --seed 7 --runs 3 --profile hermes-two-pass-deterministic
|
|
170
|
+
|
|
171
|
+
# Promote a section's review — auto-populates calibration_summary from pack-relative receipt
|
|
172
|
+
research-os review-promote 01-section --pack <pack> --profile hermes-two-pass
|
|
173
|
+
```
|
|
174
|
+
|
|
175
|
+
Quando `--runs <n>` é usado, os relatórios de cada execução são gravados em `<perfil>/runs/run-NNN.json` e um relatório agregado (com critérios baseados na mediana e detecção de falhas recorrentes) é gravado em `<perfil>/seeded-v1.{json,md}`. O relatório agregado contém `receipt_kind: 'aggregate'` para diferenciá-lo dos relatórios de execução única. O modo de execução única (`--runs 1` ou omitido) preserva o comportamento de gravação direta existente.
|
|
176
|
+
|
|
177
|
+
**Perfis de revisor determinísticos** — utilize `review_profiles.<nome>.reviewer_options` em `research.yaml` para incluir os parâmetros de amostragem do Ollama, como `temperature` e `seed`, em cada instância de `OllamaInternReviewer` no fluxo de revisão de produção. O perfil `hermes-two-pass-deterministic` é fornecido como um exemplo integrado. Consulte [`docs/experiment-6-proof.md`](docs/experiment-6-proof.md) e a [página do manual de calibração do revisor](https://mcp-tool-shop-org.github.io/research-os/handbook/reviewer-calibration/).
|
|
178
|
+
|
|
152
179
|
## Status
|
|
153
180
|
|
|
181
|
+
**v0.6.0** — publicado no npm como `@mcptoolshop/research-os@0.6.0`, 10 de maio de 2026. v0.6.0 finaliza o Experimento 6 com evidências de confiança do revisor: o research-os agora pode produzir uma baseline canônica de modelo reproduzível e rastreável. Inclui: opções de revisor determinísticas no fluxo de revisão de produção (`review_profiles.<nome>.reviewer_options` em `research.yaml`); compatibilidade retroativa do esquema de "gate" para artefatos congelados anteriores à versão 0.3.3 (F-53); a saída da revisão revela as condições de amostragem diretamente nos arquivos `review.json` e `review.md` (F-54); relatório agregado determinístico canônico incluído (`hermes-two-pass-deterministic`, `temperature:0, seed:7`). **Nenhum baseline confiável aceito.** `hermes-two-pass-deterministic=falha` (lacuna estrutural na capacidade do modelo no vocabulário de decisão, não variância). **Hermes não é promovido a `baseline confiável`.** O ganho é o mecanismo, não um relatório de aprovação. Não houve alterações nos "gates", no processo de congelamento ou nas leis de síntese. Todos os quatro pacotes congelados são idênticos em termos de bytes. 713/713 testes vitest passaram. Consulte [CHANGELOG.md](CHANGELOG.md) e [`docs/experiment-6-proof.md`](docs/experiment-6-proof.md).
|
|
182
|
+
|
|
183
|
+
**v0.5.0** — publicado no npm como `@mcptoolshop/research-os@0.5.0`, 10 de maio de 2026. A versão v0.5.0 torna a calibração de revisores mais robusta. Um perfil de revisor não é considerado confiável apenas porque foi executado uma vez; ele adquire um status através de relatórios estruturados de falhas simuladas e agregação de múltiplas execuções. Inclui: esquema de relatório de calibração estruturado (`seeded-v1.{json,md}`, validado com Zod, quatro rótulos de status); sistema de execução de múltiplas execuções (`--runs <n>`, isolamento por execução, critérios PASS/FAIL baseados na mediana, detecção de falhas recorrentes); critério de avaliação baseado na arquitetura; pesquisa de relatórios relativa ao pacote em `review-promote`. **Nenhum baseline confiável admitido:** `hermes-two-pass=failed` (agregado, 3 execuções), `mistral-nemo-two-pass=conditional_pass`, `hermes-single-pass=comparison_only`. O research-os agora pode recusar a confiança em um perfil de revisor quando falhas simuladas repetidas não suportam a confiança. **Nenhuma alteração nos gates, congelamentos ou leis de síntese. Todos os quatro pacotes congelados verificam a identidade dos bytes.** 671/671 testes vitest aprovados. Consulte [CHANGELOG.md](CHANGELOG.md).
|
|
184
|
+
|
|
154
185
|
**v0.4.0** — Publicada no npm como `@mcptoolshop/research-os@0.4.0`, 10 de maio de 2026. A versão 0.4.0 garante a durabilidade da identidade da fonte. Regras determinísticas para o tipo de fonte lidam com a maioria repetível, os registros de substituição preservam as correções do operador durante a re-coleta, e o comando `source-card audit` substitui as verificações de derivação de scripts por uma interface de linha de comando (CLI) completa. Inclui: um classificador centralizado de tipo de fonte (Componente B — `classifySourceType`, 11 fornecedores padrão, `source-type-rules.json`); um registro de substituição de cartão de fonte (Componente A — `source-card-overrides.jsonl`, subcomandos `validate` e `list`); e uma CLI para auditoria de cartão de fonte (Componente D — `research-os source-card audit --pack <dir>`, 7 tipos de detecção, artefatos JSON + Markdown, opções `--apply --from` para aplicar o caminho). Correção estética F-46: os arquivos de manifesto agora indicam a versão binária em execução, em vez da versão fixada no arquivo `research.yaml` durante a inicialização da criação do pacote. **Não há alterações nas regras de validação, congelamento ou síntese. Todos os quatro pacotes existentes passam na verificação de integridade byte a byte.** 620/620 testes vitest aprovados. Consulte o arquivo [CHANGELOG.md](CHANGELOG.md) e a página do manual de auditoria de cartão de fonte: [https://mcp-tool-shop-org.github.io/research-os/handbook/source-card-audit/](https://mcp-tool-shop-org.github.io/research-os/handbook/source-card-audit/).
|
|
155
186
|
|
|
156
187
|
**v0.3.3** — Publicado no npm como `@mcptoolshop/research-os@0.3.3`, 10 de maio de 2026. Inclui melhorias na clareza da semântica das "gates" obtidas com o Pack-3 (durabilidade da exportação/runtime do Godot, Experimento 3, pacote nº 3 de 3). A saída da "gate" agora inclui contadores específicos da seção, além dos contadores globais (F-43); a mensagem `no_source_cluster_monopoly` foi alterada de um aviso para um diagnóstico informativo (F-41). **O comportamento de aprovação/reprovação não foi alterado; os pacotes congelados existentes são verificados byte a byte.** 570/570 testes do vitest passaram. Consulte o arquivo [CHANGELOG.md](CHANGELOG.md) e o arquivo [`docs/section-scoped-waivers.md`](docs/section-scoped-waivers.md).
|
package/README.zh.md
CHANGED
|
@@ -7,7 +7,7 @@
|
|
|
7
7
|
</p>
|
|
8
8
|
|
|
9
9
|
<p align="center">
|
|
10
|
-
<a href="https://github.com/mcp-tool-shop-org/research-os/releases/tag/v0.
|
|
10
|
+
<a href="https://github.com/mcp-tool-shop-org/research-os/releases/tag/v0.6.0"><img src="https://img.shields.io/badge/version-0.6.0-blue" alt="version 0.6.0"></a>
|
|
11
11
|
<a href="https://github.com/mcp-tool-shop-org/research-os/actions/workflows/ci.yml"><img src="https://github.com/mcp-tool-shop-org/research-os/actions/workflows/ci.yml/badge.svg" alt="CI"></a>
|
|
12
12
|
<a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-green" alt="MIT License"></a>
|
|
13
13
|
<img src="https://img.shields.io/badge/node-%E2%89%A520-brightgreen" alt="Node ≥20">
|
|
@@ -149,9 +149,40 @@ discover
|
|
|
149
149
|
|
|
150
150
|
`research-os` 是一个本地优先的命令行工具。它在您指定的“研究包”目录中读取和写入文件,并在使用 `gather` 命令时,会向外部发送 HTTP 请求以获取您提供的来源 URL。它不会:运行服务器、接受传入连接、存储凭据或发送遥测数据。任何敏感信息都不会写入到包文件中。请参阅 [SECURITY.md](SECURITY.md),了解漏洞报告政策。
|
|
151
151
|
|
|
152
|
+
## 评审员校准
|
|
153
|
+
|
|
154
|
+
v0.5.0版本使评审器校准更加可靠。评审器配置文件不会因为只运行一次而被信任,而是通过结构化的、带有模拟错误的反馈和多次运行的聚合来获得信任状态。v0.6.0版本为生产环境的评审流程和校准工具添加了确定性的评审器选项。
|
|
155
|
+
|
|
156
|
+
**目前没有任何配置文件被认为是“可信任的基准”。** 仓库中的标准反馈显示:`hermes-two-pass=failed`(失败),`mistral-nemo-two-pass=conditional_pass`(条件通过),`hermes-single-pass=comparison_only`(仅进行比较),`hermes-two-pass-deterministic=failed`(失败)。这是有意为之:信任是通过重复的、带有模拟错误的数据获得的,而不是默认信任。`hermes-two-pass-deterministic`反馈存在结构上的模型能力差距(产生了2/6种决策类型,需要3/6种),这并非是方差问题。
|
|
157
|
+
|
|
158
|
+
校准结果文件位于`calibration/reviewer-profiles/<profile>/seeded-v1.{json,md}`。每个结果文件记录了针对七个方面的PASS/FAIL(通过/失败)结果,四个状态标签(`trusted_baseline`、`conditional_pass`、`failed`、`comparison_only`),并诚实地披露了测试框架无法测试的内容(`needs_contradiction_mapping`无法从`seeded-v1`访问)。请参阅[CHANGELOG.md](CHANGELOG.md)。
|
|
159
|
+
|
|
160
|
+
```bash
|
|
161
|
+
# Single-run calibration (quick local check)
|
|
162
|
+
node scripts/reviewer-calibration.mjs --model hermes3:8b --two-pass --profile hermes-two-pass
|
|
163
|
+
|
|
164
|
+
# Multi-run aggregate calibration (canonical evidence — 3 runs, median-based PASS/FAIL)
|
|
165
|
+
node scripts/reviewer-calibration.mjs --model hermes3:8b --two-pass --profile hermes-two-pass --runs 3
|
|
166
|
+
|
|
167
|
+
# Deterministic multi-run calibration (temperature + seed explicit in receipt)
|
|
168
|
+
node scripts/reviewer-calibration.mjs --model hermes3:8b --two-pass \
|
|
169
|
+
--temperature 0 --seed 7 --runs 3 --profile hermes-two-pass-deterministic
|
|
170
|
+
|
|
171
|
+
# Promote a section's review — auto-populates calibration_summary from pack-relative receipt
|
|
172
|
+
research-os review-promote 01-section --pack <pack> --profile hermes-two-pass
|
|
173
|
+
```
|
|
174
|
+
|
|
175
|
+
当使用`--runs <n>`参数时,每个运行的结果文件会被写入到`<profile>/runs/run-NNN.json`,并且会生成一个聚合结果文件(包含基于中位数的PASS/FAIL结果,以及重复失败检测),写入到`<profile>/seeded-v1.{json,md}`。聚合结果文件包含`receipt_kind: 'aggregate'`,用于区分单次运行的结果文件。单次运行模式(`--runs 1`或省略)会保留现有的直接写入行为。
|
|
176
|
+
|
|
177
|
+
**确定性的评审器配置文件**——在`research.yaml`文件中,使用`review_profiles.<name>.reviewer_options`来将`temperature`(温度)、`seed`(种子)和其他Ollama采样参数传递到生产环境评审流程中的每个`OllamaInternReviewer`实例。`hermes-two-pass-deterministic`配置文件作为内置示例提供。请参阅[`docs/experiment-6-proof.md`](docs/experiment-6-proof.md)以及[评审器校准手册页面](https://mcp-tool-shop-org.github.io/research-os/handbook/reviewer-calibration/)。
|
|
178
|
+
|
|
152
179
|
## 状态
|
|
153
180
|
|
|
154
|
-
**v0.
|
|
181
|
+
**v0.6.0**——已发布到npm,版本号为`@mcptoolshop/research-os@0.6.0`,发布日期为2026年5月10日。v0.6.0版本完成了实验6,并提供了评审器信任的证据:research-os现在可以生成可重复、可追溯的标准模型基准。包含内容:在生产环境的评审流程中,添加了确定性的评审器选项(`review_profiles.<name>.reviewer_options`,位于`research.yaml`文件中);为预v0.3.3版本的冻结资源(F-53)提供了向后兼容的schema;评审输出直接在`review.json`和`review.md`文件中显示采样条件(F-54);提交了标准的、确定性的聚合反馈(`hermes-two-pass-deterministic`,`temperature:0, seed:7`)。**目前没有任何可信任的基准。** `hermes-two-pass-deterministic=failed`(决策词汇表中的结构模型能力差距,而非方差问题)。**Hermes没有被提升为“可信任的基准”。** 关键在于机制,而不是通过了反馈。没有对gate、freeze或合成规则进行任何更改。所有四个冻结包都具有完全相同的字节标识。713/713个vitest测试通过。请参阅[CHANGELOG.md](CHANGELOG.md)以及[`docs/experiment-6-proof.md`](docs/experiment-6-proof.md)。
|
|
182
|
+
|
|
183
|
+
**v0.5.0** — 发布到npm,版本号为`@mcptoolshop/research-os@0.5.0`,发布日期:2026-05-10。v0.5.0版本使评审员校准更加可靠。评审员配置文件不会因为只运行一次而被信任,而是通过结构化的、带有预设错误的测试结果和多次运行的聚合来获得信任状态。包含:结构化的校准结果模式(`seeded-v1.{json,md}`,经过Zod验证,包含四个状态标签);多运行测试框架(`--runs <n>`,每个运行隔离,基于中位数的PASS/FAIL结果,重复失败降级);能够感知架构的决策词汇表;在`review-promote`中进行包相关的结果文件查找。**没有可信的基线:** `hermes-two-pass=failed`(聚合,3次运行),`mistral-nemo-two-pass=conditional_pass`,`hermes-single-pass=comparison_only`。research-os现在可以拒绝信任评审员配置文件,当反复的、带有预设错误的测试结果不支持信任时。**没有对网关、冻结或合成规则的更改。所有四个现有的冻结包都以字节级别的相同方式进行验证。** 671/671个vitest测试通过。请参阅[CHANGELOG.md](CHANGELOG.md)。
|
|
184
|
+
|
|
185
|
+
**v0.4.0** — 发布到npm,版本号为`@mcptoolshop/research-os@0.4.0`,发布日期:2026-05-10。v0.4.0版本使源代码身份更加可靠。基于确定性的源代码类型规则处理可重复的多数情况,覆盖账本保留了操作员的更正,并且`source-card audit`(源代码卡审计)取代了对临时脚本漂移的检查,提供了一个一流的命令行界面。包含:集中式的源代码类型分类器(组件B — `classifySourceType`,11个标准供应商,`source-type-rules.json`);源代码卡覆盖账本(组件A — `source-card-overrides.jsonl`,`validate` + `list`子命令);以及源代码卡审计命令行界面(组件D — `research-os source-card audit --pack <dir>`,7种发现类型,JSON + Markdown格式,`--apply --from`用于应用路径)。F-46:一个小的修复,现在包清单会记录实际的二进制版本,而不是冻结在`research.yaml`中的版本,该版本在包初始化时被冻结。**没有对网关、冻结或合成规则的更改。所有四个现有的冻结包都以字节级别的相同方式进行验证。** 620/620个vitest测试通过。请参阅[CHANGELOG.md](CHANGELOG.md)以及[源代码卡审计手册页面](https://mcp-tool-shop-org.github.io/research-os/handbook/source-card-audit/)。
|
|
155
186
|
|
|
156
187
|
**v0.3.3** — 已发布到 npm,版本号为 `@mcptoolshop/research-os@0.3.3`,发布日期:2026年5月10日。此版本改进了“门”机制的语义清晰度,这是Pack-3(Godot导出/运行时稳定性,实验3的第3个包)所取得的成果。现在,“门”的输出结果除了包含整个包的计数外,还包含按“门”划分的发布者和主要计数(F-43);`no_source_cluster_monopoly` 的警告信息已更改为信息性诊断信息(F-41)。**通过/失败的行为未改变;现有的冻结包在字节级别上进行验证。** 570/570 个 vitest 测试通过。请参阅 [CHANGELOG.md](CHANGELOG.md) 和 [`docs/section-scoped-waivers.md`](docs/section-scoped-waivers.md)。
|
|
157
188
|
|