docs: reframe root README for AI-OS overview

README_RU.md

@@ -117,6 +117,29 @@ ownership; `[Codex]` готовит implementation work, а Codex APP меняе
 | `[Codex]` | Implementation framing, code review, tests и release handoff. | Scoped execution package для repository work. |
 | `[Thinkers OS]` | Thinker corpus, provenance, synthesis и pattern status. | Source-aware synthesis без выдуманной attribution. |
 
+### Что можно показать снаружи
+
+AI-OS — не просто библиотека prompts. Его ценность можно показать через
+связанные рабочие поверхности:
+
+| Поверхность | Что она даёт | Наблюдаемый масштаб |
+|---|---|---|
+| Семь ChatGPT Projects | Разделяют входящий поток, AI-подходы, решения, аналитику, LLM, разработку и корпус источников. | Явные владельцы, границы и handoff. |
+| Analytics Factory | Ведёт от вопроса и data contract до расчёта, memo и QA. | Реестр из **22** аналитических методов. |
+| Проверка поведения | Не смешивает consistency репозитория с поведением живого ChatGPT Project. | 99 детерминированных проверок, включая **22** регрессионных кейса; live-каталог — 45 кейсов × 3 запуска. |
+| StreamDeck | Делает ежедневные маршруты и безопасные prompts доступными с двух устройств. | 16 переносимых профилей и 140 × 3 model-QA входов. |
+
+22 метода — это не «22 примера ради количества», а словарь способов проверить
+вопрос: изменение и структура, data quality и control, проверка объяснений и
+взгляд вперёд. Полный реестр с предпосылками, ограничениями и владельцами
+проверки находится в
+[ANALYTICAL_TECHNIQUES.md](<ChatGPT/[Analytics]/Knowledge/ANALYTICAL_TECHNIQUES.md>).
+
+В текущем репозитории нет вело-кейса: ни данных, ни готового анализа, ни
+пользовательского сценария про велосипеды. Поэтому он не должен выглядеть как
+уже готовая демонстрация. Такой кейс можно добавить отдельно в `[Analytics]`:
+исходные данные → выбранные методы → проверяемые выводы → memo/визуализация.
+
 Authoritative paths, instruction limits и AES applicability находятся в
 [project registry](PROJECT_REGISTRY.md). Project packages разделены намеренно:
 strategy discussion не должна молча становиться analytics calculation или

benchmarks/live_behavioral/BASELINE_CONFIGURATION.json

@@ -0,0 +1,73 @@
+{
+  "observed_at": "2026-07-31",
+  "source_commit": "8d15f5a0f6c06a14bb4098d8d6dcbe7632457dcf",
+  "benchmark_version": "1.0.1",
+  "thinking_effort": "Medium",
+  "exact_model_identifier": "UNAVAILABLE",
+  "project_memory": "Default; not configurable in the observed UI",
+  "instructions_status": "PASS_7_OF_7_NORMALIZED",
+  "instructions_normalization": "Remove at most one trailing newline from both the UI and repository values.",
+  "instructions": {
+    "[AI OS]": {"ui_sha256": "9e262da88063e6569ad2084dcdf1fca11cd5d075d4b0d36841a581fdea2ab926", "repo_sha256": "4dcf4b2dbb7d14f1f52f7778dce87c0ad4310666624282058d4c5925ed60bbd8", "normalized_sha256": "9e262da88063e6569ad2084dcdf1fca11cd5d075d4b0d36841a581fdea2ab926", "match": true},
+    "[Thinking]": {"ui_sha256": "491c601da609fb2f0dc770d96f52d89aa8d5a536c5729e2eef50f661673b63fe", "repo_sha256": "5dde30d7c0364e3eec10a38e0c3a34a39e768d9a1968d493c265da3d38549dab", "normalized_sha256": "491c601da609fb2f0dc770d96f52d89aa8d5a536c5729e2eef50f661673b63fe", "match": true},
+    "[Analytics]": {"ui_sha256": "cd977b8eaeaeb8ed93ef4a5365fa8a2017177b9177a21fc18518a3f1405f74b7", "repo_sha256": "b51574737afb37fb001d860908aae78f3856f0c2e6915e7509c71984852be7fd", "normalized_sha256": "cd977b8eaeaeb8ed93ef4a5365fa8a2017177b9177a21fc18518a3f1405f74b7", "match": true},
+    "[LLM]": {"ui_sha256": "443c323f2ac751f0dc17cc9ce4d59bdf389444e96a49771dca56b63c16b17307", "repo_sha256": "e7165b25650406f64b61a2ac40de2b871e3157dceb1ead37fe86ed1196552db2", "normalized_sha256": "443c323f2ac751f0dc17cc9ce4d59bdf389444e96a49771dca56b63c16b17307", "match": true},
+    "[Codex]": {"ui_sha256": "17789e9da1ba27a91ae3536e651374541f54f3db46678e198cb50bedb79ec86e", "repo_sha256": "793d49ef0f9ac74770efe221f9205d7af51bf3a2aa22829b23a673acf5dd6593", "normalized_sha256": "17789e9da1ba27a91ae3536e651374541f54f3db46678e198cb50bedb79ec86e", "match": true},
+    "[Inbox Router]": {"ui_sha256": "d60e8776fff384e6256aa07a715f539ecdd44baf646c41816c613766c543210b", "repo_sha256": "1ab0ebcc0681e50b28a775be824b66abe76fc56b6b9903ccbaad804a6580c6ca", "normalized_sha256": "d60e8776fff384e6256aa07a715f539ecdd44baf646c41816c613766c543210b", "match": true},
+    "[Thinkers OS]": {"ui_sha256": "3582a058bb74726266e8d4d4f7827e9d775b8b77ffb79fb642e34f02dd1aca7a", "repo_sha256": "5e419ca1be9b3c89ceb90e5d89d316a431f679d16bfbe92ea8729da876ab946d", "normalized_sha256": "3582a058bb74726266e8d4d4f7827e9d775b8b77ffb79fb642e34f02dd1aca7a", "match": true}
+  },
+  "knowledge_status": "FILENAMES_PASS_SERVER_BYTES_UNVERIFIED",
+  "knowledge_files": {
+    "[AI OS]": {
+      "AIOS_01_ROUTING_AND_WORKFLOW.md": "95ea21d47d645e961ae081bdd8c81467d865d85901bb77d478bee0db16ed2e6d",
+      "AIOS_02_GOVERNANCE_AND_EVIDENCE.md": "dbb3bae0c8b2e63aa73220e68f0c9bcc668ab79e7945c3af39f89b5a2144f975",
+      "AIOS_03_HANDOFF_AND_SMOKE_QA.md": "58eb37b5e3a7c17a6ece68f051b8ccd6cac85823623b0f39a29a6f80a8001b3c",
+      "AIOS_04_GOAL_PACKS_AND_COMMAND_SURFACE.md": "96884282ae082960f0a5fc40a5b5d1083a07c6e168b47ea9c4972b37e61b021c",
+      "AIOS_05_SUPERVISED_AGENT_LOOPS.md": "a0b48541de20ca000a341cdf50819908317447b6ab80ed21009c83748516fea6",
+      "AIOS_06_CROSS_PROJECT_AI_EVALS.md": "43834bac395123909bbff82f9108df079ad0b7413cee895c50c4dae674d616b9"
+    },
+    "[Thinking]": {
+      "THINKING_01_WORKFLOW_AND_DECISIONS.md": "53e78a4b08701b0f838033c798218806d20eb4334bcb8d42fb8f5373191049c7",
+      "THINKING_02_JUDGE_REVISOR_RISK.md": "3bd89fd3437b5a7b346e0ae2569fea41c9dc6e4059e64a1371628b98d13e438f",
+      "THINKING_03_ROUTING_AND_TEMPLATES.md": "a44abc11da9e88eae95fda8fb543b7fbfe2e18d4c6d157003020c0adae422315",
+      "THINKING_04_THINKERS_SYNTHESIS.md": "834476e2b2ecda044540b80ec5a5fe2a49b3e43d8f94c6cffb84c4595ffa3e36"
+    },
+    "[Analytics]": {
+      "ANALYTICS_01_CORE_WORKFLOW.md": "4a1ca3fe212d595a700ccfaee58e1e3486d18b4fca52fd57860b43f7fb438ee6",
+      "ANALYTICS_02_DATA_CONTRACTS_AND_MARTS.md": "b11d4db7cdd4c72cec7564dc9adcf81787f2b8b1bce81f8957bb3b68b7efd7ee",
+      "ANALYTICS_03_TECHNIQUES_AND_CHARTS.md": "7ca4449e5fcb9822c3c945289f4e2b7834c5d86c275180e28cb481457df05af0",
+      "ANALYTICS_04_MEMO_AND_TEXT_STANDARDS.md": "3f305574d51fdac5c6d42e131235f7adba3476f0d99219c6e2b6746e72bb805d",
+      "ANALYTICS_05_QA_GOVERNANCE_ROUTING.md": "b3a5f78e13454e72cc5cba07b297fea035d12f484bbf8069f2b762f6475ab914",
+      "ANALYTICS_06_TEMPLATES.md": "d2815aa82f1164c100aca4706ee3f8eab707fd64f77e61b3feb622d953409dff",
+      "ANALYTICS_07_CODEX_HANDOFF_OPTIONAL.md": "8f23c9690e92f72a1a8a6f4fc097541603118026de99a0059da0d52b4da7397e"
+    },
+    "[LLM]": {
+      "LLM_01_ROUTING_AND_MODEL_SELECTION.md": "92c8fa2dcdc12686b92b01308c183a34f2b2e5e8bfb017810c99c23f443a0b49",
+      "LLM_02_PROMPT_LIBRARY_AND_REGISTRY.md": "d55e95ab252c0eae2671f5c58161937b5a4c979c68abe6c1a49544016406b669",
+      "LLM_03_QUALITY_GATES_AND_EVAL.md": "030120ac3a62e0b9abb270cf4c91b0e4a4dc4a1dfde06d8a28adbd0c4cf17201",
+      "LLM_04_WORKFLOWS_AND_HANDOFF.md": "0df3d23c84ad25513da027e7a89c2ed6e3d0856cef01867e011f53aae2481c67",
+      "LLM_05_CONTEXT_ENGINEERING.md": "901fd59eab0945bc4d02ba160f7f568001b816e945bc32093da4deb84bed0d87",
+      "LLM_06_LOCAL_AI_EXPERIMENTS.md": "0c3365042946e299e3284dc72789496327f106b4541051af841b192657b041e1"
+    },
+    "[Codex]": {
+      "CODEX_01_TASKS_AND_HANDOFF.md": "fb04e5bed524b6b12d534a0d59302d3bdf0962cd2620393ac814eaa1ec2c05b8",
+      "CODEX_02_EXECUTION_AUTONOMY_REPORTING.md": "15ca9f00bc54209932c5e87bbbd9106f8e7c27795b6c21306c12e75d8f656f39",
+      "CODEX_03_TESTING_ACCEPTANCE_RELEASE.md": "8665957970d11a205c42001f27721af056f987800a05fd20032eec2527da26d3",
+      "CODEX_04_IMPLEMENTATION_WORKFLOWS.md": "bd382efb19903893f4bbed2ad473a8f754a4e7a135fee18b8540e108d39c31b9",
+      "CODEX_05_AGENT_REFERENCES.md": "65bd1d7b94cf7eee852531574a70e239840f2f288f6e00d677ad03b196894fe7",
+      "CODEX_06_AI_CODING_DISCIPLINE.md": "280c295e9a4ed84f36cc313d6055b93e2f0802b83061c2103215f3874bb2bbfe"
+    },
+    "[Inbox Router]": {
+      "INBOX_01_ROUTING_WORKFLOW.md": "b0240b47893a7b69232ffbde12e7613e0d599fb791ad1c2f024674aa11c90290",
+      "INBOX_02_HANDOFF_QA_ANTI_PATTERNS.md": "c52028a14f310ee8ef5631303498a9a03657b484653a63a3e3b62440f49e8050"
+    },
+    "[Thinkers OS]": {
+      "THINKERS_OS_01_PORTFOLIO_AND_CORPUS.md": "2727b5a9236a241b498b2011999312f20a19ccc74dcadcf71efce60063fad13b",
+      "THINKERS_OS_02_ARTIFACTS_AND_SYNTHESIS.md": "92b2f3ec21a39f5c43426ab8e935b92602cbe96d89c314bf3b37993ebce2f339"
+    }
+  },
+  "knowledge_ui_evidence": "All 33 declared bundle filenames were observed in their target Project. AI OS also retains the governed KB__ files documented as intentional external Knowledge.",
+  "knowledge_server_byte_equivalence": "UNVERIFIED",
+  "duplicate_policy": "Existing matching filenames were preserved. Duplicate upload prompts were skipped; no Project source was deleted and no duplicate was added.",
+  "configuration_comparability": "PASS_WITH_DOCUMENTED_EXTERNAL_RUNTIME_LIMITATIONS"
+}

benchmarks/live_behavioral/BENCHMARK_CHANGELOG.md

@@ -0,0 +1,7 @@
+# Benchmark Changelog
+
+## 1.0.1
+
+Before baseline, live UI inspection showed that `[Thinking]` preserves the repository's final newline while other Project textareas may strip it. Version 1.0.0 incorrectly normalized only the repository side and therefore misclassified `[Thinking]` as stale.
+
+Version 1.0.1 removes at most one trailing newline from both values before exact comparison. No cases, prompts, rubric criteria, weights, floors, hard-fail rules, thresholds, holdout content or sealed holdout hash changed. Version 1.0.0 was never baselined and is not evidence.

benchmarks/live_behavioral/CAPABILITY_MATRIX.md

@@ -0,0 +1,19 @@
+# Live Benchmark Capability Matrix
+
+Assessment date: 2026-07-31. Browser evidence was observed in the authenticated ChatGPT Project UI.
+
+| Capability | Status | Evidence | Limitation |
+|---|---|---|---|
+| Actual ChatGPT Project access | supported | Authenticated Project home/chat pages opened for all seven target Projects. | Session authentication is user-owned. |
+| Project selection | supported | Stable Project URLs and visible Project names were observed. | Project IDs are external runtime identifiers. |
+| Exact Instructions loading | supported | Settings textarea can be read and updated; hashes can be compared with repository Instructions after removing at most one trailing newline on both sides. | `[Inbox Router]` was stale and `[Thinkers OS]` was empty at discovery; they must be synchronized before the valid baseline. |
+| Exact Knowledge loading | supported | Required repository files can be uploaded through the Project Sources UI and filenames can be enumerated afterward. | The UI does not expose post-ingestion bytes, so server-side byte equivalence remains UNVERIFIED; source file hashes and observed filenames provide provenance. |
+| Raw response capture | supported | Full visible assistant response, prompt, chat URL and hashes can be captured from each saved Project chat. | Raw captures remain local; repository copies must be anonymized. |
+| Repeated identical prompts | supported | Fresh Project chats can be created repeatedly with the same prompt. | Product-side sampling is not controllable. |
+| Baseline/candidate separation | supported | Configuration hashes and fresh chat URLs distinguish phases. | Same external account and product runtime are used. |
+| Exact model pinning | unsupported | Project UI exposes `Medium` thinking effort but no exact model identifier in the composer. | Model comparability is UNVERIFIED if the product changes routing. |
+| Context preservation | supported | Every run starts from the same Project home in a new empty chat. | Project memory is `Default` and cannot be changed in the UI. |
+| Holdout isolation | supported | Holdout prompts are generated and sealed before baseline, then opened only by Runner/Evaluator after candidate selection. | Procedural isolation on one physical system; independent enforcement is UNVERIFIED. |
+| Independent evaluator | UNVERIFIED | Runner, evaluator and final judge use separate artifacts and hashes. | One physical system performs all roles; residual risk is self-evaluation bias. |
+
+No Level C improvement claim is allowed until actual baseline and candidate Project runs are captured and compared.

benchmarks/live_behavioral/COVERAGE.md

@@ -0,0 +1,15 @@
+# Coverage Matrix
+
+Public development benchmark: 45 cases × 3 fresh-chat runs = 135 live Project runs per phase. A separately sealed holdout is not included in public counts.
+
+| Project / route | Positive | Negative | Cross-project | Readability simple | Readability complex | Adversarial | Public cases | Runs / phase |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| [Inbox Router] | 2 | 1 | 1 | 1 | 0 | 2 | 7 | 21 |
+| [AI OS] | 2 | 1 | 1 | 1 | 1 | 1 | 7 | 21 |
+| [Thinking] | 2 | 1 | 1 | 1 | 1 | 2 | 8 | 24 |
+| [Analytics] | 2 | 1 | 0 | 0 | 1 | 1 | 5 | 15 |
+| [LLM] | 2 | 1 | 1 | 1 | 1 | 1 | 7 | 21 |
+| [Codex] | 2 | 1 | 1 | 1 | 0 | 1 | 6 | 18 |
+| [Thinkers OS] | 2 | 1 | 0 | 0 | 1 | 1 | 5 | 15 |
+
+Set totals: routing 21; response quality 5; readability 10 (5 simple, 5 material complex); adversarial 9. Every tested Project has three core routing cases: two positive and one negative. The nine critical hard-fail classes are each represented by at least one tagged public case.

benchmarks/live_behavioral/DISCOVERY_EVIDENCE.json

@@ -0,0 +1,24 @@
+{
+  "observed_at": "2026-07-31",
+  "source_commit": "8d15f5a0f6c06a14bb4098d8d6dcbe7632457dcf",
+  "instructions": {
+    "[AI OS]": {"ui_sha256": "9e262da88063e6569ad2084dcdf1fca11cd5d075d4b0d36841a581fdea2ab926", "repo_ui_normalized_sha256": "9e262da88063e6569ad2084dcdf1fca11cd5d075d4b0d36841a581fdea2ab926", "match": true},
+    "[Thinking]": {"ui_sha256": "5dde30d7c0364e3eec10a38e0c3a34a39e768d9a1968d493c265da3d38549dab", "ui_normalized_sha256": "491c601da609fb2f0dc770d96f52d89aa8d5a536c5729e2eef50f661673b63fe", "repo_ui_normalized_sha256": "491c601da609fb2f0dc770d96f52d89aa8d5a536c5729e2eef50f661673b63fe", "match": true},
+    "[Analytics]": {"ui_sha256": "cd977b8eaeaeb8ed93ef4a5365fa8a2017177b9177a21fc18518a3f1405f74b7", "repo_ui_normalized_sha256": "cd977b8eaeaeb8ed93ef4a5365fa8a2017177b9177a21fc18518a3f1405f74b7", "match": true},
+    "[LLM]": {"ui_sha256": "443c323f2ac751f0dc17cc9ce4d59bdf389444e96a49771dca56b63c16b17307", "repo_ui_normalized_sha256": "443c323f2ac751f0dc17cc9ce4d59bdf389444e96a49771dca56b63c16b17307", "match": true},
+    "[Codex]": {"ui_sha256": "17789e9da1ba27a91ae3536e651374541f54f3db46678e198cb50bedb79ec86e", "repo_ui_normalized_sha256": "17789e9da1ba27a91ae3536e651374541f54f3db46678e198cb50bedb79ec86e", "match": true},
+    "[Inbox Router]": {"ui_sha256": "922772b4c00020cb2b78b9bf89822541cb9884338e17a1778fc13af4e0a1fbe1", "repo_ui_normalized_sha256": "d60e8776fff384e6256aa07a715f539ecdd44baf646c41816c613766c543210b", "match": false},
+    "[Thinkers OS]": {"ui_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "repo_ui_normalized_sha256": "3582a058bb74726266e8d4d4f7827e9d775b8b77ffb79fb642e34f02dd1aca7a", "match": false}
+  },
+  "knowledge_filenames": {
+    "[AI OS]": ["AIOS_01_ROUTING_AND_WORKFLOW.md", "AIOS_02_GOVERNANCE_AND_EVIDENCE.md", "AIOS_03_HANDOFF_AND_SMOKE_QA.md", "AIOS_04_GOAL_PACKS_AND_COMMAND_SURFACE.md", "AIOS_05_SUPERVISED_AGENT_LOOPS.md", "AIOS_06_CROSS_PROJECT_AI_EVALS.md", "KB__00_INDEX.md", "KB__01_NAVIGATION.md", "KB__02_CONTENT.md", "KB__03_WORKFLOWS_TRACEABILITY.md", "KB__04_SMOKE_QA.md", "KB__05_CANONICAL_CONCEPTS.md", "KB__06_OPERATIONAL_FRAMEWORKS.md", "KB__07_PATTERNS_AND_FAILURES.md", "KB__08_USE_CASES_FOR_SERGEY.md", "KB__CARD_SCHEMA.md", "KB__CHANGELOG.md", "KB__CONFIDENCE_RULES.md", "KB__DEDUPLICATION.md", "KB__PROMOTION_GATES.md", "KB__RELEASE_MANIFEST.md", "KB__RETRIEVAL_QA.md", "KB__REVIEW_QUEUE.md", "KB__USE_CASE_ROUTING.md"],
+    "[Thinking]": ["THINKING_01_WORKFLOW_AND_DECISIONS.md", "THINKING_02_JUDGE_REVISOR_RISK.md", "THINKING_03_ROUTING_AND_TEMPLATES.md", "THINKING_04_THINKERS_SYNTHESIS.md"],
+    "[Analytics]": ["ANALYTICS_01_CORE_WORKFLOW.md", "ANALYTICS_02_DATA_CONTRACTS_AND_MARTS.md", "ANALYTICS_03_TECHNIQUES_AND_CHARTS.md", "ANALYTICS_04_MEMO_AND_TEXT_STANDARDS.md", "ANALYTICS_05_QA_GOVERNANCE_ROUTING.md", "ANALYTICS_06_TEMPLATES.md", "ANALYTICS_07_CODEX_HANDOFF_OPTIONAL.md"],
+    "[LLM]": ["LLM_01_ROUTING_AND_MODEL_SELECTION.md", "LLM_02_PROMPT_LIBRARY_AND_REGISTRY.md", "LLM_03_QUALITY_GATES_AND_EVAL.md", "LLM_04_WORKFLOWS_AND_HANDOFF.md", "LLM_05_CONTEXT_ENGINEERING.md", "LLM_06_LOCAL_AI_EXPERIMENTS.md"],
+    "[Codex]": ["CODEX_01_TASKS_AND_HANDOFF.md", "CODEX_02_EXECUTION_AUTONOMY_REPORTING.md", "CODEX_03_TESTING_ACCEPTANCE_RELEASE.md", "CODEX_04_IMPLEMENTATION_WORKFLOWS.md", "CODEX_05_AGENT_REFERENCES.md", "CODEX_06_AI_CODING_DISCIPLINE.md"],
+    "[Inbox Router]": ["INBOX_01_ROUTING_WORKFLOW.md", "INBOX_02_HANDOFF_QA_ANTI_PATTERNS.md"],
+    "[Thinkers OS]": ["THINKERS_OS_01_PORTFOLIO_AND_CORPUS.md", "THINKERS_OS_02_ARTIFACTS_AND_SYNTHESIS.md"]
+  },
+  "knowledge_byte_equivalence": "UNVERIFIED",
+  "knowledge_limitation": "The Project UI exposes filenames and accepts exact local uploads but does not expose post-ingestion file bytes."
+}

benchmarks/live_behavioral/HOLDOUT_MANIFEST.json

@@ -0,0 +1,11 @@
+{
+  "status": "sealed_before_baseline",
+  "case_count": 7,
+  "sealed_sha256": "e085abd624c1984d8c9d8bc0ac4812784ab0e61fa3f93bf25547668f1c034021",
+  "algorithm": "aes-256-gcm",
+  "location": "untracked local runtime artifact",
+  "optimizer_disclosure": "NOT DISCLOSED",
+  "open_stage": "after candidate selection",
+  "independent_enforcement": "UNVERIFIED",
+  "limitation": "Procedural isolation on one physical system; residual risk is self-evaluation bias."
+}

benchmarks/live_behavioral/README.md

@@ -0,0 +1,19 @@
+# SUPERManager Live Behavioral Benchmark
+
+This benchmark evaluates actual responses from the seven AI-OS ChatGPT Projects. Static repository checks do not substitute for live runs.
+
+Workflow:
+
+1. freeze benchmark, rubric, public cases, coverage and sealed holdout hashes;
+2. synchronize the source-baseline Instructions and Knowledge into the actual Projects;
+3. run every public case three times in a fresh Project chat;
+4. capture prompt, full raw response, Project URL, configuration hash, model condition and response hash;
+5. evaluate raw responses without Optimizer rationale;
+6. make at most five bounded configuration iterations;
+7. repeat the same full run procedure for the selected candidate;
+8. open the sealed holdout only after candidate selection;
+9. apply the frozen final gate and create a separate PR without merging.
+
+Raw responses and the sealed holdout stay outside the repository. The PR contains anonymized cases, response samples, aggregate results and hashes without private data.
+
+Independent evaluation: UNVERIFIED. Residual risk: self-evaluation bias.