docs: explain 21-model observations and evaluator sensitivity

PUBLICATION_MANIFEST.json

@@ -6,7 +6,7 @@
     "Technical-preview source; technical acceptance does not authorize publication.",
     "Verified-base evidence is scoped to the accepted clean source tree. Changed candidates require an exact-source controlled acceptance record and successful hosted CI.",
     "Project name, licence split and author/copyright attribution are owner-confirmed; know-how/rights review and public visibility authorization remain separate human decisions.",
-    "Scope revision 0.3.0: the original prepublication blanket results ban is preserved in Git history. The owner directed the bounded retrospective descriptive table after being informed that independent human validation and the old claims gate remained pending. This explicit exception does not authorize official release or establish independent approval; the owner decision is separately recorded in controlled records outside public Git."
+    "Scope revision 0.3.0: the original prepublication blanket results ban is preserved in Git history. The owner directed the bounded retrospective descriptive table after being informed that independent human validation and the old claims gate remained pending. This explicit exception does not authorize official release or establish independent approval; the owner decision is separately recorded in controlled records outside public Git. Scope extension: the owner-authorized 21-model, two-series descriptive study and approved aggregates, charts and safe semantic summaries are limited to docs/studies/2026-10-aggregates.json, docs/studies/2026-10-case-churn.svg, docs/studies/2026-10-compliance-diagnostic.csv, docs/studies/2026-10-goal-strict.svg, docs/studies/2026-10-model-flows.csv, docs/studies/2026-10-models.csv, docs/studies/2026-10-observations.md. This does not authorize future results, raw traces or billing disclosure; independent human validation and official-release boundaries remain unchanged."
   ],
   "release": {
     "project": "OperateBench",
@@ -144,7 +144,7 @@
     {
       "class": "unvalidated_results",
       "description": "Unsupported official leaderboards, composite capability scores, universal rankings, production capability claims and general named-model comparisons; unvalidated results outside the explicit descriptive-table exception.",
-      "reason": "Independent human semantic validation and official-release claims review remain pending. Only docs/MODEL_RESULTS.md is excepted under the bounded owner-directed descriptive scope; this is not a validation PASS."
+      "reason": "Independent human semantic validation and official-release claims review remain pending. docs/MODEL_RESULTS.md is the primary historical path excepted under the bounded owner-directed descriptive scope; this is not a validation PASS. Scope extension: the owner-authorized 21-model, two-series descriptive study and approved aggregates, charts and safe semantic summaries are limited to docs/studies/2026-10-aggregates.json, docs/studies/2026-10-case-churn.svg, docs/studies/2026-10-compliance-diagnostic.csv, docs/studies/2026-10-goal-strict.svg, docs/studies/2026-10-model-flows.csv, docs/studies/2026-10-models.csv, docs/studies/2026-10-observations.md. This does not authorize future results, raw traces or billing disclosure; independent human validation and official-release boundaries remain unchanged."
     },
     {
       "class": "source_history",
@@ -198,7 +198,7 @@
       "value": true,
       "path": "docs/MODEL_RESULTS.md",
       "scope": "owner_directed_retrospective_descriptive_only",
-      "statement": "One English table reports all 19 exact IDs: a 16-model, 704-assignment historical campaign and three separately executed successors, with distinct cohorts and no ranking across cohorts. One trial per model on each of 44 related synthetic cases; assigned and scored denominators are disclosed, with technical exclusions excluded from scored denominators rather than counted as failures. Costs are estimates, not invoices. Raw traces remain private. Independent human semantic validation is pending; the owner decision is separately recorded, not independent approval. No official Lifecycle Track leaderboard, universal ranking or production capability claim is authorized.",
+      "statement": "One English table reports all 19 exact IDs: a 16-model, 704-assignment historical campaign and three separately executed successors, with distinct cohorts and no ranking across cohorts. One trial per model on each of 44 related synthetic cases; assigned and scored denominators are disclosed, with technical exclusions excluded from scored denominators rather than counted as failures. Costs are estimates, not invoices. Raw traces remain private. Independent human semantic validation is pending; the owner decision is separately recorded, not independent approval. No official Lifecycle Track leaderboard, universal ranking or production capability claim is authorized. Scope extension: the owner-authorized 21-model, two-series descriptive study and approved aggregates, charts and safe semantic summaries are limited to docs/studies/2026-10-aggregates.json, docs/studies/2026-10-case-churn.svg, docs/studies/2026-10-compliance-diagnostic.csv, docs/studies/2026-10-goal-strict.svg, docs/studies/2026-10-model-flows.csv, docs/studies/2026-10-models.csv, docs/studies/2026-10-observations.md. This does not authorize future results, raw traces or billing disclosure; independent human validation and official-release boundaries remain unchanged.",
       "human_semantic_validation": "pending",
       "official_release_authorized": false
     }

PUBLIC_RELEASE_CHECKLIST.md

@@ -151,7 +151,7 @@ and confirm none of the following appears in any form:
       BoundaryBench run is Lifecycle evidence.
 
 **Current-scope exception (manifest revision 0.3.0).** The original blanket ban
-remains historical in Git. Only the independently labelled, owner-directed
+remains historical in Git. The independently labelled, owner-directed
 retrospective [descriptive table](docs/MODEL_RESULTS.md) is excepted: one English
 table of 19 exact IDs, comprising the 16-model, 704-assignment historical campaign
 and three separately executed successors. Cohorts remain distinct, with no
@@ -164,6 +164,21 @@ and is not independent approval. This exception does not close G8, G11 or G12,
 permit official/general claims, or bypass exact-commit review, privacy, secrets,
 licensing or other official-release blockers.
 
+The bounded exception also covers the owner-authorized 21-model, two-series
+[descriptive study](docs/studies/2026-10-observations.md), approved aggregates,
+charts and safe semantic summaries, limited to these six companions:
+
+- `docs/studies/2026-10-aggregates.json`
+- `docs/studies/2026-10-case-churn.svg`
+- `docs/studies/2026-10-compliance-diagnostic.csv`
+- `docs/studies/2026-10-goal-strict.svg`
+- `docs/studies/2026-10-model-flows.csv`
+- `docs/studies/2026-10-models.csv`
+
+This extension authorizes no future results, raw traces or billing disclosure.
+Independent human semantic validation remains pending; the same G8, G11, G12 and
+official-release boundaries above apply.
+
 Positively present:
 
 - [ ] Prior art conceded in `docs/RELATED_WORK.md`, with sources.

README.md

@@ -12,6 +12,8 @@ external events change independently of the agent.
 > has one Semantic Scenario with three frozen scenario variants of synthetic
 > Maintenance. Commerce and Property Compliance are separate SDK incubators,
 > not additional official benchmark evidence. Boundary is a separate track.
+> Read the [21-model, two-series October observation study](docs/studies/2026-10-observations.md) for goal versus strict checks, case churn and a separate Compliance correction diagnostic. The historical 19-model table below is preserved.
+>
 > The [model results page](docs/MODEL_RESULTS.md)
 > reports a descriptive full44 comparison of 19 exact IDs.
 > No official Lifecycle model results or leaderboard are published. No proven incremental lifecycle

docs/MODEL_RESULTS.md

@@ -1,5 +1,7 @@
 # Model results — descriptive observations: 44 related synthetic cases per model, one trial per case; separate sources/settings; no independent human review yet; not a general-capability ranking
 
+See the [21-model, two-series October observation study](studies/2026-10-observations.md) for outcome versus strict checks, changed cases and a separate Compliance correction diagnostic. This historical 19-model source table and its scores are preserved.
+
 | Model | Series (distinct runs/settings) | Scored / assigned | Unscored (excluded / aborted) | Goal achieved | All quality checks | Estimated cost (USD, not invoice) |
 | --- | --- | --- | --- | --- | --- | --- |
 | `gpt-6-astra` | [Baseline original campaign](https://github.com/DwellyOrg/OperateBench/commit/58b36826894c4306aa3ab) | 44/44 | — | 44/44 (100.0%) | 42/44 (95.5%) | $239 |