sinapse-ai 1.21.0 → 1.22.0

package/.sinapse-ai/install-manifest.yaml

@@ -7,9 +7,9 @@
 # - SHA256 hashes for change detection
 # - File types for categorization
 #
-version: 1.21.0
+version: 1.22.0
 generator: scripts/generate-install-manifest.js
-file_count: 1154
+file_count: 1155
 files:
   - path: cli/commands/config/index.js
     hash: sha256:bfa83cb1dc111b0b30dd298dc0abc2150b73f939b6cd4458effa8e6d407bc9e2
@@ -192,9 +192,9 @@ files:
     type: cli
     size: 5908
   - path: core-config.yaml
-    hash: sha256:3ab9c9a61ed541123178c1a13ea35dd12549de8fd5812ed2266a9ab77bfad5b4
+    hash: sha256:cf8a5e6baaff0c26eaeebcdb876f6f329278e81e6b5448499b622c47772c3bf1
     type: config
-    size: 13160
+    size: 15012
   - path: core/atlas/atlas-data.js
     hash: sha256:ab6ef91bd2de1498dbe3892a80b5f18d67624db1fb2da74b0e53f83de902f383
     type: core
@@ -392,9 +392,9 @@ files:
     type: core
     size: 3671
   - path: core/doctor/checks/index.js
-    hash: sha256:898106a6aae49bdc7b4ca7518c8e7d2ba8d0614a8bfb7ecd7515862d8ed61764
+    hash: sha256:1bbf4cff649bbf88729c33df95391d342ebd7c0fa55082f7c2421c460f254659
     type: core
-    size: 1513
+    size: 1540
   - path: core/doctor/checks/manifest-version-parity.js
     hash: sha256:ffb698eed47364abccf28ce9d3a07529ac00096f6993a4d03359998b364e8329
     type: core
@@ -496,9 +496,9 @@ files:
     type: core
     size: 34474
   - path: core/execution/build-orchestrator.js
-    hash: sha256:56bbde48aaa4eb0de3ea7c95d5d1584a9ec8b753931097102fa696fb72d11848
+    hash: sha256:d115167d203651b7288c1ee2f2a3b2d1bc0b09d624d8e7bd0123c27574890b53
     type: core
-    size: 43980
+    size: 44135
   - path: core/execution/build-state-manager.js
     hash: sha256:3b2fe56bf0e1a480f663a02afbb83c6d11cf89fb56815381c7ccafab948a63ae
     type: core
@@ -512,9 +512,9 @@ files:
     type: core
     size: 14652
   - path: core/execution/subagent-dispatcher.js
-    hash: sha256:c291a8d9dea32273577b5c726626f4bc7a94b6aa67a52c9f312d6448ca6ab910
+    hash: sha256:d7c440006585bd750d88da6926a766055e4b5bcb01974f30b4fc1fe8812c91a5
     type: core
-    size: 32197
+    size: 32206
   - path: core/external-executors/delegate-cli.js
     hash: sha256:e56313ca703a4fccf64e0eec958ceb9f84e5d7a3f1c84c22b12375a9b5b58bc1
     type: core
@@ -776,9 +776,9 @@ files:
     type: core
     size: 7763
   - path: core/ideation/ideation-engine.js
-    hash: sha256:e4973a25c61101d278f102e4ed449421e50d876a274b2c4c2b150d00df97cb39
+    hash: sha256:9fbbd66ea38bc08696e42abca30810ae55e3c80aca703ec2ab0f7657eb3566be
     type: core
-    size: 27573
+    size: 27657
   - path: core/ids/circuit-breaker.js
     hash: sha256:d9ff8c6e540902457b17aba668ff67c04d40564ed479993ad3c4807fb8c09c3f
     type: core
@@ -936,9 +936,9 @@ files:
     type: core
     size: 10096
   - path: core/orchestration/cli-commands.js
-    hash: sha256:a3fc11597f2dd2098ffa0b30acbe283506c4e64132c198ddf5c233c9791ab2b0
+    hash: sha256:55fe402101964487a8ad91c99a3cf1f2af3832a58bda294244cf5abb46fd8654
     type: core
-    size: 23762
+    size: 24239
   - path: core/orchestration/condition-evaluator.js
     hash: sha256:c90c2b1506d72ce8c468b8b2b90cf4116fd1fc441eb591e8ce09dbcb75e28c00
     type: core
@@ -956,9 +956,9 @@ files:
     type: core
     size: 10880
   - path: core/orchestration/doc-first-resolver.js
-    hash: sha256:e309338d1ca98d70953563a2cf92bae261264fac8001f389afdb7fbf968be353
+    hash: sha256:70ed2b0522825e74cdd229ee2e5c9a3908a026f229a7803edfee9aa7c8a69ec5
     type: core
-    size: 14762
+    size: 19360
   - path: core/orchestration/epic-context-accumulator.js
     hash: sha256:9f38c7e3b4de7fbcfeae5fe1dc88aa3f0ae38667f1dc1da025519d40b3b9daa9
     type: core
@@ -972,13 +972,13 @@ files:
     type: core
     size: 11379
   - path: core/orchestration/executors/epic-3-executor.js
-    hash: sha256:da06c2b492104e453a8b18b1333bd2fd24123efb4933ab623045bff657b6d1d2
+    hash: sha256:adbf4790f4381727be0a5e6842d05471f92bbf6475495a1241f8799a95559add
     type: core
-    size: 9463
+    size: 10209
   - path: core/orchestration/executors/epic-4-executor.js
-    hash: sha256:ffb124ce4751d3b6189cf03210eac9ee93be3754b760cfdb0188880f03ed928f
+    hash: sha256:9cf89dad552b9ea369fd9274fe1d14d3860f5150154c73d30a7883c18ce89931
     type: core
-    size: 11680
+    size: 11868
   - path: core/orchestration/executors/epic-5-executor.js
     hash: sha256:865224cd6cb1f80c1228e149349aea5c335440201c9047f849d2ec85b73d3f03
     type: core
@@ -1008,9 +1008,9 @@ files:
     type: core
     size: 44211
   - path: core/orchestration/index.js
-    hash: sha256:ee215f855804448dcb02375128f4ff1f39e6861f37d1762dde19c80e77b8f4a6
+    hash: sha256:8e147f32b272bb7e1f1431773a788152b76277e6c51f3954e78a02a15fcf9978
     type: core
-    size: 9205
+    size: 9601
   - path: core/orchestration/lock-manager.js
     hash: sha256:b269b3c17d575f2dbca8e749585ad9133ab60dbfd7d74a85e6a776d2bb1909b4
     type: core
@@ -1043,6 +1043,10 @@ files:
     hash: sha256:321eee21d44c8c2548a4a989b4fd9bd13782dc95f917098d67c514e0886000f5
     type: core
     size: 10564
+  - path: core/orchestration/spec-complexity.js
+    hash: sha256:98263efa7189b42e8747929886eddb84b047083369d5636a401141de5b13385a
+    type: core
+    size: 4631
   - path: core/orchestration/subagent-prompt-builder.js
     hash: sha256:1b566f9cdd35df9ab33e13f36adb4e356c538665421637b2f4e57fc32bb67483
     type: core
@@ -1364,9 +1368,9 @@ files:
     type: data
     size: 9671
   - path: data/entity-registry.yaml
-    hash: sha256:c3e211656dbfb4dd5628e403728bc90ace0c1d2a1faa55385d60c8c873e46e08
+    hash: sha256:4da795f0c48efdb220f5fba360a0ee5537a6664b612db4692db94319c1f5c402
     type: data
-    size: 557940
+    size: 558557
   - path: data/learned-patterns.yaml
     hash: sha256:1a4cd045c087b9dfd7046ff1464a9d2edb85fba77cf0b6fba14f4bb9004c741e
     type: data
@@ -1472,21 +1476,21 @@ files:
     type: agent
     size: 14918
   - path: development/agents/architect.md
-    hash: sha256:b25f7a5d927b8645af278828d99af3b46e0d21b2d5dd443dbec0e8af2d195e1c
+    hash: sha256:eb26e12277ff48e41af5147a72bed8ddf96139546b07a9c75b29970edb7a912b
     type: agent
-    size: 23749
+    size: 23345
   - path: development/agents/data-engineer.md
-    hash: sha256:8f965782df49716c4d043327f8cdf430980794c0bd4984b1b3fc0dc774b2323c
+    hash: sha256:de529096dfe0e8b1cb05f664e0caa9e58b7a941bb497580cd8a4bb51d0756867
     type: agent
-    size: 23210
+    size: 22822
   - path: development/agents/developer.md
-    hash: sha256:da98be997a4b86f96b8c4f822ae189dcd4d2698d19609c6d1d2b7a56d879eeb8
+    hash: sha256:5de86f5515bf49214328bd32983bba73e7789c0864e66b18af51d2534213a199
     type: agent
-    size: 28149
+    size: 27322
   - path: development/agents/devops.md
-    hash: sha256:9c566f77c11e95edc5537e13c5f5fda65949c4cef6a1f142f81d7b35347b808b
+    hash: sha256:e3a95759b775fdf27934fe5b2a6448b7eb1c98904c2a974c3db20dd7859e8e02
     type: agent
-    size: 26954
+    size: 25968
   - path: development/agents/product-lead.md
     hash: sha256:a0d13d7e04c602000eae91dff54bf4030f11e29b4182e04c763c381086b9e7b7
     type: agent
@@ -1496,9 +1500,9 @@ files:
     type: agent
     size: 16959
   - path: development/agents/quality-gate.md
-    hash: sha256:a2a4ec39d9d75b0f9872104157094f2846453c962974f551fee1d7a2dccfbec1
+    hash: sha256:0d1dadd456826d1604d9009d12549bd0bb0b9e289775c791ad5d3094c5a8a388
     type: agent
-    size: 22302
+    size: 21342
   - path: development/agents/snps-orqx.md
     hash: sha256:8e4145cdf25f57d366e0f6b3563e907c7a3ba8b2d6d3af0a6e573927f4a0321c
     type: agent
@@ -1596,9 +1600,9 @@ files:
     type: development
     size: 10414
   - path: development/knowledge-base/token-economy-guide.md
-    hash: sha256:cd92d1becb3ea5e728a041fef5d7b850e66996ce1d2222a1ff1604758867d948
+    hash: sha256:bb10715f48778453f5f44bb99244cc4c000de06117ef4e123fe5795a2d30887f
     type: development
-    size: 5932
+    size: 6004
   - path: development/README.md
     hash: sha256:cb3aa87366eba1adebeb1f033c62e3cfc2a65430dc456914808b2c108884ebf2
     type: development
@@ -2496,9 +2500,9 @@ files:
     type: task
     size: 8687
   - path: development/tasks/resolve-github-issue.md
-    hash: sha256:e4773924e8452f0d55712eca51ed1168c916f1864e1d2791830c926bfe9ae167
+    hash: sha256:288ab27dfcc8337e4502bfc4e713b03f39635bc3bdd25a78ff1c3953ac02e0a1
     type: task
-    size: 18286
+    size: 18277
   - path: development/tasks/review-contributor-pr.md
     hash: sha256:e5a4aabbb2d35bdbc8fd8c624df1c38bff3106eadb6d1b961143d27c8b139b57
     type: task
@@ -4240,9 +4244,9 @@ files:
     type: template
     size: 10990
   - path: product/templates/story-tmpl.yaml
-    hash: sha256:43a050de5ef51f53f890fd20f77ac567b297a40597f467b3bc6d00c8193c7fa6
+    hash: sha256:3dbcd522e69b86715f32eaf0cd13ce3a2631b5d2fc4f068c33df9593b6dbbfc2
     type: template
-    size: 16686
+    size: 17050
   - path: product/templates/story.hbs
     hash: sha256:126c55c30e2e49816acc7c1712e71e7c26874b010edfe34e8d7890bac67095b2
     type: template

package/.sinapse-ai/product/templates/story-tmpl.yaml

@@ -97,7 +97,12 @@ sections:
   - id: acceptance-criteria
     title: Acceptance Criteria
     type: numbered-list
-    instruction: Copy the acceptance criteria numbered list from the epic file
+    instruction: >-
+      Write each acceptance criterion in executable Given/When/Then form, tracing
+      to the epic's numbered ACs. Example: "AC1: Given <context/state>, When
+      <action>, Then <observable outcome>." GWT is the preferred format
+      (story-lifecycle.md) — `npm run validate:story-acs` flags non-GWT ACs as
+      advisory. Free-form ACs remain valid but should be the exception.
     elicit: true
     owner: scrum-master
     editors: [scrum-master]
@@ -307,11 +312,11 @@ sections:
       - If there were important notes from previous story that are relevant to this one, include them here
       - Put enough information in this section so that the dev agent should NEVER need to read the architecture documents, these notes along with the tasks and subtasks must give the Dev Agent the complete context it needs to comprehend with the least amount of overhead the information to complete the story, meeting all AC and completing all tasks+subtasks
 
-      **Opus 4.7 Optimization — Upfront Spec Block (REQUIRED):**
+      **Frontier-Model Optimization — Upfront Spec Block (REQUIRED):**
       The Dev Notes MUST begin with an "Upfront Spec" structured block (see `upfront-spec` sub-section).
       This consolidates intent + constraints + AC summary + key file locations at the top of Dev Notes,
       so the developer agent loads the full problem specification in one read instead of discovering
-      it mid-implementation. Adaptive thinking in Opus 4.7 benefits strongly from upfront specification.
+      it mid-implementation. Adaptive thinking in frontier models (Opus/Fable family) benefits strongly from upfront specification.
     elicit: true
     owner: scrum-master
     editors: [scrum-master]
@@ -322,7 +327,7 @@ sections:
           **REQUIRED — appears at the TOP of Dev Notes.**
 
           Consolidate the full problem specification in a single structured block, so the developer
-          agent has everything needed before making the first tool call. This is an Opus 4.7 best
+          agent has everything needed before making the first tool call. This is a frontier-model best
           practice: upfront specification reduces mid-task re-planning and improves adaptive thinking
           quality.
 
@@ -348,11 +353,11 @@ sections:
       - id: prompt-hints
         title: Prompt Hints by Story Type
         instruction: |
-          **Opus 4.7 Adaptive Thinking Hints (STRUCTURED FIELDS — survive YAML parse).**
+          **Adaptive Thinking Hints (STRUCTURED FIELDS — survive YAML parse).**
 
           Emit the appropriate hint based on story type classification. These are NOT comments —
           they are structured fields consumed by the developer agent at story load time to calibrate
-          thinking depth for Opus 4.7's adaptive reasoning.
+          thinking depth for the frontier model's adaptive reasoning.
 
           Story type classification:
           - `critical` → migration, security, architecture, data integrity, irreversible operations

package/.sinapse-ai/quality/judge-calibration/README.md

@@ -0,0 +1,75 @@
+# LLM-Judge Calibration
+
+> Story `mesa2-llm-judge-calibration` · AF-20260704 Mesa #2.
+> Addresses the finding "the LLM judge decides Done without calibration" — the literal
+> gap (no golden set, no documented process) is now closed; the semantic layer (v2) has a
+> baseline + procedure but no recorded agreement number yet (it has not been run).
+
+## What the judge is
+
+The agent that decides whether a story is **Done** is `@quality-gate`, invoked by
+`Epic6Executor` (`.sinapse-ai/core/orchestration/executors/epic-6-executor.js`). It emits
+free-form review prose ending in a line `VERDICT: APPROVED | NEEDS_REVISION | BLOCKED`.
+`_parseVerdict(output)` turns that prose into the binary verdict — and `APPROVED` is what
+marks the work Done. So there are **two layers** to a correct verdict:
+
+1. **Semantic judgment (the LLM):** did the agent reach the *right* verdict for the change?
+2. **Verdict extraction (deterministic code):** did `_parseVerdict` correctly read the verdict
+   the agent expressed? This is the seam where "LLM prose" becomes "system decides Done."
+
+## The golden set
+
+`golden-set.json` is the **human baseline**: realistic review outputs, each labeled with the
+verdict a human considers correct, with a rationale. It is balanced across all three verdicts and
+includes the tricky edges (no explicit `VERDICT:` line, spaced/hyphenated tokens, conflicting
+signals where an explicit verdict must win, empty output that must never read as approval, and the
+"a blocker dominates" precedence).
+
+A golden label is **law**. If a case is ambiguous, fix the label — do not loosen the gate.
+
+## v1 — deterministic calibration (automated, CI-gated)
+
+`npm run calibrate:judge` runs the **real** `_parseVerdict` (imported via
+`Epic6Executor.prototype`, never re-implemented) over every golden output and requires **100%**
+agreement with the human labels. It reports an expected→predicted confusion map. Because
+`_parseVerdict` is deterministic, any disagreement means one of two things — the parser regressed,
+or a label is wrong — and both demand action. This is also enforced in CI by
+`tests/scripts/calibrate-judge.test.js`.
+
+What this layer guarantees is the **safe direction** of the extraction seam: for the recognized
+verdict vocabulary it reads the expressed verdict faithfully, and outside it, it **never fabricates
+approval and never silently flips an explicit verdict** — out-of-vocabulary or prose-only approvals
+fall to the safe non-approving default by design.
+
+## v2 — semantic calibration (live, periodic, NOT in CI)
+
+Calibrating the LLM's *judgment* requires running the real `@quality-gate` agent, which is
+non-deterministic and costly (and spawns a nested `claude` process that is unreliable on some
+platforms — see the executor's Windows note). It is therefore a **documented periodic procedure**,
+not a CI gate:
+
+1. Take the `scenario` behind each golden case (the code change + test state that would produce
+   such a review) from `scenarios.json`, feed it to the real `@quality-gate` judge.
+2. Compare the agent's emitted verdict (extracted with the production `_parseVerdict`) to the
+   golden `expected`.
+3. Track agreement over time. Target: **≥ 90%** agreement on the golden scenarios; below that, the
+   judge prompt/rubric needs work.
+
+### Latest measurement
+
+**First recorded run — 2026-07-06: 40/42 = 95.2% agreement** (13/14 scenarios majority-correct;
+**perfect 15/15 on all CRITICAL/blocking cases**). Above the 90% target; the single divergence is a
+known-ambiguous MEDIUM-vs-HIGH severity boundary, not a systematic gap. Full breakdown, confusion
+matrix, and divergence analysis in [`calibration-log.md`](./calibration-log.md).
+
+> **Honesty note.** v1 does not auto-calibrate the LLM's judgment — it locks the deterministic seam
+> and provides the human baseline + procedure for the semantic layer. Do not claim the LLM judge is
+> "calibrated" from a green CI alone. The v2 number above is a point-in-time live measurement, not a
+> CI-enforced guarantee; re-run periodically (labels are law — a systematic miss becomes rubric work,
+> never a label edit).
+
+## Feedback loop
+
+When the judge decides Done **wrongly** in production (approves broken work, or blocks good work),
+capture that review output as a **new golden case** with the correct human label. The golden set
+grows from real misses, so the calibration tracks reality instead of a frozen synthetic snapshot.

package/.sinapse-ai/quality/judge-calibration/calibration-log.md

@@ -0,0 +1,70 @@
+# LLM-Judge Calibration Log
+
+Versioned record of semantic (v2) calibration runs. Each entry measures the real LLM judge's
+agreement with the human-labeled scenarios in `scenarios.json`, using the production
+`_parseVerdict` for extraction. Labels are law; a systematic miss becomes rubric analysis here,
+never a label edit to inflate agreement.
+
+---
+
+## Run 2026-07-06 — first recorded measurement
+
+| Field | Value |
+|-------|-------|
+| Date | 2026-07-06 |
+| Judge model | frontier tier (Opus-class), via the QA executor prompt contract + quality-gate severity rubric |
+| Scenarios | 14 (human-labeled: 4 approved, 5 blocked, 5 needs_revision) |
+| Judgments | 42 (3 blind independent judgments per scenario) |
+| Extraction | production `Epic6Executor.prototype._parseVerdict` (no re-implementation) |
+| **Agreement** | **40/42 = 95.2%** |
+| Scenarios majority-correct | 13/14 |
+| Target (README v2) | ≥ 90% — **met** |
+
+### Confusion matrix (expected → predicted)
+
+| Expected \ Predicted | approved | needs_revision | blocked |
+|---|---|---|---|
+| **approved** (12) | 12 | 0 | 0 |
+| **needs_revision** (13) | 2 | 13 | 0 |
+| **blocked** (15) | 0 | 0 | 15 |
+
+Read: every `blocked` scenario (all 15 judgments across 5 critical scenarios — SQL injection,
+hardcoded secret, RLS dropped, auth bypass, destructive migration) was caught with **zero misses**.
+Every clear `approved` and `needs_revision` scenario was unanimous. The only divergence is a single
+borderline scenario, below.
+
+### Divergences (2 of 42)
+
+**`stale-todo-and-debug-logs`** — human label `needs_revision`, judge said `approved` on 2 of 3
+independent judgments (the 3rd agreed: `needs_revision`).
+
+- Scenario: implementation meets all ACs and 34 tests pass, but the diff leaves 6
+  `console.log('DEBUG cart=', cart)` in the payment path and 2 unresolved `TODO: 3DS fallback` on
+  code that throws a generic error.
+- Judge reasoning (approvers): treated both as **MEDIUM technical debt** — "remove the debug logs in
+  a fast-follow cleanup and open a tracked story… since ACs are met, tests green, no security/data
+  risk, this does not block completion."
+- Judge reasoning (the dissenter): "the payment-path debug logs are a cheap fix I expect done before
+  merge, which puts this just short of approval."
+- **Analysis:** this is a legitimate **MEDIUM-vs-HIGH severity boundary call**, not a miscalibration.
+  The human label treats "ship debug noise + unresolved TODOs in the changed lines" as fix-before-merge
+  (HIGH → needs_revision); the judge majority treats it as document-as-debt (MEDIUM → approved). The
+  split judgment (2 approve / 1 revise) shows the model itself sees the case as borderline. Reasonable
+  human reviewers disagree here too.
+
+### Verdict on this run
+
+The judge is **well-calibrated** at 95.2%, comfortably above the 90% target, with a **perfect record
+on all CRITICAL/blocking cases** — the highest-stakes class. The one divergence is a known-ambiguous
+MEDIUM/HIGH boundary, not a systematic gap. **No rubric change is warranted from this run.** If a
+future run shows the judge systematically down-grading fix-before-merge issues to debt (a pattern, not
+a single borderline case), that would justify a story to sharpen the severity rubric's MEDIUM/HIGH
+line — captured here as a watch item, not an action.
+
+### Method notes
+
+- Judgments were **blind**: scenarios were fed without their expected labels; each of the 3 judgments
+  per scenario ran as an independent agent with no shared context.
+- Prompt contract mirrors production (`epic-6-executor.js` `_reviewViaAgent`): free-form review prose
+  ending in a single `VERDICT:` line, plus the quality-gate severity policy.
+- Raw judgment outputs are not committed (bulk); this log is the durable aggregate.

package/.sinapse-ai/quality/judge-calibration/golden-set.json

@@ -0,0 +1,105 @@
+{
+  "$schema": "./golden-set.schema.md",
+  "description": "Human-labeled golden set for LLM-judge verdict calibration (Story mesa2-llm-judge-calibration, AF-20260704 Mesa #2). Each case is a realistic @quality-gate review OUTPUT and the verdict a human considers CORRECT for it. The harness runs the REAL Epic6Executor._parseVerdict over each output and must agree 100% — a disagreement means either the parser regressed or a label is wrong; both demand action.",
+  "verdicts": ["approved", "needs_revision", "blocked"],
+  "cases": [
+    {
+      "id": "clear-approve-explicit",
+      "output": "All acceptance criteria met, tests green, no security concerns.\nVERDICT: APPROVED",
+      "expected": "approved",
+      "rationale": "Explicit approving verdict line with clean review.",
+      "source": "synthetic"
+    },
+    {
+      "id": "clear-block-explicit",
+      "output": "Critical: SQL built via string interpolation in user path — injection risk.\nVERDICT: BLOCKED",
+      "expected": "blocked",
+      "rationale": "Explicit blocking verdict; critical security defect.",
+      "source": "synthetic"
+    },
+    {
+      "id": "needs-revision-explicit",
+      "output": "Logic is correct but missing tests for the error branch and a stale comment.\nVERDICT: NEEDS_REVISION",
+      "expected": "needs_revision",
+      "rationale": "Explicit needs-revision; minor, fixable issues.",
+      "source": "synthetic"
+    },
+    {
+      "id": "needs-revision-spaced",
+      "output": "Coverage dipped on the new module.\nVERDICT: NEEDS REVISION",
+      "expected": "needs_revision",
+      "rationale": "Verdict written with a space instead of underscore — still a needs-revision.",
+      "source": "synthetic"
+    },
+    {
+      "id": "lowercase-prose-verdict",
+      "output": "looks good to me, ship it. verdict: approved",
+      "expected": "approved",
+      "rationale": "Case-insensitive extraction: a lowercase verdict line still means approved.",
+      "source": "synthetic"
+    },
+    {
+      "id": "verdict-with-trailing-text",
+      "output": "Solid refactor, interfaces stable.\nVERDICT: APPROVED — nice work on the edge cases.",
+      "expected": "approved",
+      "rationale": "Trailing prose after the verdict token must not change the decision.",
+      "source": "synthetic"
+    },
+    {
+      "id": "conflicting-praise-but-blocked",
+      "output": "Great structure overall and readable code. However VERDICT: BLOCKED because auth is bypassable.",
+      "expected": "blocked",
+      "rationale": "When an explicit verdict is present it is authoritative, even amid praise.",
+      "source": "synthetic"
+    },
+    {
+      "id": "no-verdict-prose-blocked",
+      "output": "This change is BLOCKED: it drops the RLS policy on a public table.",
+      "expected": "blocked",
+      "rationale": "No verdict line, but an unambiguous block signal in prose → block.",
+      "source": "synthetic"
+    },
+    {
+      "id": "no-verdict-prose-approved",
+      "output": "Everything checks out and it is APPROVED for merge.",
+      "expected": "approved",
+      "rationale": "No verdict line, unambiguous approval word → approved.",
+      "source": "synthetic"
+    },
+    {
+      "id": "no-verdict-ambiguous-concerns",
+      "output": "The reviewer raised a few concerns worth discussing before we proceed.",
+      "expected": "needs_revision",
+      "rationale": "No verdict and no clear approve/block signal → safe non-approving default.",
+      "source": "synthetic"
+    },
+    {
+      "id": "empty-output",
+      "output": "   ",
+      "expected": "needs_revision",
+      "rationale": "Nothing to go on must never be read as approval.",
+      "source": "synthetic"
+    },
+    {
+      "id": "no-verdict-both-words-block-wins",
+      "output": "Parts look approved, but this is blocked by a missing migration rollback.",
+      "expected": "blocked",
+      "rationale": "With no verdict line and both signals, a block must dominate — you do not ship over an unresolved blocker.",
+      "source": "synthetic"
+    },
+    {
+      "id": "hyphenated-needs-revision-defaults-safe",
+      "output": "Tighten the naming.\nVERDICT: NEEDS-REVISION",
+      "expected": "needs_revision",
+      "rationale": "A hyphenated token is not matched as an explicit verdict, but the safe default lands on the same human-intended needs-revision.",
+      "source": "synthetic"
+    },
+    {
+      "id": "multiline-review-then-approve",
+      "output": "Reviewed correctness, tests, and NFRs.\n- Correctness: traces to all ACs.\n- Tests: added for the new branch.\n- Security: input validated.\nNo blocking issues found.\nVERDICT: APPROVED",
+      "expected": "approved",
+      "rationale": "Realistic multi-line review ending in an approving verdict.",
+      "source": "synthetic"
+    }
+  ]
+}

package/.sinapse-ai/quality/judge-calibration/scenarios.json

@@ -0,0 +1,90 @@
+{
+  "description": "Semantic golden set for LLM-judge calibration v2 (Story mesa2-judge-calibration-v2-run). Each case is a review SCENARIO (story + code changes + test state) with the verdict a competent human reviewer — applying the framework's severity policy (CRITICAL → blocked; HIGH/fixable issues → needs_revision; clean → approved) — considers correct. The live procedure feeds each scenario to the real LLM judge (executor prompt contract), extracts the verdict with the production _parseVerdict, and compares against these labels. Labels are law: a systematic judge miss becomes rubric analysis in calibration-log.md, never a label adjustment to inflate agreement.",
+  "severity_policy": "APPROVED = no CRITICAL/HIGH issues, ACs met, tests green. NEEDS_REVISION = fixable issues (missing tests, coverage drop, unmet AC, failing tests). BLOCKED = critical issues (security vulnerability, data exposure, destructive operation without safeguards).",
+  "cases": [
+    {
+      "id": "clean-approve",
+      "expected": "approved",
+      "rationale": "Everything a reviewer checks is green — nothing to fix.",
+      "scenario": "Story 'user-profile-page' with 3 ACs (render profile data, handle missing avatar, unit tests for both). Diff: 2 new components + 1 test file. All 3 ACs implemented and traceable. CI: 214/214 tests pass, lint clean, typecheck clean. No security-sensitive surface touched. Code follows existing project patterns."
+    },
+    {
+      "id": "small-fix-with-regression-test",
+      "expected": "approved",
+      "rationale": "Root cause fixed with a regression test covering it — the ideal bug-fix shape.",
+      "scenario": "Bug-fix story: date parser crashed on single-digit months. Diff: 1-line fix in parseDate() plus a new regression test that reproduces the original crash input and passes. Full suite green (1,102/1,102). No other files touched."
+    },
+    {
+      "id": "docs-only-clean",
+      "expected": "approved",
+      "rationale": "Documentation-only change with valid links and no runtime surface.",
+      "scenario": "Story 'update-install-guide': diff touches only README.md and docs/installation/guide.md — fixes 4 outdated CLI flag names and adds a troubleshooting section. Link checker passes, markdown lint passes. No code files changed, test suite untouched and green."
+    },
+    {
+      "id": "stable-refactor",
+      "expected": "approved",
+      "rationale": "Behavior-preserving refactor: interfaces stable, tests prove no regression.",
+      "scenario": "Refactor story: extract duplicated pagination logic from 3 API handlers into a shared helper. Public interfaces unchanged (verified by unchanged integration tests). All 890 tests pass unmodified. Coverage steady at 84%. Diff is -120/+45 lines."
+    },
+    {
+      "id": "sql-injection",
+      "expected": "blocked",
+      "rationale": "SQL built by string concatenation from user input — critical injection vulnerability.",
+      "scenario": "Story 'search-endpoint': new GET /search builds its query as \"SELECT * FROM products WHERE name LIKE '%\" + req.query.q + \"%'\" — user input concatenated directly into SQL, no parameterization. Tests pass (16/16) because none exercise malicious input. Everything else in the diff is fine."
+    },
+    {
+      "id": "hardcoded-secret",
+      "expected": "blocked",
+      "rationale": "Plaintext credential committed to the repository — critical exposure.",
+      "scenario": "Story 'payment-webhook': the diff hardcodes a live Stripe secret key (a real 'sk_live_'-prefixed production credential, redacted here as <STRIPE_LIVE_KEY>) directly in src/payments/webhook.js instead of reading it from environment variables. The webhook logic itself is correct and its 8 tests pass."
+    },
+    {
+      "id": "rls-dropped",
+      "expected": "blocked",
+      "rationale": "Migration removes row-level security from a user-data table — critical data exposure.",
+      "scenario": "Story 'fix-orders-query-perf': to make a slow query faster, the migration includes ALTER TABLE orders DISABLE ROW LEVEL SECURITY; and drops the tenant-isolation policy. Query benchmarks improve 40x. All existing tests pass because the test suite runs as service role."
+    },
+    {
+      "id": "praise-worthy-but-auth-bypass",
+      "expected": "blocked",
+      "rationale": "Elegant code cannot outweigh an authentication bypass — a blocker dominates praise.",
+      "scenario": "Story 'admin-dashboard': beautifully structured code, clean componentization, 96% coverage on the new module. However, the /api/admin/* routes check authorization via a client-supplied header ('x-is-admin: true') with no server-side session validation — any user can set the header and access admin endpoints."
+    },
+    {
+      "id": "destructive-migration-no-safeguard",
+      "expected": "blocked",
+      "rationale": "DROP TABLE without safeguards or rollback in a production migration — critical destructive operation.",
+      "scenario": "Story 'remove-legacy-cart': migration runs DROP TABLE legacy_cart; directly — no IF EXISTS, no backup step, no rollback script, and the table still has 12k rows referenced by an archived-orders report job. Code changes removing the ORM model are otherwise clean and tests pass."
+    },
+    {
+      "id": "missing-error-branch-tests",
+      "expected": "needs_revision",
+      "rationale": "Correct logic but the error path is untested — fixable gap, not a blocker.",
+      "scenario": "Story 'csv-import': the happy path is implemented correctly and covered by 6 tests. The error branch (malformed row → collect error and continue) is implemented but has zero test coverage; the try/catch was never exercised by any test. No security issues. Lint/typecheck clean."
+    },
+    {
+      "id": "coverage-drop-new-module",
+      "expected": "needs_revision",
+      "rationale": "New module lands with far lower coverage than the project floor — fix before merge.",
+      "scenario": "Story 'notifications-service': the new src/notifications/ module lands with 31% line coverage while the project enforces a soft floor of 75%. The 4 tests that exist pass; core send/retry logic is untested. Functionality demoed manually and works. No security-sensitive code."
+    },
+    {
+      "id": "failing-tests-on-ci",
+      "expected": "needs_revision",
+      "rationale": "Red suite means not done — but failures are ordinary logic bugs, fixable, not a security block.",
+      "scenario": "Story 'timezone-support': implementation is complete per ACs, but CI shows 3 failing tests in date-formatting for UTC-negative offsets (off-by-one on the day boundary). No security surface. The other 428 tests pass."
+    },
+    {
+      "id": "unmet-acceptance-criterion",
+      "expected": "needs_revision",
+      "rationale": "Green tests do not compensate for a story AC that was never implemented.",
+      "scenario": "Story 'export-report' has 3 ACs: (1) export CSV, (2) export PDF, (3) email the file. The diff implements CSV and PDF with passing tests (22/22 green), but AC3 (email delivery) is entirely absent — no code, no test, not mentioned in the PR. No defects in what exists."
+    },
+    {
+      "id": "stale-todo-and-debug-logs",
+      "expected": "needs_revision",
+      "rationale": "Functional but ships debug noise and unresolved TODOs in the changed lines — clean up before merge.",
+      "scenario": "Story 'checkout-flow': implementation meets all ACs and 34 tests pass. The diff, however, leaves 6 console.log('DEBUG cart=', cart) statements in the payment path and 2 'TODO: handle 3DS fallback properly' comments on code that currently throws a generic error for that case. No security issues."
+    }
+  ]
+}