instar 1.3.927 → 1.3.928

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "instar",
-  "version": "1.3.927",
+  "version": "1.3.928",
   "description": "Coherence infrastructure for self-evolving AI agents — on the Claude Code or Codex subscription you already have.",
   "type": "module",
   "main": "dist/index.js",

package/src/data/benchmarkPredictions.json

@@ -1,6 +1,6 @@
 {
  "_doc": "Content-free INSTAR-Bench predictions mirror (benchmark-divergence-detector FD1). Counts, not just rates, so the prediction carries its own sampling uncertainty. Regenerate by re-running the battery against the CURRENT prompt templates and re-stamping benchedPromptHash.",
- "capturedAt": "2026-07-23T22:30:00.000Z",
+ "capturedAt": "2026-07-24T01:20:00.000Z",
  "tasks": {
   "tone-gate": {
    "perModel": {
@@ -28,6 +28,30 @@
    "benchedPromptSource": "src/core/MessagingToneGate.ts:TONE_GATE_PROMPT_TEMPLATE",
    "benchedPromptHash": "4ae0cf6b02663c21a695393727bd05662334f0d8f1a9d3e45d8d9b2141208d54",
    "capturedAt": "2026-07-23T22:30:00.000Z"
+  },
+  "zombie-classify": {
+   "perModel": {
+    "gpt-5.5": {
+     "passRate": 0.9,
+     "passes": 9,
+     "deterministic": 10
+    },
+    "gpt-5.4-mini": {
+     "passRate": 0.9,
+     "passes": 9,
+     "deterministic": 10
+    },
+    "gemini-2.5-flash": {
+     "passRate": 0.8333,
+     "passes": 5,
+     "deterministic": 6
+    }
+   },
+   "benchedPromptSource": "src/monitoring/ExternalHogClassifierPrompt.ts:EXTERNAL_HOG_CLASSIFIER_PROMPT_TEMPLATE",
+   "benchedPromptHash": "98b026dbe3ee48cab1403ef24218e9e24f6ea2b3d598c23e383bed4e894cbb56",
+   "capturedAt": "2026-07-24T01:20:00.000Z",
+   "scoredPopulation": "production-candidate only (10 of 17 cases). Excluded by design: 4 floor-excluded (never allowlist-matched, so no model is ever consulted about them), 1 invalid-unwinnable (renders byte-identical to another case with the opposite expected answer — a prompt defect, see ACT-1212), 2 contested-expectation (the models made the better argument and the expectation was withdrawn). Excluded cases are retained in the task file as evidence, not deleted.",
+   "note": "gemini-2.5-flash n=6 not 10: four calls errored during the run, and the shortfall is recorded rather than papered over — the detector is noise-aware and should treat this prediction as correspondingly weak."
   }
  }
 }

package/src/data/builtin-manifest.json

@@ -1,8 +1,8 @@
 {
   "$schema": "./builtin-manifest.schema.json",
   "schemaVersion": 1,
-  "generatedAt": "2026-07-24T01:38:30.603Z",
-  "instarVersion": "1.3.927",
+  "generatedAt": "2026-07-24T04:02:28.049Z",
+  "instarVersion": "1.3.928",
   "entryCount": 202,
   "entries": {
     "hook:session-start": {

package/upgrades/1.3.928.md

@@ -0,0 +1,38 @@
+# Upgrade Guide — vNEXT
+
+<!-- assembled-by: assemble-next-md -->
+<!-- bump: patch -->
+
+## What Changed
+
+The benchmark predictions mirror now carries a `zombie-classify` entry, so the Benchmark-Divergence Detector has something to compare against for the second of its two enrolled decision points. It had been reporting `precondition-failed` for that pair since enrollment — correctly, since no baseline existed. Half the enrolled surface was measuring nothing.
+
+## What to Tell Your User
+
+Nothing to do. This is internal measurement plumbing — the detector is observe-only, dev-gated, dark on the fleet, and gates nothing.
+
+## Summary of New Capabilities
+
+- `GET /benchmark-divergence` can now produce a real verdict for the external-hog kill/leave classifier instead of a standing `precondition-failed`.
+
+## Evidence
+
+The baseline could not simply be re-run, because the existing case set was not testing the live decision.
+
+**Prompt drift, and a shape change.** The benched template was 2029 chars against a live prompt of 1373. The old cases fed the model `pid`, `parentPid`, `cpuPercent` and `elapsed`. The live prompt carries seven derived booleans plus the matched allowlist class, and deliberately withholds the identity tuple — the model does not need it, and omitting it denies an injection payload a concrete target to name in its logged reason.
+
+**Population mismatch.** Only an allowlist-matched candidate ever reaches a model: `identityFor` returns null otherwise and the scan tick surfaces the process without classifying it. Four of the eight original cases describe processes no model is ever consulted about.
+
+The boundary is the allowlist match **alone**, not the whole floor — an easy thing to get wrong, and it was gotten wrong once during this work. `evaluateKillFloor` runs separately, and a kill requires `floor.permitted && verdict === 'kill'`. A candidate the floor will veto is still a real question put to a real model, and its answer still worth grading. Reading floor-veto as "never asked" would have silently deleted three legitimate cases.
+
+**Two findings from the rebuilt set:**
+
+One case is unpassable by construction. `buildClassifierPrompt` renders every boolean with `=== true`, so an *unknown* owner state and a *known-false* owner state both print `owner_app_running: false`. Two cases carrying opposite expected answers render byte-for-byte identical prompts. That is a defect in the prompt, not the models — logged as ACT-1212. It is harmless today: the floor vetoes an unknown required field outright (`field-unknown:ownerAppRunning`), so the model's answer there is never enacted. Which is precisely why it would never surface from outcomes.
+
+All three models tested, across two families, recommend `kill` on a case whose fact block states plainly `sustained_high_cpu: false`. Also not enactable — the floor hard-vetoes on that exact fact. Also invisible from outcomes alone.
+
+**Scored population** is the 10 `production-candidate` cases. Excluded and retained in the task file as evidence rather than deleted: 4 `floor-excluded` (no model consulted), 1 `invalid-unwinnable` (above), and 2 `contested-expectation` — cases authored expecting `leave` for argv describing work in flight, where every model answered `kill` and had the better argument, since reaching a kill requires an orphaned owner and a language server indexing for a closed editor is burning cores on results nobody will collect. Those expectations were withdrawn rather than scored; encoding them would have published a confident and false claim about model behaviour.
+
+`gemini-2.5-flash` records `n=6` rather than 10 because four calls errored. Stamped as 6 rather than padded — the detector's noise-awareness only protects anything if the sample size it is given is true.
+
+**Not changed:** the classifier prompt itself. Fixing the unknown-vs-false flattening alters the prompt hash and invalidates this baseline, so it belongs in its own gated change with its own re-stamp, not bundled here.