session-orchestrator 5.2.0 → 5.3.0

package/skills/autopilot/SKILL.md

@@ -235,6 +235,7 @@ import { runLoop, parseFlags } from '$PLUGIN_ROOT/scripts/lib/autopilot.mjs';
 import { buildLiveSignals } from '$PLUGIN_ROOT/scripts/lib/build-live-signals.mjs';
 import { selectMode } from '$PLUGIN_ROOT/scripts/lib/mode-selector.mjs';
 import { probe, evaluate } from '$PLUGIN_ROOT/scripts/lib/resource-probe.mjs';
+import { detectPeers } from '$PLUGIN_ROOT/scripts/lib/session-registry.mjs';
 
 const flags = parseFlags(process.argv.slice(2));
 
@@ -250,14 +251,27 @@ const modeSelector = async () => {
   return selectMode(signals);
 };
 
+let cachedProbeSnapshot = null;   // written by peerCounter, read by resourceEvaluator
+
 const resourceEvaluator = () => {
-  const snapshot = probeSync();   // or cached snapshot if probe is async
-  return evaluate(snapshot, thresholds);
+  // Synchronous by contract: never calls probe() itself (it is async) — reads
+  // the snapshot peerCounter refreshed on the prior iteration.
+  if (cachedProbeSnapshot === null) {
+    return { verdict: 'warn', reasons: ['probe not yet available'], recommended_agents_per_wave_cap: null };
+  }
+  return evaluate(cachedProbeSnapshot, thresholds);
 };
 
-const peerCounter = () => {
-  // Synchronous-friendly count from a recent probe snapshot.
-  return latestSnapshot.claude_processes_count ?? 0;
+const peerCounter = async () => {
+  // A SESSION count, not a process count (host-resources.md HR-103) —
+  // claude_processes_count runs ~6 processes per session. `ownSessionId` is
+  // this coordinator's registry id, so detectPeers() excludes it.
+  const [peers, snapshot] = await Promise.all([
+    detectPeers({ sessionId: ownSessionId, freshnessMin: 15 }),
+    probe(),
+  ]);
+  cachedProbeSnapshot = snapshot;
+  return peers.length;
 };
 
 const sessionRunner = async ({ mode, autopilotRunId }) => {
@@ -289,8 +303,9 @@ The in-process driver has Claude (the coordinator) call `/session <mode>` betwee
 - **Pro:** zero new infra. Reuses canonical kill-switch logic. Validates `buildLiveSignals`
   against real Phase 7.5 swap before headless complexity. Each iteration carries
   inter-session memory through STATE.md / sessions.jsonl / learnings.
-- **Con:** not truly autonomous — Claude must stay in the chat. Doesn't deliver
-  walk-away UX. That's Phase C-5's job.
+- **Con:** not truly autonomous — Claude must stay in the chat. The in-process
+  driver does not deliver walk-away UX; that is the shipped headless driver
+  (`--headless`, `scripts/autopilot.mjs`, Phase C-5; see § Headless Driver Wiring).
 
 ### Headless Driver Wiring (Option A — `scripts/autopilot.mjs`)
 

package/skills/claude-md-drift-check/SKILL.md

@@ -65,7 +65,7 @@ Check 7 (issue #600) is the **only** check that intentionally reads BOTH instruc
 
 The same predicate guards Check 9's probe 2a, which iterates `['CLAUDE.md', 'AGENTS.md']`: with an aliased `AGENTS.md` it scans `CLAUDE.md` only, because otherwise every dangling `.claude/rules/*.md` citation would be reported TWICE — one defect, two errors, and in `--mode strict` a doubled blocking surface against a file the operator is not allowed to edit.
 
-Check 9 (`rule-scoping`) validates `.claude/rules/*.md` frontmatter against the `scripts/lib/rule-loader.mjs` contract, catching the class of defect where a rule silently drifts out of the activation pipeline the loader actually implements. Six probes: **(1) paths-presence** — a top-level `paths:` frontmatter key is not a key `rule-loader.mjs` recognises (it only reads `globs:`), so a rule with `paths:` silently loads ALWAYS-ON regardless of intended file scope; flagged as an error. **(2) cited-but-missing** — `(a)` `.claude/rules/<name>.md` citations inside `CLAUDE.md`/`AGENTS.md` that don't resolve to a file on disk, and `(b)` bare `<name>.md` tokens in a rule's own `## See Also` footer that don't exist as sibling rule files (tokens carrying a path separator, e.g. `../../skills/_shared/state-ownership.md`, are cross-directory references and explicitly out of scope); both flagged as errors. **(3) zero-match-globs** — a `globs:` pattern matching zero files in `git ls-files` (falls back to a manual directory walk when git is unavailable); flagged as a WARNING, not an error, because library/exemplar repos legitimately carry dead stack rules. **(4) foreign-glob** — a glob pattern containing a PascalCase product-like token (regex `[A-Z][a-z]+[A-Z]`, e.g. `WalkAITalkieTests`) — a likely copy-paste leftover from another project's rule scope; flagged as a WARNING. **(5) unreadable-file** — a rule file that could not be read (permissions, a race with a concurrent delete); flagged as a WARNING rather than silently skipped, so a completeness audit does not drop files, and so an unreadable file cannot brick the gate under `mode: hard`. **(6) fleet-intent-glob** — a `globs:` pattern DECLARED as fleet intent goes to `notes[]` (see below), never to `warnings[]`, and the declaration also settles probe 4 for that same pattern: a pattern the rule author declared is by construction not a copy-paste leftover. Glob matching reuses the same picomatch-with-inline-fallback resolution `scripts/lib/rule-loader.mjs` uses (`parseGlobsFrontmatter` is imported directly; the picomatch resolution itself is duplicated locally since `rule-loader.mjs` does not export a public matcher function). The check is skipped silently (no `checks_run` entry, no `checks_skipped` entry) when `.claude/rules/` is absent, or explicitly via `--skip-rule-scoping`.
+Check 9 (`rule-scoping`) validates `.claude/rules/*.md` frontmatter against the `scripts/lib/rule-loader.mjs` contract, catching the class of defect where a rule silently drifts out of the activation pipeline the loader actually implements. Six probes: **(1) paths-presence** — since #795 `paths:` is a full same-shape alias for `globs:` in `rule-loader.mjs` (`globs:` wins silently when both are present), so a well-formed `paths:`-only rule loads correctly SCOPED and is NOT flagged (#840). The probe fires only when the frontmatter textually carries a top-level `paths:` key AND `parseGlobsFrontmatter` — the loader's own parser, imported rather than re-derived — returns `globs === null` (e.g. a malformed frontmatter line makes it throw); that parse mismatch means the rule may silently load ALWAYS-ON, and is flagged as an error. **(2) cited-but-missing** — `(a)` `.claude/rules/<name>.md` citations inside `CLAUDE.md`/`AGENTS.md` that don't resolve to a file on disk, and `(b)` bare `<name>.md` tokens in a rule's own `## See Also` footer that don't exist as sibling rule files (tokens carrying a path separator, e.g. `../../skills/_shared/state-ownership.md`, are cross-directory references and explicitly out of scope); both flagged as errors. **(3) zero-match-globs** — a `globs:` pattern matching zero files in `git ls-files` (falls back to a manual directory walk when git is unavailable); flagged as a WARNING, not an error, because library/exemplar repos legitimately carry dead stack rules. **(4) foreign-glob** — a glob pattern containing a PascalCase product-like token (regex `[A-Z][a-z]+[A-Z]`, e.g. `WalkAITalkieTests`) — a likely copy-paste leftover from another project's rule scope; flagged as a WARNING. **(5) unreadable-file** — a rule file that could not be read (permissions, a race with a concurrent delete); flagged as a WARNING rather than silently skipped, so a completeness audit does not drop files, and so an unreadable file cannot brick the gate under `mode: hard`. **(6) fleet-intent-glob** — a `globs:` pattern DECLARED as fleet intent goes to `notes[]` (see below), never to `warnings[]`, and the declaration also settles probe 4 for that same pattern: a pattern the rule author declared is by construction not a copy-paste leftover. Glob matching reuses the same picomatch-with-inline-fallback resolution `scripts/lib/rule-loader.mjs` uses (`parseGlobsFrontmatter` is imported directly; the picomatch resolution itself is duplicated locally since `rule-loader.mjs` does not export a public matcher function). The check is skipped silently (no `checks_run` entry, no `checks_skipped` entry) when `.claude/rules/` is absent, or explicitly via `--skip-rule-scoping`.
 
 **Fleet-intent declaration (Check 9, probe 6).** A rule may declare that a `globs:` pattern is aimed at CONSUMER repos rather than this one, via its own frontmatter key `fleet-intent-globs:`. Accepted forms are block style (`fleet-intent-globs:` plus indented `  - "<pattern>"` lines), flow style (`fleet-intent-globs: ["a", "b"]`) and a single inline value; surrounding quotes are stripped, blank and `#` comment lines are ignored. The key is inert for rule loading — `rule-loader.mjs`'s `parseGlobsFrontmatter` skips unknown top-level keys and their continuation lines. A declared pattern is reported in `notes[]` with its file, the pattern, whether it matches `≥1` or `0` tracked files here, and the declaration `source` (`frontmatter`). The grain is per pattern AND per rule file: declaring one glob amnesties neither a sibling glob in the same rule nor the same pattern in another rule. There is deliberately NO checker-side exemption list: a rule that does not declare the pattern itself is not exempt, so the declaration always sits with the rule it describes.
 

package/skills/convergence-monitoring/README.md

@@ -19,7 +19,14 @@ Implemented from spec since upstream skill not yet published as of 2026-04-30.
 
 ## Invocation
 
-**Explicit:** `/convergence-monitoring` — standalone assessment from current session history.
+`user-invocable: false` — there is no `/convergence-monitoring` slash command.
+
+**Platform-native (primary, CC 2.1.105+):** registered as a plugin monitor
+(`monitors/monitors.json`); auto-starts in the background on `wave-executor`
+invocation and tails `.orchestrator/metrics/events.jsonl` for convergence signals.
+
+**Direct/model invocation:** the model dispatches this skill by name for a
+standalone assessment from current session history (fallback for harness < 2.1.105).
 
 **Embedded:** Set `convergence-monitoring: true` in Session Config. `wave-executor` and
 `/autoresearch` will call this skill automatically after each Impl-Core/Impl-Polish wave.

package/skills/convergence-monitoring/SIGNALS.md

@@ -208,12 +208,56 @@ VEL_LINES_PRIOR=<read from events.jsonl for wave N-1>
 
 The `velocity_drop` NDJSON signal counts **one `orchestrator.agent.stopped`
 record per dispatched agent** (wave read from `wave` or `wave_number`) and fires
-when the later wave stopped fewer agents than the earlier one. Measured
-2026-09-05 over `.orchestrator/metrics/events.jsonl`: 11,754
-`orchestrator.agent.stopped` records versus **0** for the `agent.dispatched` type
-the reader also accepts — `agent.stopped` is the only per-agent record this repo
-actually emits with a wave number. The git-based `VEL_COMMITS` / `VEL_LINES`
-recipe above is the SKILL-level computation and is independent of the monitor.
+when the later wave stopped fewer agents than the earlier one. `agent.stopped`
+is the only per-agent record this repo actually emits with a wave number — but
+the type is BIMODAL (#939/#949), so only records carrying a **non-empty `agent`
+field** are counted; the rest are the phantom-stop class and counting them
+inflated every per-wave count ~10.5x (#1379 P10).
+
+Measured 2026-09-18 over `.orchestrator/metrics/events.jsonl` plus the legacy
+`.jsonl.1` backup that then sat beside it — that backup was destroyed on
+2026-09-19, so the counts below are a record, not something this repo's ledger
+still reproduces.
+
+**Read the ledger through `readEventsWithRotations(repoRoot, opts)`**
+(`scripts/lib/events.mjs`), never by `cat`-ing rotation files together. It
+returns the active file plus every archive in time order, counts unreadable
+lines (`malformed_lines`), and — the reason it exists — reports a rotation
+whose archive is no longer on disk as an entry in `gaps` with `complete:
+false`, instead of silently returning a shorter history. Rotation itself no
+longer writes a `.1`..`.N` ring: since #1401 it renames the active file to
+`_archive/events-<firstTs>_<lastTs>.jsonl` and makes an
+`orchestrator.events.rotated` record naming that archive the first line of the
+new active file (`scripts/lib/events-rotation.mjs:230-258`); the reader still
+reads a legacy ring left on disk by older versions
+(`scripts/lib/events.mjs:512-539`).
+
+```bash
+node --input-type=module -e '
+import { readEventsWithRotations } from "./scripts/lib/events.mjs";
+const r = readEventsWithRotations(process.cwd());
+console.error(JSON.stringify({ gaps: r.gaps, malformed_lines: r.malformed_lines, complete: r.complete }));
+for (const e of r.events) process.stdout.write(JSON.stringify(e) + "\n");
+' | jq -s '{
+  total_stopped:      [.[]|select(.event=="orchestrator.agent.stopped")]|length,
+  with_wave:          [.[]|select(.event=="orchestrator.agent.stopped" and ((.wave//.wave_number)!=null))]|length,
+  with_wave_no_agent: [.[]|select(.event=="orchestrator.agent.stopped" and ((.wave//.wave_number)!=null) and ((.agent//"")==""))]|length,
+  with_wave_and_agent:[.[]|select(.event=="orchestrator.agent.stopped" and ((.wave//.wave_number)!=null) and ((.agent//"")!=""))]|length }'
+# 2026-09-18 (active + `.jsonl.1`):
+# → { total_stopped: 16438, with_wave: 6553, with_wave_no_agent: 5930, with_wave_and_agent: 623 }
+```
+
+i.e. 5930 of 6553 wave-scoped stops (90.5%) are phantoms, against **0** records
+of the `agent.dispatched` type the reader also accepts. The git-based
+`VEL_COMMITS` / `VEL_LINES` recipe above is the SKILL-level computation and is
+independent of the monitor.
+
+The `console.error` line is the honesty half of the reading and belongs in every
+re-measurement: a count taken from a ledger whose `gaps` is non-empty is a count
+over an unknown fraction of the history. Its blind spot is bounded and named —
+an archive deleted BEFORE #1401 left no tombstone and is undetectable by
+construction (`scripts/lib/events.mjs:556-561`), which is exactly how the
+53,896-line `.jsonl.1` vanished on 2026-09-19 without any analysis noticing.
 
 ### Thresholds
 

package/skills/convergence-monitoring/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: convergence-monitoring
-user-invocable: true
+user-invocable: false
 tags: [autoresearch, convergence, loop-control, wave-executor]
 model: haiku
 model-preference: sonnet
@@ -32,9 +32,17 @@ For harness < 2.1.105 (no monitor support), the skill's manual probes documented
 
 ## When to Invoke
 
-**Explicit invocation (`/convergence-monitoring`):** Standalone assessment — reads wave
-history from `.orchestrator/metrics/events.jsonl`, computes all three signals, reports
-verdict + recovery options.
+This skill is `user-invocable: false` — there is no `/convergence-monitoring` slash
+command. It is reached two ways:
+
+**Platform-native monitor (primary path, CC 2.1.105+):** the background watcher below
+auto-starts on `wave-executor` invocation and streams verdicts as it tails
+`events.jsonl` — no explicit call needed.
+
+**Direct/model invocation (fallback, harness < 2.1.105, or a standalone assessment):**
+the model dispatches this skill by name — reads wave history from
+`.orchestrator/metrics/events.jsonl`, computes all three signals, reports verdict +
+recovery options.
 
 **Embedded (loop context):** `/autoresearch` and `wave-executor` invoke this skill after
 each Impl-Core or Impl-Polish wave when `convergence-monitoring: true` is set in Session
@@ -48,9 +56,10 @@ toward done or oscillating/stalling.
 
 ## Phase 0: Activation Gate
 
-### 0.1 Explicit Invocation
+### 0.1 Direct Invocation
 
-If invoked via `/convergence-monitoring`, skip to Phase 1 unconditionally.
+If invoked directly (not embedded from `wave-executor` or `/autoresearch`), skip to
+Phase 1 unconditionally.
 
 ### 0.2 Embedded Invocation
 

package/skills/eval/SKILL.md

@@ -15,7 +15,7 @@ args-schema:
   - flag: --verify
     description: "Re-evaluate a stored run-id and diff per-dimension for scoring drift (exit 1 on drift)"
 description: >
-  Use this skill to run an honest session-process evaluation (Standard v1, aiat-llm-eval/1.0) — score the last completed orchestrator session against the pre-registered rubric-v1 dimensions, run /eval, evaluate this session, produce an eval report, or re-verify a stored eval run for reproducibility. Deterministic-first with an optional advisory LLM judge; never produces a global score.
+  Use this skill to run an honest session-process evaluation (Standard v1, aiat-llm-eval/1.0) — score the last completed orchestrator session against the pre-registered rubric-v2 dimensions, run /eval, evaluate this session, produce an eval report, or re-verify a stored eval run for reproducibility. Deterministic-first with an optional advisory LLM judge; never produces a global score.
 ---
 
 > **Platform Note:** State files use the platform's native directory: `.claude/` (Claude Code), `.codex/` (Codex CLI), or `.cursor/` (Cursor IDE). Shared metrics + the eval journal live in `.orchestrator/metrics/`. See `skills/_shared/platform-tools.md`.
@@ -23,14 +23,17 @@ description: >
 # Eval Skill — Session-Process Evaluation (aiat-llm-eval/1.0)
 
 On-demand, honest measurement of ONE completed orchestrator session against the
-pre-registered **rubric-v1** check set. The deterministic engine
+pre-registered **rubric-v2** check set. The deterministic engine
 (`scripts/eval-session.mjs` → `scripts/lib/eval/engine.mjs`) reads only local
-metrics files (`sessions.jsonl` + `events.jsonl`), scores the five deterministic
+metrics files (`sessions.jsonl` + `events.jsonl`), scores the six deterministic
 dimensions, appends a `session-eval` record to the journal, and optionally
-renders an HTML report. An opt-in LLM judge overlays two advisory dimensions.
+renders an HTML report. An opt-in LLM judge overlays ONE advisory dimension
+(`instruction-adherence`; `report-quality` was retired in rubric-v2, #1381).
 
 The standard this skill implements is [`docs/eval/aiat-llm-eval-v1.md`](../../docs/eval/aiat-llm-eval-v1.md);
-the frozen, content-hashed check set is [`skills/eval/rubric-v1.md`](./rubric-v1.md).
+the frozen, content-hashed check set is [`skills/eval/rubric-v2.md`](./rubric-v2.md)
+(stored records written before 2026-09-19 carry `rubric-v1` and are read against
+[`rubric-v1.md`](./rubric-v1.md), which is never edited again).
 
 ## Invocation
 
@@ -43,7 +46,7 @@ Invoked as `/eval [--session <id>] [--no-write] [--verify <run-id>]` with argume
 
 **On-demand `/eval` runs regardless of `eval.enabled`** — that flag gates only the automatic session-end eval phase (see Phase 1.1).
 
-**Seams used:** `scripts/eval-session.mjs` (deterministic CLI) · `runEvalJudge` / `mergeJudgeDimensions` (`scripts/lib/eval/judge.mjs`, opt-in) · `writeEvalReport` (`scripts/lib/eval/report.mjs`) · `appendEvalRecord` (`scripts/lib/eval/sink.mjs`) · the `eval` config block · [`skills/eval/rubric-v1.md`](./rubric-v1.md) (frozen check set).
+**Seams used:** `scripts/eval-session.mjs` (deterministic CLI) · `runEvalJudge` / `mergeJudgeDimensions` (`scripts/lib/eval/judge.mjs`, opt-in) · `writeEvalReport` (`scripts/lib/eval/report.mjs`) · `appendEvalRecord` (`scripts/lib/eval/sink.mjs`) · the `eval` config block · [`skills/eval/rubric-v2.md`](./rubric-v2.md) (frozen check set).
 
 ## Posture Contract (load-bearing — read before executing)
 
@@ -52,9 +55,9 @@ Invoked as `/eval [--session <id>] [--no-write] [--verify <run-id>]` with argume
 - **Never guess.** Missing source data yields `cannot-determine` (a first-class,
   non-error verdict) with an honest reason — never a fabricated `pass`/`fail`.
   Do NOT "fill in" a missing KPI or infer a gate result the events do not show.
-- **Deterministic before judge.** The five deterministic dimensions are complete
-  on their own. The judge (Phase 3) is opt-in, ADVISORY, and `uncalibrated` in
-  v1 — never blend a judge verdict into the deterministic tally.
+- **Deterministic before judge.** The six deterministic dimensions are complete
+  on their own. The judge (Phase 3) is opt-in, ADVISORY, and `uncalibrated` —
+  never blend a judge verdict into the deterministic tally.
 - **Journal is SSOT; the report is a derived view.** The append-only
   `.orchestrator/metrics/eval.jsonl` is authoritative. The HTML report is
   rebuildable from any stored record and is never authoritative over the journal.
@@ -148,10 +151,13 @@ node scripts/eval-session.mjs [--session <id>] --json \
   On exit `1` (e.g. "no completed session found"), surface the message and stop —
   do not retry with fabricated inputs.
 
-Parse the emitted JSON record. It carries `dimensions[]` (5 deterministic
-entries), `kpis{}`, `provenance.rubric_sha256` (non-null once `rubric-v1.md`
-exists), `model`, `harness`, and `run_id`. Unless `--no-write` was passed, the
-record is already appended to `.orchestrator/metrics/eval.jsonl` by the CLI.
+Parse the emitted JSON record. It carries `dimensions[]` (6 deterministic
+entries — the two reported-only ones, `guard-friction` and `efficiency-kpis`,
+are always `not-applicable`), `kpis{}`, `provenance.rubric_sha256` (the sha256
+of `rubric-v2.md`; `null` means the rubric file was not found and the append
+will fail validation), `model`, `harness`, and `run_id`. Unless `--no-write` was
+passed, the record is already appended to `.orchestrator/metrics/eval.jsonl` by
+the CLI.
 
 **Contamination check:** if the human-render/summary reports a peer-overlapped
 window, note it — `verification-evidence` and `gate-health` will read
@@ -187,9 +193,13 @@ const merged = mergeJudgeDimensions(record, dimensions);
 appendEvalRecord(merged, { path: '.orchestrator/metrics/eval.jsonl' });
 ```
 
-- Every judge dimension arrives `advisory: true` + `calibration_status:
-  "uncalibrated"` (the schema firewall rejects any other shape). Keep them
-  visibly separated from the deterministic five in the summary.
+- The judge dimension arrives `advisory: true` + `calibration_status:
+  "uncalibrated"` (the schema firewall rejects any other shape). Keep it
+  visibly separated from the deterministic six in the summary.
+- `runEvalJudge` pre-computes a `facts` block (`computeRecordFacts`) from the
+  deterministic evidence strings and hands it to the judge OUTSIDE the
+  untrusted-data fence — anything countable is counted in code, never inferred
+  by the model. Do not reimplement that here.
 - If `runEvalJudge` returns a non-ok `status` (e.g. dispatch failed), keep the
   deterministic record as-is and note the judge was unavailable — the
   deterministic evaluation is complete without it.
@@ -226,20 +236,20 @@ const res = writeEvalReport(record, { generatedAt: new Date().toISOString() });
 Emit a compact, honest per-dimension summary. Status lines only — no global score.
 
 ```
-## /eval — <session_id>  (self-evaluation, aiat-llm-eval/1.0 · rubric-v1 · n=1, no CI)
+## /eval — <session_id>  (self-evaluation, aiat-llm-eval/1.0 · rubric-v2 · n=1, no CI)
 
 Deterministic:
   verification-evidence   PASS   <one-line evidence>
   plan-fidelity           PASS   completion_rate=1.0 (score)
   gate-health             PASS   <one-line evidence>
-  process-safety          PASS   <one-line evidence + guard-emission disclosure>
+  process-safety          PASS   <one-line evidence + both guard disclosures>
+  guard-friction          N/A    (reported: blocked=… warned=… loop.warning=… <attribution>)
   efficiency-kpis         N/A    (reported: duration=…s waves=… agents=… tok_in=… tok_out=… carryover=…)
 
 Judge (advisory, uncalibrated)  [only when eval.judge != off]:
-  instruction-adherence   <verdict>   advisory
-  report-quality          <verdict>   advisory
+  instruction-adherence   <verdict>   advisory   (rule <n> applied)
 
-cannot-determine: <k> of 5 deterministic dimensions (<reasons>)
+cannot-determine: <k> of 6 deterministic dimensions (<reasons>)
 Report:  .orchestrator/eval/reports/<run_id>.html
 Journal: .orchestrator/metrics/eval.jsonl  (appended: <yes|--no-write>)
 Re-verify: node scripts/eval-session.mjs --verify <run_id>
@@ -266,6 +276,11 @@ node scripts/eval-session.mjs --verify <run-id> --json
   changed since the record was written — investigate, do not overwrite.
 - `--verify` reproduces the stored model + timestamp verbatim (no env override),
   so a MATCH is a real reproducibility proof of the scoring, not of model output.
+- **A cross-version DRIFT is not a defect.** A stored `rubric-v1` record
+  re-scored by today's rubric-v2 engine necessarily differs on `process-safety`
+  and reports `present-in-fresh-only: guard-friction`. Read the record's
+  `rubric_version` before treating a diff as a regression (#1400 replaces that
+  report with an explicit version verdict).
 
 ---
 
@@ -282,9 +297,9 @@ platform. Only the judge phase needs harness-specific tooling.
   `skills/_shared/platform-tools.md` § Agent Dispatch Pattern.
 - **`harness.platform`** on the record is resolved from `$SO_PLATFORM`
   (falls back to `claude-code`) inside the engine — no skill action needed.
-- The deterministic five dimensions + the HTML report + `--verify` are fully
-  available on all platforms; the judge overlay is a Claude-Code-only enrichment
-  in v1.
+- The deterministic six dimensions + the HTML report + `--verify` are fully
+  available on all platforms; the judge overlay is a Claude-Code-only
+  enrichment.
 
 ---
 

package/skills/eval/rubric-v1.md

@@ -5,6 +5,7 @@
 - **Conforms to standard:** `aiat-llm-eval/1.0` — see [`docs/eval/aiat-llm-eval-v1.md`](../../docs/eval/aiat-llm-eval-v1.md)
 - **Reference engine:** [`scripts/lib/eval/engine.mjs`](../../scripts/lib/eval/engine.mjs) (the executable scorers this document mirrors verbatim)
 - **Hash binding:** the sha256 of THIS FILE is written to every record's `provenance.rubric_sha256`.
+- **Superseded for NEW records by [`rubric-v2.md`](./rubric-v2.md) (2026-09-19, issue #1037).** This file stays frozen and authoritative for every record already carrying `rubric_version: "rubric-v1"` — its formulas are the ones those verdicts were produced under and must not be edited. v2 changed exactly two things: `process-safety` no longer fails on `destructive_guard.blocked`, and the guard counts moved to a new reported-only `guard-friction` dimension. The reasoning and the measurement live in rubric-v2 § Änderungen gegenüber v1.
 
 > **Pre-Registration (leading principle, standard §1.1).** The checks below are
 > **fixed BEFORE the first scored run executes against this rubric**. This document