auto-decorrelation — скилл для Claude Code: проверка изменений

---
name: auto-decorrelation
description: Recruits cross-family verifier sidecars (codex, agy, local 4090 over Tailscale) for adversarial verification of load-bearing changes, maximizing model-family diversity against the orchestrator. Mechanically discovers the available sidecar panel, recruits at least one cross-family verifier when present, and degrades gracefully when none are. The governor (Claude) keeps the terminal verdict; sidecar findings must be source-grounded before acceptance. Opt-in via one-time consent, stored in the UAP; fires only on load-bearing changes. Triggered by "recruit a cross-family check", "decorrelate this verification", "use the idle sidecars to verify", "auto-decorrelation".
user-invocable: true
allowed-tools: ["Read", "Bash", "Grep", "Glob"]
model-note: session-inherit — Sonnet base is first-class (sonnet_floor_doctrine.md); depth-critical judged steps route to dispatch (opus agent / cross-family sidecar, consent-gated), never a substrate requirement
---

# auto-decorrelation — Cross-Family Verifier Recruitment

The lever for catching a model's blind spots is **decorrelation, not repetition** — a model cannot find
its own family's flaws by repeating itself. FH's adversarial verification (Axis-2 challenger) dispatches
in-session at the session tier, **same family** as the governor. When frontier CLIs are installed and a
local GPU box is live, leaving them idle wastes the one thing that actually raises the verification
ceiling. This skill recruits a **cross-family** verifier for load-bearing changes — and degrades
honestly when none is available.

It is a thin recruitment-and-degrade **policy** that REUSES existing machinery: `agent-composer` does the
actual dispatch, `cross-ecosystem-synergy-detection` inventories tools, `token-budget-gate` gates spend,
the headless CLI patterns and the floor-tier canary already exist. Net-new = the discovery + the
family-diversity recruitment + the degrade ladder only. (A **composing policy skill** — it fans out to
several callees, so its token floor is higher than a leaf skill by design; this is composition, not
reinvention, and the chain is return-path-closed — Step 5 gates acceptance on the callee's back-trace.)

## Triggers

- "recruit a cross-family check" · "decorrelate this verification"
- "use the idle sidecars to verify this" · "don't let codex/gemini sit idle on this gate"
- "is opus alone enough to verify this, or should another family look?"
- "auto-decorrelation"

## Step 0 — Consent gate (default OFF)

External-CLI calls and credit spend are never silent. Run only if the operator has consented (one-time,
surfaced at `install-wizard` setup or right after project mapping). Consent is stored in the UAP
(`tracks/_meta/user_adaptation_profile.md` — behavioral pref, not domain content). No UAP / no consent
(incl. ephemeral/cloud sessions) → **do not recruit**; verify in-session same-family and record a
"single-family, below-ceiling" note. Consent wording is honest by requirement: *"멀티-CLI 사이드카로
단일 모델을 넘어서는 효과를 기대할 수 있습니다 — 100%는 아님(decorrelation은 향상이지 보장 아님).
가용 사이드카 없으면 자동으로 단일 세션으로 축소됩니다."*

## Step 1 — Load-bearing predicate (consent gates eligibility; this gates firing)

Fire only on **load-bearing** changes — the same predicate as the target-tier sim gate: gate infra ·
onboarding scaffolds · destructive/publish paths · routing/gate skills · **or any change whose output
another skill/gate consumes as trusted input** (a new synthesis/analysis skill, a corpus writer) — the
predicate tracks *downstream trust*, not only whether the artifact is itself a gate. Ordinary additive
edits stay in-session (cost guard). Consent ON ≠ always-fire: a small/low-stakes check silently stays
single-session even when sidecars are available.

## Step 2 — Sidecar discovery (mechanical, the anti-power-waste core)

Build the available panel at run time; absent tools **and unreachable endpoints** drop off silently:
```bash
# Sidecar-callable CLIs only. **Scope discipline**: the probe list is NOT "every agent CLI that exists".
# FH's main is Claude Code (vendor-native — `[[feedback_vendor_native_harness]]`), and the recommended
# cross-family sidecars are **codex and gemini/agy**. Other agent CLIs on the machine (opencode, qwen,
# hermes, cursor-agent …) are **runtimes a user works INSIDE**, not verifiers FH calls out to — being
# installed is not a reason to probe them. Adding one costs a maintenance surface and dilutes the panel;
# add only when a concrete task needs that family. (2026-07-19: four were added off a general CLI
# catalog and removed the same session — installed ≠ belongs in the panel.)
command -v codex >/dev/null && echo "codex"          # GPT family CLI
command -v agy   >/dev/null && echo "agy"            # serves Gemini AND GPT-OSS — probe the model
command -v gemini>/dev/null && echo "gemini"
command -v gh >/dev/null && gh copilot --help >/dev/null 2>&1 && echo "copilot"
                                                     # ↑ gh EXTENSION, not a binary — `command -v copilot` misses it, and
                                                     #   `gh copilot --help` shows only the LAUNCHER (its sole flag is
                                                     #   --remove); the real flags live behind `--`.
                                                     # Call form (verified 2026-07-19 by live call — credits were consumed):
                                                     #   gh copilot -- -p '<prompt>' --model <model> --allow-all-tools
                                                     # **Same class as codex/agy**: `--model` selects among several families
                                                     #   behind ONE CLI ('auto' lets Copilot pick), so family MUST come from
                                                     #   the pinned/probed model (Step 3), NEVER from the CLI name.
                                                     # ★ SEAT TIER CHANGES ITS VALUE ENTIRELY — probe, never assume:
                                                     #   · free seat        → narrow model choice; treat as ONE extra family
                                                     #   · enterprise seat  → serves GPT, Gemini AND Claude behind the single
                                                     #     CLI: a THREE-FAMILY panel with no other CLI installed
                                                     #     ([[reference_corp_env_decorrelation_panel]]).
                                                     #     ⚠️ BUT each family runs on Copilot's harness, not its vendor-native
                                                     #     one — Claude-via-copilot ≠ Claude Code, GPT-via-copilot ≠ codex,
                                                     #     Gemini-via-copilot ≠ agy. Per `[[feedback_vendor_native_harness]]`
                                                     #     a non-native harness costs depth. So copilot buys **breadth cheaply,
                                                     #     not depth**: use it to widen the panel, and route the decisive
                                                     #     check to the vendor-native CLI when one is reachable. Same shape as
                                                     #     the local canary tier (breadth ≠ terminal depth, measured 2026-07-19).
                                                     #   Because the panel it yields depends on the seat, Step 3's model probe
                                                     #   is not optional here: enumerate what this seat actually serves before
                                                     #   claiming family diversity.
                                                     # Cost shape: paid-seat credits. That makes it a strong *sidecar* but a
                                                     #   poor main driver — seat quota is spent faster than it is worth when
                                                     #   it drives the whole harness. Recruit it for decisive checks, not bulk.
                                                     # Residual: the launcher may fetch the CLI body on first call, so on a
                                                     #   cold machine the first recruit pays a download.
# Local ollama serving-paths = canary tier (electricity-only). mac localhost is public → probed
# UNCONDITIONALLY. Any extra path (e.g. a Tailscale GPU box) is an operator-private token → read from a
# gitignored binding, NEVER hardcoded in this public file. Both mac-serving (H2) and 4090-serving (평시)
# are covered: whichever box is not serving simply fails the probe and drops off.
# probe() validates the /api/tags SCHEMA, not just a reachable port: -f rejects HTTP 4xx/5xx and the
# `"models"` grep rejects a non-ollama server or an empty/overloaded instance — else a dead box reports
# live (false-positive discovery). Endpoints are only ever curl-probed here, never eval'd.
probe() { curl -fsS -m"${2:-6}" "http://$1/api/tags" 2>/dev/null | grep -q '"models"'; }
probe localhost:11434 && echo "ollama-local(mac)"
EP="$FH_SIDECAR_EXTRA"
[ -z "$EP" ] && [ -f tracks/_meta/sidecar_endpoints.env ] && \
  EP="$(grep '^OLLAMA_EXTRA=' tracks/_meta/sidecar_endpoints.env | cut -d= -f2- | tr -d '"')"
# Split on NEWLINES via `while read`, never on an unquoted `for e in $EP`. Word-splitting an unquoted
# variable is a BASH behavior; zsh does NOT word-split, so under zsh a space-separated multi-endpoint
# list arrives as ONE string, hits the charclass below (space is not in it) and is dropped — silently
# reporting "no extra sidecars", i.e. the single-family fail-open this skill exists to prevent.
# Measured 2026-08-11: `EP="host1:11434 host2:11434"` → bash probes 2, zsh probes 0.
# Accepts either separator: commas/spaces are normalized to newlines first.
printf '%s\n' "$EP" | tr ' ,' '\n\n' | while IFS= read -r e; do
  [ -z "$e" ] && continue
  case "$e" in *[!0-9a-zA-Z.:-]*)
    # NEVER a silent `continue` — a dropped endpoint must be visible, or an unprobed panel is
    # indistinguishable from an unavailable one (DEGRADED_PANEL_UNUSED vs DEGRADED_SINGLE_FAMILY).
    echo "sidecar-endpoint DROPPED (not host:port form): [$e]" >&2; continue;;
  esac
  probe "$e" 10 && echo "ollama-extra($e)"             # -m10: a sleeping GPU box may wake slower than 6s
done
```
**Shell note**: this block is `sh`-portable and is written so bash and zsh behave identically. Any
future edit that reintroduces `for e in $EP` re-opens the zsh drop above. A drop line on stderr is a
**finding, not noise** — carry it into the Step 6 degrade verdict rather than discarding stderr.
Endpoint resolution is a **mechanical env/file read** (not a prose instruction the runner must remember),
so this discovery is tier-independent — no target-tier sim owed. The extra-endpoint binding lives only in
the gitignored `tracks/_meta/sidecar_endpoints.env` (auto-synced to the companion store); the public skill
carries the probe logic, never the address.

## Step 3 — Family map by runtime model probe (NOT CLI name)

Tag each verifier by the model family **actually serving the request**, probed — not the CLI binary name.
`agy` serves both Gemini and GPT-OSS 120B, so "agy = Gemini" is wrong; resolve via the CLI's
`--model`/served-model. The binary→family table is only a fallback default. Recruit to **maximize family
diversity vs the orchestrator** (orchestrator = Claude/opus → recruit GPT or Gemini or local-Qwen).

## Step 4 — Recruit (cost-aware) + spend gate

- **Cost order**: local 4090 = electricity-only (cheapest — use for breadth pre-screen) · codex headless
  = hard-capped credit pool · agy = operator Gemini sub. So: local-Qwen for cheap breadth, then ONE
  frontier cross-family for the decisive check.
- **Spend gate (S-1)**: the **free local tier auto-fires** silently; any **paid** draw (codex/agy)
  surfaces a one-line `token-budget-gate` ask per run (*"recruiting codex (~N) — proceed?"*) unless the
  operator has set `paid_auto: true` in the UAP. One-time feature-consent ≠ consent to this spend now.
- Dispatch via `agent-composer` (no re-implementation of dispatch).
- **Wait mechanically — `scripts/sidecar_wait.sh` is the required form, not a suggestion (S-1b).**

  **The rule, before the command, because a Sonnet-tier blind sim of this section said the command
  reads as "the content" and everything after it as "color commentary" — and named peeking at the
  output file as the first thing it would do wrong in a hurry:**

  > **Never judge a sidecar by looking at its output file.** A live process and a dead one produce
  > the same zero bytes. The only readable verdict is the typed `SIDECAR_VERDICT=` line, and it
  > does not exist until the process has exited.

  ```bash
  printf '%s' "$prompt" | bash scripts/sidecar_wait.sh out.txt 900 -- codex exec -m gpt-5.5 -
  # → SIDECAR_VERDICT=COMPLETE exit=0 bytes=48489      (read out.txt)
  # → SIDECAR_VERDICT=TIMEOUT  waited=900s bytes=0     (still alive — NOT a result)
  # → SIDECAR_VERDICT=EMPTY    exit=0                  (the only state that means "it said nothing")
  ```

  The runner **refuses to emit a verdict while the process is alive**, so "the sidecar returned
  nothing" becomes unsayable until it has actually exited. Grep the typed `SIDECAR_VERDICT=` line;
  never judge by looking at the output file.

  **Why this is mechanical rather than a habit** — the bullets below already described bounding a
  sidecar, and a session that had them loaded still got it wrong on 2026-07-29: it backgrounded
  `codex exec` and `agy -p`, read the output files after **1 s and 30 s**, found them empty, and
  recorded *"both sidecars returned 0-output"* into a gate marker, a PR body, a session card, a
  memory file and a handoff. Both were running normally and both answered — codex with 48 KB and
  three findings, agy with a further HIGH, and **all four were real**; one of them showed the change
  under review was over-applied. So the measurement error nearly retired a working mechanism.

  **That is a second failure mode this section did not cover.** The bullets below describe a *hung*
  sidecar. An impatient read of a *healthy* one produces the identical observation — zero bytes —
  and only one of the two is a fault. Distinguishing them requires process state, which is exactly
  what a human eye on an output file cannot see and the runner always reports.

- **Liveness / hang-catch (the runner does the waiting; this is how to read a `TIMEOUT`).** A backgrounded CLI that hangs
  (stuck on a sandbox/file-tool prompt, auth, or network) **does not exit**, so the background-completion
  signal *never fires* — passive waiting is the wrong model and silently stalls the run (observed
  2026-06-27: a `codex exec` that asked to read repo files hung at 0-output with no session log, and the
  turn waited on a notification that could not come). So **bound it actively, never wait open-endedly**:
  - The timeout is the runner's second argument; pick it from the model's real latency, not from
    impatience (a reasoning model can be silent for minutes and still be working).
  - On `TIMEOUT`, watch a **progress signal**, not just process-alive: output bytes growing **and** the CLI's own
    session/log advancing (e.g. `~/.codex/sessions/<today>`). 0 output **and** no session created after a
    short bound (≈2–3 min for codex/agy) = **hung, not slow** → kill and recover, do not keep waiting.
  - **Recover, don't stall**: kill → diagnose (a file-tool/sandbox hang is the common cause) → retry with
    the **robust pattern** — inline the needed file content into the prompt via **stdin** so the sidecar
    needs no file tools (verified fix 2026-06-27: `{ instructions; cat file1; cat file2; } | codex exec -`
    succeeded where reading-files-itself hung) — or degrade to single-session with a recorded note.
  This is the sidecar twin of `mcp-circuit-breaker` (stop a stuck external call instead of hanging on it).

## Step 4.5 — Residency screen (mandatory, immediately before ANY external-family send)

**Every** payload this skill sends outward — the diff, the marker (Step 5's payload), the prompt
text itself — crosses a family boundary the moment a codex/agy/gemini sidecar is recruited.
"Sanitize before external-family dispatch" used to be a prose reminder with no consumer reading it
back (`[[feedback_half_externalization_slot_without_consumer]]` shape — a slot that exists but
nothing downstream checks). This step makes it mechanical:

```bash
python3 scripts/residency_closure_scan.py --files src/a.py docs/b.md   # every file you are about to send, as separate args (zsh: ${=var}, never an unquoted "$var")
```

`--files` (not the default closure-walk) because what matters is **what you are actually
sending**, not this skill's own dependency graph. Read the exit code, not the words:

| Exit | Meaning | What to do |
|---|---|---|
| 0 | CLEAN — no organizational identifier found **within the scanner's own stated limits** (see its header — this is a screening result, not a safety guarantee) | Send. Record `residency=CLEAN(files=N)` in the crossfamily grounds (Step 6). |
| 1 | TAINTED — at least one payload file matched a residency pattern | Do not send as-is. Either strip/redact the tainted file(s) from the payload and re-scan, or route to an in-boundary panel member. If it truly cannot be cleaned, record `residency=TAINTED(N files, stripped=no)` and do NOT dispatch — the crossfamily value becomes a `DEGRADED_*` rung, never `panel(...)`. |
| 10 | HARNESS_ERROR — the scan itself could not run (commonly: no operator-private `.claude/rules/.residency-patterns` override on this machine — the shipped `.defaults` file alone is deliberately insufficient to claim CLEAN, per the scanner's own header) | Do not send. Record `residency=NOT_SCANNED(<reason>)` — same as exit 1's outcome (no dispatch), different reason (*could not screen* vs *screened and found something*). |

🟥 **On a fresh install with no operator pattern file, this WILL be exit 10 every time** — that is
the scanner's own documented fail-closed default (`residency_closure_scan.py`'s header: "오버라이드
파일이 없으면 exit 10 이다"), not a bug introduced here. It means: until an operator configures
their own residency patterns, this skill cannot claim a panel dispatch was screened, so it honestly
degrades to a `DEGRADED_*` rung with `residency=NOT_SCANNED(...)` instead of silently sending. That
is not a new block on fresh installs — the crossfamily field already tolerates `DEGRADED_*` without
hard-failing the commit; it is one more honest reason to be in that state, not a novel one.

**Where the result lands**: `templates/.git-hooks/pre-commit`'s `validate_crossfamily_leg` now
requires a `residency=CLEAN(...)` token inside the grounds of any `panel(...)` value (grace date
`RESIDENCY_TOKEN_GRACE_DATE`, no retroactivity on markers dated earlier), and accepts an optional,
format-checked `residency=(CLEAN|TAINTED|NOT_SCANNED)(...)` token on `DEGRADED_*`/`UNKNOWN`/
`declined`. Fixtures: `scripts/test_marker_crossfamily_lanes.sh` (`r1`–`r13`). This is the SAME
typed field Step 6 already emits into — one channel, not a second marker line.

## Step 4.6 — `evidence=` : say WHAT the panel received (mandatory with `panel(...)`, 2026-09-12)

`residency=` says the payload was safe to send. `evidence=` says **what was sent** — and that is the
axis the outside world just measured as the dominant one. `arXiv:2609.10969` (48 task templates →
2,880 scenarios, **fixed call budget** so the comparison is about the axis and not about spending
more): a cross-model vote over **shared** evidence approved **62.9 %** of unsafe proposals against
**22.9 %** with an **independent source** — source effect **40.9 pp** vs **11.3 pp** for model
diversity, i.e. **3.6×**.

Decide it from **how this run actually dispatched**, not from intent:

| What happened | Token |
|---|---|
| Every member got the **same** payload (one prompt file / one diff fanned out) | `evidence=SHARED(<name the artifact>)` |
| Each member produced its **own** evidence — own checkout, own run, own extraction | `evidence=INDEPENDENT(<what each got>)` |
| Some of each | `evidence=MIXED(<which member got which>)` |

🟥 **This skill's default flow is `SHARED`, by construction.** Step 4 builds **one** payload and fans
it out, so a panel dispatched the ordinary way through this skill is the **62.9 % arm** — the axis it
raises is model family, which the paper prices at 11.3 pp. That is not a reason to stop using it: a
finder is worth its recall, and this skill's value was never precision
(`[[feedback_decorrelation_is_a_cheap_finder]]`). It **is** a reason to stop recording that run as if
it carried the stronger axis. Write `SHARED` and mean it.

🟢 **`SHARED` is a legal, expected answer.** The hook does not block it; it blocks a `panel(...)` that
does not say which it was, a value outside the closed three, a malformed or duplicated token, and a
**vacuous body** on `SHARED`/`MIXED` (name the artifact — *"same staged diff"*, *"one findings.jsonl"*,
*"identical prompt file"*). `INDEPENDENT` is self-describing and is not body-checked, deliberately:
over-blocking the honest strong answer would train the override.

**Where it lands**: the same grounds line as `residency=`, in the same `crossfamily:` field —
`crossfamily: panel(codex) — residency=CLEAN(files=3) · evidence=SHARED(same staged diff to both) · R1, 3 findings`.
Hook: `validate_crossfamily_leg` (grace `EVIDENCE_TOKEN_GRACE_DATE`, no retroactivity).
Fixtures: `scripts/test_marker_crossfamily_lanes.sh` `e1`–`e10`.

⚠️ **Named gap, not built today**: this skill has **no independent-source mode**. Giving it one
(each member reconstructs the evidence itself rather than reading the author's payload) is the change
that would move it onto the 40.9 pp axis; it is a design question, not a wiring one, and it is
recorded as a signal rather than improvised here.

🟥 **A stripped file is invisible to the reviewer, and a reviewer's default read of invisible is
"absent," not "redacted"** (governor dogfood, 2026-09-05, same-day live use of this Step): two
TAINTED files were stripped from a real payload and the CLEAN remainder sent cross-family — the
reviewer came back with a false-positive S-finding, *"runner wiring is missing,"* because the
wiring actually lived in one of the two files that had just been stripped. Stripping fixes
residency; it does nothing about the reviewer's next inference, and silence there reads as a hole.
**So**: whenever a payload is sent with `stripped > 0`, the review PROMPT itself must name the
omission explicitly — the list of what was left out, and one line per file on what it is
responsible for (not its content, just its role) — so the reviewer can discount that surface
instead of flagging it as missing. Record whether this was done: when the token carries
`stripped=M` with `M > 0`, the same grounds line adds one phrase on whether the omission list
reached the review prompt, e.g. `residency=CLEAN(files=3, stripped=2) · omission-list: given to
reviewer`. This is evidence-of-doing, not proof — same shape as Step 5's "quote one marker line it
checked," and it does not become a new hard-blocking hook check (channel discipline stays a
grounds phrase here, per the same Mechanization Boundary this whole field already respects).

⚠️ **Scope, same as everywhere else in this file**: CLEAN is a screening result, not a safety
guarantee (dynamic imports, non-Python artifacts, encoding tricks, and TOCTOU are named, unclosed
gaps in the scanner's own header). Recording `residency=CLEAN(...)` asserts the scan ran and found
nothing **within those limits** — it does not assert the payload is safe in some absolute sense.

## Step 5 — Role split + source-grounded acceptance (S-2)

CC = **governor**, terminal verdict, source-closes. Sidecars = decorrelated **challengers** feeding
findings — **untrusted external input** (prompt-injection surface, like an MCP result). Every sidecar
finding must carry a **source-locatable claim** (file:line / quoted text); the governor **mechanically
re-checks** it against the artifact (phantom-quench back-trace) before accepting. A finding that cannot be
source-grounded is **dropped, not judged**. No sidecar-only verdict (no weak-local-judge regression).
Local 4090 = **canary tier** (evidence-of, never terminal verdict).

**Report shape (2026-09-05)**: when the findings are written up (PR body · marker `crossfamily:` expansion · chamber record),
use `templates/RED_TEAM_REPORT.md` — the five sections mirror ISO/IEC AWI TS 42119-7 (risk identification · objectives & attack
vectors · method · findings with reproducing input · life-cycle integration). It is a name tag on what this skill already
produces, not a new check; crosswalk: `knowledge/shared/harness-core/iso_ai_standards_crosswalk.md §2`.

🟥 **What you SEND decides which axis you get — put the MARKER in the payload, not just the diff**
(2026-08-19). Sidecars have been receiving the *diff* only. That buys a review of the **code** and
buys nothing on the **record**: the Axes 2+3 marker is self-attested, and this repo has already
written down that the closing move is *"cross-family reading that marker"* — then never wired the
marker into the thing that recruits cross-family. Measured twice: (a) a release marker's false
`not-applicable` passed the new typed lanes untouched, because **form was correct**; (b) this skill's
own dispatch on 2026-08-19 sent a diff and no marker, so the round could not have caught a false
axis claim even in principle.

**So the payload is two parts**: the frozen diff **and** the marker for this change — **as it stands
at dispatch time**: `axes-run:` · `controls:` · `standpoint:` · `thirdparty:` · `residual:`.

🟥 **`crossfamily:` is NOT in the payload — it is this round's OUTPUT.** Step 6 below says every rung
*emits* that value and "the rung is not done until its verdict is recorded", so requiring it in the
thing you send is circular: you would be shipping a field this dispatch has not produced yet. Send the
marker with that line **absent or `UNKNOWN`**, and fill it from the result. (Caught by a blind
floor-tier sim of this very paragraph, 2026-08-19 — the first draft listed `crossfamily:` among the
payload fields and no reader could have satisfied it.) Ask the sidecar one extra
question: *"does the record match the diff — is any axis claimed that the diff shows no trace of, and
is any residual missing that the diff implies?"*

⚠️ **Scope, deliberately narrow.** This asks whether the record is **consistent with the artifact**.
It does NOT ask the sidecar to score honesty — «is this marker truthful» is a *conclusion*, and
§Mechanization Boundary forbids freezing that into machinery. Consistency is checkable from two
documents; honesty is not.
⚠️ **You cannot observe that the sidecar READ it — say so rather than implying otherwise.** Putting the
marker in the payload and the sidecar actually using it are different events, and nothing here
distinguishes them: a silently-ignored marker looks exactly like a marker that was read and raised
nothing. The cheap partial anchor is to require the return to **quote one marker line it checked** —
a reply that quotes nothing did not demonstrably read it. That is evidence-of-reading, not proof, and
it does not close §Mechanization Boundary's named self-attestation residual. Three independent
readers flagged this same gap on the day the paragraph was written (a peer session, and two blind
floor-tier sims), which is why it is stated here instead of left to the reader to notice.

⚠️ **Prose, not a check** — measured recurrence is 2, below this repo's own N≥3 bar
(`[[feedback_mechanize_at_repetition_prose_before]]`). On the third occurrence, mechanize it here.
⚠️ **Residency still governs**: a marker can name company assets. Run Step 4.5's screen against the
marker file too, not just the diff — the marker is not exempt because it is metadata. (This line was
pure prose until 2026-09-05; Step 4.5 above is what closes it — the scan itself still carries the
named, unclosed gaps in its own header, so "screened" is not "guaranteed safe".)

*Origin*: sister-asset read of `raphaelchristi/harness-evolver`'s `harness-critic` agent, whose whole
role is auditing the **evaluator** rather than the artifact. Its detection signatures (score jumps,
suspiciously fast convergence) do **not** port — FH markers carry no score — but the *target* does.
Full assessment incl. what was deliberately not imported: `tracks/_audit/proposal_2026-08-19_sister_harness-evolver.md`.

**Target freeze — a prior drop reason, checked before any of the above (2026-08-17).** Grounding a
finding against *today's* tree proves nothing if the sidecar reviewed *yesterday's*. Pin before
dispatch and verify on return:

```bash
bash scripts/target_freeze.sh pin    "$REPO" "$ROUND_LABEL"   # 발주 직전
bash scripts/target_freeze.sh verify "$REPO" "$ROUND_LABEL"   # 회수 직후
```

`rc 1` = **WRONG-TARGET → drop the whole round's findings**, not individual ones — the round read a
tree that moved under it, so which findings survive is not decidable from the output. `rc 10` =
UNKNOWN (could not measure) — that is **not** a pass; re-pin and re-dispatch. Measured trigger: three
occurrences in one day, one of which was sending a sidecar a **pre-repair diff** and then grounding
its findings against the repaired tree — every finding "failed to ground", and the instrument, not
the reviewer, was wrong.

🟥 **What this does NOT cover**: §L191's inlined-file pattern. When the prompt carries file contents
on stdin (the robust form for sidecars that have no file tools), the tree freeze says nothing about
**what bytes you actually sent**. That binding is still owed and is not claimed here.

## Step 6 — Degrade ladder (the intelligent scale-down)

**Consent br...