dsh-aris-panel 0.1.0

package/skills/kill-argument/SKILL.md

@@ -0,0 +1,437 @@
+---
+name: kill-argument
+description: "Two-thread adversarial review: a fresh reviewer constructs the strongest 200-word rejection memo, then a second fresh reviewer defends the paper point-by-point and surfaces still-unresolved critical issues. Use when user says \"kill argument\", \"adversarial review\", \"hostile review\", \"rebuttal preparation\", \"reviewer-2 simulation\", or before submitting a theory paper that has already passed standard review rounds."
+argument-hint: "[paper-directory]"
+allowed-tools: Bash(*), Read, Write, Edit, Grep, Glob, mcp__codex__codex
+---
+
+# Kill Argument Exercise: Adversarial Attack-Defense Review
+
+> 🔒 **Do not wrap this skill in `/loop`, `/schedule`, or `CronCreate`.** It is
+> verdict-bearing — it produces an adversarial accept/reject verdict (attack →
+> adjudication). Re-firing it on a wall-clock timer adds no new signal (the
+> attack changes only when the *paper* changes). Schedule the *external wait
+> that precedes it* — draft stable → then run this **once** before submission.
+> See
+> [`shared-references/external-cadence.md`](../shared-references/external-cadence.md).
+
+Stress-test the headline claims of a paper against the strongest possible rejection argument: **$ARGUMENTS**
+
+## Why This Exists
+
+Standard score-based reviews (`/research-review`, `/auto-paper-improvement-loop`) tend to produce **balanced** weakness lists.  Each weakness gets ~equal attention, ranked CRITICAL > MAJOR > MINOR.  Empirically, this misses one specific failure mode: the **single most damaging argument** a reviewer would write in a rejection paragraph — the one sentence that, if a senior area chair reads it, kills the paper.
+
+A balanced reviewer might list "scope-overclaim risk" as MAJOR alongside 3-5 other MAJORs, never quite committing.  An adversarial reviewer **must commit**: their entire job is to convince the area chair to reject in 200 words.
+
+This skill runs that adversarial pass deliberately, then forces a second fresh reviewer to defend point-by-point, classify each rejection as already-fixed / partially-fixed / still-unresolved, and surface what's actually load-bearing.
+
+**Empirical motivation:** in a real submission run, after several rounds of standard improvement (score 7-8/10), the kill-argument exercise surfaced framing weaknesses that no prior review caught (e.g., a setting being mostly conditional rather than truly general, or a baseline being irrelevant to real systems).  Author rebuttal forced explicit scope qualifications in abstract and discussion that weren't visible from the score-based reviews alone.
+
+## How This Differs From Other Review Skills
+
+| Skill | What it asks the reviewer | Output |
+|-------|---------------------------|--------|
+| Standard peer review | "Score this paper, list weaknesses by severity" | balanced weakness list |
+| `/research-review` | "Deep technical review of methods + claims" | structured deep critique |
+| `/proof-checker` | "Is this theorem actually proved?" | per-step proof obligation audit |
+| `/paper-claim-audit` | "Does the paper report numbers truthfully?" | per-claim evidence verification |
+| `/citation-audit` | "Are citations real and used in correct context?" | per-entry KEEP/FIX/REPLACE/REMOVE |
+| **`/kill-argument`** | **"Write the single strongest rejection paragraph; then defend it."** | **attack memo + per-point defense + unresolved surfaced** |
+
+This skill is **complementary**, not a replacement.  Run after standard reviews when you want to know what the worst-case reviewer paragraph would look like, before camera-ready or rebuttal preparation.
+
+## When To Use
+
+- After 1-2 rounds of `/auto-paper-improvement-loop` settled at a stable score, but before submission.  Surfaces what additional fixes would close the headline-attack gap.
+- During rebuttal preparation, to predict reviewer-2's strongest objection so you can prepare the response in advance.
+- For theory papers with a high-level title that may oversimplify the actual theorem (the most common reject-attack pattern).
+- For papers where a reviewer might attack scope, assumption-vs-claim mismatch, missing proof obligations, or evidence-vs-headline gaps.
+
+This skill is most valuable for **theory papers** with ≥5 theorem-class environments (so the headline depends on real proof obligations).  For empirical papers without theorems, use `/research-review` instead.
+
+## Constants
+
+- **REVIEWER_MODEL** = `gpt-5.6-sol` (default; `gpt-5.5` is the capability fallback, `gpt-5.4` only as an explicit legacy override).  Reviewer reasoning effort = `ultra` for the attack / defense / adjudication threads (deep-audit tier; capability fallback per `shared-references/reviewer-routing.md`, never below `xhigh`).  Beast-mode axis probes stay at `xhigh`.
+- **CONTEXT_POLICY** = `fresh` (REVIEWER_BIAS_GUARD).  Each thread is a fresh `mcp__codex__codex` call.  **Never** use `mcp__codex__codex-reply`.  No prior review summary, fix list, or executor explanation enters either prompt.
+- **ATTACK_LENGTH** = approximately 200 words (do not exceed 250).  Single coherent argument, not a list.
+- **DEFENSE_DECOMPOSITION** = 3-7 atomic rejection points extracted from the attack memo.  Each gets its own classification.
+- **CLASSIFICATION** = `answered_by_current_text` / `partially_answered` / `still_unresolved`.  (Names chosen so the adjudicator does not assume "fixed" implies prior history of patching — they read the paper as a fresh reviewer would.)
+- **OUTPUT** = `KILL_ARGUMENT.md` (human-readable) + `KILL_ARGUMENT.json` (machine-readable) in the paper directory.
+- **RENDER_HTML = true** — When `true` (default), auto-render `KILL_ARGUMENT.md` to HTML after writing the report. Uses **full Codex review gate** (audit-class artifact — full render-fidelity check matches the skill's cross-model audit invariant; the sidecar `KILL_ARGUMENT.json` is also passed to the renderer). Set `false` to skip, or pass `— render html: false`.
+
+## Workflow
+
+### Step 1: Discover paper files
+
+Locate the paper directory and inventory the source.
+
+```bash
+PAPER_DIR="$ARGUMENTS"   # e.g., paper-overleaf/ or paper/
+cd "$PAPER_DIR"
+
+# Find the LaTeX entry point
+ENTRY=$(grep -lE '^\\documentclass' *.tex 2>/dev/null | head -1)
+echo "Entry: $ENTRY"
+
+# Find all source files codex should read
+find . -name "*.tex" -not -path "./.git/*" 2>/dev/null
+find . -name "*.bib" -not -path "./.git/*" 2>/dev/null
+find figures/ -name "*.pdf" -o -name "*.png" 2>/dev/null
+ls -la *.pdf 2>/dev/null  # compiled PDF
+```
+
+If a compiled PDF is missing, the skill should still run on .tex source alone, but the prompt should mention this so the reviewer doesn't waste cycles trying to extract from a non-existent PDF.
+
+### Step 2: Attack memo (Thread 1, fresh codex)
+
+Invoke `mcp__codex__codex` (NOT `codex-reply`) with the following prompt structure:
+
+```
+mcp__codex__codex:
+  model: gpt-5.6-sol
+  config: {"model_reasoning_effort": "ultra"}
+  sandbox: read-only
+  cwd: <paper directory>
+  prompt: |
+    You are simulating a hostile NeurIPS / ICLR / ICML reviewer for a paper.
+    This is a kill-argument adversarial check — your task is NOT to give a
+    balanced review but to construct the **single strongest argument for
+    rejecting this paper**.
+
+    ## Files to read
+    - LaTeX entry: <ENTRY>
+    - All section files under sections/ or wherever they live
+    - Macro files (math_commands.tex, etc.)
+    - Compiled PDF: <main.pdf> (if available)
+
+    Read the source carefully. Do not consult any prior reviews, fix lists,
+    or summaries; this must be a fresh, zero-context adversarial pass.
+
+    ## Your task
+    Construct the single best argument to reject this paper in approximately
+    200 words. Your goal is to write the worst-case rejection memo a senior
+    NeurIPS area chair would produce after reading the paper.
+
+    Focus on these axes (pick the most damaging combination, do not list all):
+    1. Theorem validity: are central theorems actually proved as stated?
+    2. Assumption-vs-claim mismatch: does the body silently retreat to a
+       narrower object than the title/abstract advertise?
+    3. Missing proof obligations: is a fundamental lemma invoked but not
+       proved (e.g., concentration, generic position, prefactor envelope)
+       that the headline depends on?
+    4. Limit-order ambiguity: are limits in K/n/d/eps composed in a way the
+       paper does not commit to?
+    5. Claim-vs-evidence gap: is the empirical/numerical evidence too narrow
+       to support the breadth of the stated theorem or take-away?
+    6. Scope overclaim: does the title or abstract sell a result substantially
+       broader than what the body proves?
+
+    ## Constraints
+    - Approximately 200 words total (do NOT exceed 250).
+    - Single argument, not a list — pick the most damaging line of attack
+      and develop it.
+    - Cite specific file:line locations or equation numbers when accusing.
+    - Tone: dispassionate but uncompromising. Do NOT hedge. Do NOT acknowledge
+      mitigations the paper might have made elsewhere. This is the rejection
+      paragraph; the defense gets the next pass.
+    - Do NOT reference prior review rounds, fix lists, or any context outside
+      the current paper files.
+
+    Output: just the rejection memo, nothing else.
+```
+
+Save the returned `threadId` for the trace; do NOT pass it to Thread 2.  Save the attack memo verbatim — both Thread 2 and the human-readable report use it.
+
+### Step 2.5 (optional, `beast` effort): multi-axis attack fan-out
+
+**Default OFF.** The deliverable of this skill *is* a verdict — the single
+strongest rejection paragraph — and
+[`shared-references/fan-out-pattern.md`](../shared-references/fan-out-pattern.md)
+is explicit: **do not fan out the verdict; fan out only the evidence that
+feeds it.** The default single-commitment attack (Step 2) is deliberate —
+forcing one paragraph produces sharper feedback than a balanced list (see *Why
+This Exists*). Do **not** replace it with a list.
+
+Under `beast` effort you may widen the *evidence* the commitment draws on
+without diluting the commitment:
+
+1. **Axis probes (evidence breadth).** Run the six attack axes (theorem
+   validity / assumption-vs-claim / missing obligation / limit-order /
+   claim-vs-evidence / scope-overclaim) as **separate fresh-codex probes**,
+   each asked for the strongest ~120-word thrust *on that axis alone*. These
+   are evidence-gathering, not the verdict. Probes run at `xhigh` (not
+   `ultra`) — six serial delegating calls would multiply cost for evidence
+   that the ultra-tier commit re-judges anyway.
+   - **These are NOT Claude subagents, and there is deliberately NO `Agent`
+     grant.** Each probe is a fresh `mcp__codex__codex` call — the adversary
+     must be cross-model (non-Claude). Codex MCP is **serial** (concurrent
+     codex calls hang), so the probes run **sequentially** — Tier-3 in the
+     fan-out ladder. This is exactly why `kill-argument` lists no `Agent` in
+     `allowed-tools`: it spawns nothing; it threads codex calls.
+2. **Commit (the verdict, still single).** A final fresh-codex synthesis reads
+   the six probes plus the paper and must **commit to the single most damaging
+   ~200-word rejection paragraph** — selecting and fusing at most two axes, NOT
+   listing all six. The Step-2 commitment requirement is unchanged; the probes
+   only ensure no axis was overlooked before committing.
+
+The adjudication (Step 3) then runs against this committed attack exactly as in
+the default flow. Cost: `beast` adds ~6 extra serial codex calls — use it for
+the final pre-submission pass on a high-stakes paper, not routinely.
+
+Tracing: record each probe's `threadId` (`axis_probe_thread_ids[]`) and the
+synthesis `threadId` in the trace, the same way Steps 2–3 save their thread
+ids. The committed attack memo, not the six probes, is what Step 3 consumes.
+
+### Step 3: Adjudication memo (Thread 2, fresh codex with attack + paper)
+
+Invoke a second `mcp__codex__codex` call (still NOT `codex-reply` — Thread 2 is independent of Thread 1's codex history):
+
+```
+mcp__codex__codex:
+  model: gpt-5.6-sol
+  config: {"model_reasoning_effort": "ultra"}
+  sandbox: read-only
+  cwd: <paper directory>
+  prompt: |
+    You are an independent area-chair adjudicator examining whether the
+    current paper text answers a hostile reviewer's rejection memo.
+    You are NOT the paper's defender — your job is to read the attack
+    point-by-point and rule, from the current source files alone,
+    whether each point stands or falls. Fresh, zero-context adjudication;
+    do not reference any prior reviews / fix lists.
+
+    ## Paper files
+    [list paths same as Step 2]
+
+    ## The hostile reviewer's rejection memo (the "attack")
+    > <attack memo verbatim from Thread 1>
+
+    ## Your task
+    The attack is one continuous argument, but it makes multiple distinct
+    rejection points that you must adjudicate separately. Decompose the
+    attack into its atomic rejection points (3-7 of them), then for each
+    point classify it:
+
+    - answered_by_current_text: the current paper source already mitigates
+      this point (cite specific file:line evidence)
+    - partially_answered: paper has some response but not enough to refute
+      the attack as written
+    - still_unresolved: paper has no effective response
+
+    The label `answered_by_current_text` is intentional — "fixed" implies
+    history of patching and biases toward optimism. You are reading the
+    paper as a reviewer would, with no knowledge of prior round drafts.
+
+    For each rejection point, output:
+    ### Point P_n: <short label>
+    **Attack claim**: <the specific accusation, ~30 words>
+    **Verdict**: answered_by_current_text | partially_answered | still_unresolved
+    **Evidence (or lack of)**: <cite file:line, ~50 words>
+    **Severity if unresolved**: critical | major | minor
+    **If unresolved, recommended fix**: <one specific actionable sentence>
+
+    After per-point analysis, output:
+
+    ## Summary
+    Total rejection points: N
+    - answered_by_current_text: X
+    - partially_answered: Y
+    - still_unresolved: Z
+
+    ## Net assessment
+    <one short paragraph: would this paper survive a senior area-chair read
+    of the attack memo, given only what is in the current source? Be honest —
+    if Y or Z > 0 and they hit the headline, say so.>
+
+    ## Top action items (in priority order, max 3)
+    1. ...
+    2. ...
+    3. ...
+
+    ## Constraints
+    - Do NOT consult any prior round reviews or fix lists. Adjudication must
+      be made strictly from current paper files.
+    - If the paper cannot refute a point, do NOT minimize — keep severity
+      honest.
+    - If a point reflects an author-chosen position (e.g., conscious title
+      scope decision), classify as `partially_answered` with a note that the
+      position is intentional, AND say whether this position is sustainable
+      under the attack — do NOT auto-grade as `answered_by_current_text`
+      just because it is intentional.
+    - Be specific. No flattery, no hedging, no rationalizing on the paper's
+      behalf.
+```
+
+Save the returned `threadId`.
+
+### Step 4: Write KILL_ARGUMENT.md and KILL_ARGUMENT.json
+
+Compose the human-readable report `<paper-dir>/KILL_ARGUMENT.md`:
+
+```markdown
+# Kill Argument Report — <paper title>
+
+**Date**: <YYYY-MM-DD>
+**Reviewer model**: <resolved pair that actually ran — target gpt-5.6-sol ultra>, fresh threads (no codex-reply)
+**Attack thread**: <threadId 1>
+**Adjudicator thread**: <threadId 2>
+**Verdict**: <PASS / WARN / FAIL / NOT_APPLICABLE / BLOCKED / ERROR> (`reason_code: <...>`)
+
+## Net assessment
+
+<paragraph from adjudicator memo's "Net assessment">
+
+## Attack memo (verbatim)
+
+> <attack memo from Thread 1>
+
+## Adjudication (per-point)
+
+<copy verbatim from Thread 2 — uses labels answered_by_current_text / partially_answered / still_unresolved>
+
+## Top action items
+
+<copy from Thread 2>
+
+## Recommendation
+
+If P_4 (or whatever still_unresolved critical) is research-level, record
+it as a known open problem in the conclusion / limitations. If it is
+writing-level, queue for next /auto-paper-improvement-loop round.
+```
+
+Compose the machine-readable `<paper-dir>/KILL_ARGUMENT.json` per the
+ARIS Audit Artifact Schema (`shared-references/assurance-contract.md`):
+
+```json
+{
+  "audit_skill": "kill-argument",
+  "verdict": "PASS | WARN | FAIL | NOT_APPLICABLE | BLOCKED | ERROR",
+  "reason_code": "<see verdict mapping below>",
+  "summary": "<one-line summary, ~80 chars>",
+  "audited_input_hashes": {
+    "main.tex":                          "sha256:<...>",
+    "sec/0.abstract.tex":                "sha256:<...>",
+    "sec/<each-section>.tex":            "sha256:<...>",
+    "references.bib":                    "sha256:<...>",
+    "main.pdf":                          "sha256:<...>"
+  },
+  "trace_path": ".aris/traces/kill-argument/<date>_run<NN>/",
+  "thread_id": "<defense threadId — primary; attack threadId in details>",
+  "reviewer_model": "<resolved — the model that actually ran (target: gpt-5.6-sol)>",
+  "reviewer_reasoning": "<resolved — the effort that actually ran (target: ultra)>",
+  "generated_at": "<UTC ISO-8601>",
+  "details": {
+    "attack_thread_id": "<threadId 1>",
+    "defense_thread_id": "<threadId 2 — same as top-level thread_id>",
+    "attack_memo": "<verbatim>",
+    "decomposed_points": [
+      {
+        "id": "P_1",
+        "label": "<short label>",
+        "attack_claim": "<...>",
+        "verdict": "answered_by_current_text | partially_answered | still_unresolved",
+        "evidence": "<file:line citation>",
+        "severity_if_unresolved": "critical | major | minor",
+        "recommended_fix": "<...>"
+      }
+    ],
+    "counts": {
+      "answered_by_current_text": <int>,
+      "partially_answered":       <int>,
+      "still_unresolved":         <int>
+    },
+    "net_assessment": "<adjudicator memo's net assessment>",
+    "top_action_items": ["...", "...", "..."]
+  }
+}
+```
+
+**Hash inputs** (`audited_input_hashes`): use paper-relative paths,
+`sha256` of every `.tex` consumed plus `references.bib` and the
+compiled `main.pdf` if it exists. The verifier rehashes these on
+`verify_paper_audits.sh` and flags `STALE` if the user edited the
+paper after running the audit.
+
+**Verdict mapping** (every (counts, severity) tuple must hit exactly one row):
+
+| Verdict | reason_code | Trigger |
+|---|---|---|
+| `FAIL` | `unresolved_critical` | ≥1 `still_unresolved` at `critical` severity |
+| `WARN` | `unresolved_major_or_minor` | ≥1 `still_unresolved` at `major` or `minor` severity (and no `critical`) |
+| `WARN` | `partial_critical_or_repeated_major` | 0 `still_unresolved`, AND ≥1 `partially_answered` at `critical` or `major` |
+| `PASS` | `defense_survives_with_minor_partial_only` | 0 `still_unresolved`, AND ≥1 `partially_answered`, all at `minor` severity |
+| `PASS` | `defense_survives` | 0 `still_unresolved`, AND 0 `partially_answered` |
+| `NOT_APPLICABLE` | `not_theory_or_scope_paper` | Paper has <2 `\begin{theorem\|lemma\|proposition\|corollary}` AND no scope / generality claims in abstract |
+| `NOT_APPLICABLE` | `headline_unstable` | Title or abstract changed within the last 2 commits — re-run after headline stabilizes |
+| `BLOCKED` | `paper_compile_failed` | Compiled PDF missing AND `main.tex` does not compile clean — adjudication needs source fidelity |
+| `BLOCKED` | `source_files_missing` | `main.tex` not found, or no `sec/*.tex` files |
+| `ERROR` | `codex_api_error` | `mcp__codex__codex` call failed |
+| `ERROR` | `decomposition_parse_failed` | Adjudicator thread did not return parseable per-point structure |
+| `ERROR` | `trace_save_failed` | Trace directory write failed |
+
+`PASS` requires `still_unresolved == 0`. With `still_unresolved == 0`, any
+`partially_answered` at `major` or higher makes the best available verdict
+`WARN` — never `PASS`.
+
+The verdict is computed from the per-point counts; do NOT let the
+defense thread output the top-level verdict directly (that would let
+it self-grade). The skill code does the verdict mapping.
+
+### Step 5: Print summary
+
+To the user:
+
+```
+🗡  Kill Argument complete.
+
+  Attack: <one-sentence summary of the rejection thrust>
+
+  Adjudication breakdown:
+    answered_by_current_text:   X
+    partially_answered:         Y
+    still_unresolved:           Z   ← critical: <names>
+
+  Verdict: <PASS / WARN / FAIL / NOT_APPLICABLE / BLOCKED / ERROR>
+  Reason:  <reason_code, e.g., defense_survives, unresolved_critical>
+
+  Top action items:
+  1. ...
+  2. ...
+  3. ...
+
+  Full report: <paper-dir>/KILL_ARGUMENT.md
+```
+
+## Output Contract
+
+- `<paper-dir>/KILL_ARGUMENT.md` — human-readable report
+- `<paper-dir>/KILL_ARGUMENT.json` — machine-readable ledger
+- `.aris/traces/kill-argument/<date>_runNN/` — per-thread codex traces (Attack memo + Adjudication memo)
+- Optional: applied fixes if user explicitly requests; default is **detect-only, do not auto-modify**.
+- `<paper-dir>/KILL_ARGUMENT.html` (when `RENDER_HTML = true`, default) — single-file HTML view auto-rendered via `/render-html "<paper-dir>/KILL_ARGUMENT.md" --json "<paper-dir>/KILL_ARGUMENT.json"`. Full review gate applies. The `.review.json` sidecar carries the render-fidelity verdict. **Non-blocking**: if `/render-html` fails (helper missing, Codex MCP unavailable, file write error), log the failure and treat the skill as complete — the HTML view is a convenience, not a prerequisite for the kill-argument verdict.
+
+## Key Rules
+
+- **Fresh thread per call.**  Both Attack and Adjudication use `mcp__codex__codex`, never `codex-reply`.  Thread 1 and Thread 2 must not share codex context.
+- **Zero prior context.**  Neither thread receives prior round reviews, fix lists, executor summaries, or improvement-loop logs.
+- **Attack must commit.**  Single argument, ~200 words.  No "consider also" hedge.  The whole value is in forcing the reviewer to pick the most damaging line.
+- **Adjudicator must classify, not minimize.**  `still_unresolved` is honest if the paper has no effective response.  Don't downgrade to `partially_answered` unless evidence is real.
+- **Author-chosen positions** (e.g., deliberate title scope, deliberate omission of qualifier): mark `partially_answered` with note that the position is intentional, AND say whether the position is sustainable under the attack.  Don't auto-grade as `answered_by_current_text` just because it's intentional.
+- **Verdict is computed by the skill, not by the adjudicator.**  The Codex thread emits per-point classifications; the skill code maps those to one of the 6 audit verdicts via the table in Step 4.  Never let the adjudicator self-grade the top-level verdict.
+- **Detect-only by direct invocation; can be invoked by `/auto-paper-improvement-loop` Step 5.5 which then merges unresolved findings into its fix list.**  When a user runs `/kill-argument paper/` directly, the output is informational and the human decides whether to act.  When the skill is invoked from inside the auto-improvement loop, the loop reads `KILL_ARGUMENT.json`, deduplicates against its existing weakness list, and feeds novel `still_unresolved` points into Step 6 fixes — `/kill-argument` itself never edits paper files.
+
+## When NOT to Use
+
+- Empirical papers without theorems / scope claims — `/research-review` is more useful.  The skill emits `NOT_APPLICABLE` with `reason_code: not_theory_or_scope_paper` in this case.
+- Very early drafts where the headline isn't stable yet — fix the headline first.  The skill emits `NOT_APPLICABLE` with `reason_code: headline_unstable` if the title or abstract changed within the last 2 commits.
+- Papers with ongoing experiments — wait until results stabilize, then run.
+- (`/auto-paper-improvement-loop` Step 5.5 used to run this protocol inline; as of May 2026 it now invokes `/kill-argument` and reads `KILL_ARGUMENT.json` instead, so there is no longer a "do not invoke from inside auto-loop" exclusion.)
+
+## Review Tracing
+
+After each `mcp__codex__codex` reviewer call, save the trace following `shared-references/review-tracing.md` (Policy C — forensic; never silently skip).  Use `save_trace.sh` (resolved per the chain in `shared-references/integration-contract.md` §2) or write files directly to `.aris/traces/kill-argument/<date>_run<NN>/`.  Both threads' raw responses should be preserved.
+
+## Notes
+
+This skill was extracted as a standalone primitive from `/auto-paper-improvement-loop` Step 5.5 in May 2026, after the protocol proved valuable in surfacing headline-vs-body scope gaps that score-based reviews missed.  The attack-then-defense pattern was kept exactly because of empirical evidence that asking one model to "write the rejection memo" produces qualitatively different feedback than asking it to "review and grade" — the former forces commitment, the latter encourages hedging.