adversarial-review-research

name	adversarial-review-research
description	Research-mode review procedure for adversarial-reviewer. Fires when the target is a `.md`/`.txt` file, URL, or raw claim text. Invoked by the agent's turn-1 dispatch.

Adversarial Review — Research Mode

This is a procedure file, not a standalone skill. It is loaded by the adversarial-reviewer agent after turn-1 dispatch resolves the target as research mode.

Vocabulary authority: Evidence labels ([FETCHED], [CODEBASE], [TRAINING], [REASONING], [ABSENT]), tunable constants (TIER_1_CONFIDENCE_FLOOR, TIER_2_CONFIDENCE_FLOOR, TIER_1_HARD_CAP, TRAINING_LABEL_CONFIDENCE_CAP, FETCHED_QUOTE_MAX_CHARS, WITHDRAWAL_CONFIDENCE_FLOOR), and the 4-tier verdict scale are defined in ~/.pi/agent/agents/adversarial-reviewer.md. Do not redefine any of these here.

---

Pipeline Execution Order

S1 → S2 → [S3 opt-in] → S6 → S4 → S5 → S7 → S8

Execute all steps before writing the output artifact. Do not interleave composition with execution — complete all steps, then assemble D6 sections.

---

S1 — Restate the Claim

Purpose: Confirm the target before attacking it. Produces a crisp restatement in the reviewer's own words that locks in what is being reviewed.

How:

1. Read the full target (file, URL content, or free-form text).
2. Write 2–4 sentences restating the central claim(s) as you understand them.
3. Note the scope: what the work claims to establish (not what it does establish).

Output: Internal working note only — not surfaced as a numbered section. Feeds Bottom Line (S8 → Section 2). If the restatement reveals the claim is ambiguous or unspecified, surface that ambiguity as an objection in S7.

---

S2 — Key Assumptions Check

Purpose: Enumerate every load-bearing implicit assumption the work makes and rate each by confidence × impact.

How:

1. List every assumption the central claim depends on — include hidden premises.
2. For each assumption:
   • Confidence rating (0–100%) — how confident are you the assumption is actually true?
   • Impact if false (High / Medium / Low) — would the claim collapse, weaken, or survive?
3. High-impact, low-confidence assumptions are likely to surface as Tier-1 objections in S7 — flag them for promotion.

Output format: Table for Section 5.

| # | Assumption | Confidence | Impact if False | Notes |
|---|-----------|-----------|-----------------|-------|
| A1 | ... | 40% | High | flagged for S7 |

Output composition: → Section 5 (Key Assumptions Surfaced)

---

S3 — Steelman (Opt-In)

Only execute if --steelman flag was passed.

Purpose: Articulate the strongest possible version of the claim before attacking it. Prevents attacking a straw man.

How:

1. Restate the claim in its most defensible form — fill obvious gaps, assume best-faith readings.
2. Identify the 2–3 strongest pieces of evidence or arguments in its favor.
3. Note what would need to be true for the claim to fully survive scrutiny.

Position: Execute after S2, before S6. Doubles output size — only enable when the claim is vaguely stated or the author is not present to defend.

Output composition: → Not a numbered D6 section. Used internally to calibrate S7 objection thresholds. Optionally append as a labeled sub-section under Section 2 (Bottom Line) with header ### Steelman (requested).

---

S6 — Citation Shepardize

Purpose: Verify that every cited source actually says what the author claims it says.

How: Invoke the adversarial-shepardize skill. Do not reimplement the procedure here — load and follow ~/.pi/agent/skills/adversarial-shepardize/SKILL.md.

The shepardize skill produces a completed Section 4 table (Citation Shepardizing). Fold that table directly into the artifact.

Cross-Section Escalation (CRITICAL)

When the shepardize subroutine returns any row with verdict MISREPRESENTED or CHERRY-PICKED, this skill — not the subroutine — owns the promotion to Section 3 (Objections):

1. For each MISREPRESENTED / CHERRY-PICKED row (citation ID C<N>):

   • Create a top-level objection O<N> in Section 3.
   • Tier: always Tier-1. Shepardize findings bypass all intensity thresholds (per adversarial-reviewer.md D8 — "Citation shepardizing findings (S6): bypass thresholds — any misrepresented citation is auto-Tier-1 regardless of confidence").
   • Objection block:

     **O<N> — Misrepresented citation: <source short name>** [FETCHED]
     Tier: 1 (auto — shepardize finding)
     Confidence: [use confidence from the shepardize row, or 80% if not stated]
     Consequence if true: The claim's evidentiary basis is weaker than presented; readers following the citation will not find support for the stated conclusion.
     See C<N> in Citation Shepardizing table (Section 4).
     

   • Do NOT duplicate the full table row. The cross-reference is sufficient.

2. These O<N> IDs are reserved first. Regular S7 objections are numbered after them.

Output composition:

• Section 4 (Citation Shepardizing table) ← table produced by adversarial-shepardize
• Section 3 (Objections) ← MISREPRESENTED / CHERRY-PICKED rows promoted as auto-Tier-1 O<N>

---

S4 — Pre-Mortem

Purpose: Generate a list of plausible failure modes by assuming the claim has already been disproved. Unfiltered — inclusion criterion is "plausible enough to be worth defending against."

How:

1. Start with: "Assume this claim has been conclusively disproved. List every reason why."
2. Generate an unfiltered list — do not apply confidence thresholds yet.
3. For each failure mode, ask: "Can I label this with evidence?" If yes, it's a candidate for S7. If it's speculation only, it may still be worth surfacing if plausible.

S4 in Research Mode — Resolution

The D6 artifact has a Pre-mortem Failure Modes section (Section 6) only in change mode. In research mode, Section 6 does not appear. This is intentional: the research artifact emphasizes citation integrity (S6) and competing hypotheses (S5) over speculative pre-mortem.

Resolution: S4 pre-mortem findings in research mode are NOT written to a standalone section. Instead:

• If a finding meets the S7 threshold (confidence ≥ TIER_2_CONFIDENCE_FLOOR) AND can carry an evidence label → promote it to S7 as a regular objection.
• If a finding is below threshold or is pure speculation → dropped. Count it in the S7 dropped-count footer.

Output composition: → No dedicated section in research mode. Qualifying findings feed into Section 3 (Objections) via S7. Below-threshold findings feed the dropped-count footer.

---

S5 — Analysis of Competing Hypotheses (ACH)

Purpose: Enumerate alternative explanations or hypotheses that would equally or better account for the same evidence. Forces acknowledgment of alternatives the work ignores.

How:

1. Identify the central hypothesis (H1 = the work's claim).
2. Generate at least 2 alternative hypotheses (H2, H3, ...) that could explain the same evidence.
3. Build an evidence × hypothesis matrix:
   • Rows: pieces of evidence cited in the work
   • Columns: each hypothesis (H1 through H)
   • Cells: consistent (+), inconsistent (−), neutral (~), or unknown (?)
4. Score each hypothesis: count − marks (inconsistencies). Fewer inconsistencies = more viable.
5. Note which hypotheses the work fails to consider.

Output format:

### Hypotheses

- **H1** (author's claim): <statement>
- **H2**: <alternative>
- **H3**: <alternative>

### Evidence × Hypothesis Matrix

| Evidence | H1 | H2 | H3 |
|----------|----|----|-----|
| E1: ...  | +  | +  | −   |
| E2: ...  | +  | ~  | +   |
| ...      |    |    |     |

**Inconsistency count:** H1: 0, H2: 1, H3: 1
**Conclusion:** <1-2 sentences on which alternatives are viable and whether H1 is uniquely supported>

Output composition: → Section 7 (Competing Hypotheses / ACH matrix)

---

S7 — Structured Objection List

Purpose: Produce the tiered objection list that drives the verdict.

How:

1. Reserve O-numbers for S6 shepardize promotions (auto-Tier-1). These are already filed.
2. Collect all remaining objection candidates from S2 (flagged assumptions), S4 (promoted pre-mortem findings), S5 (unaddressed alternatives), and any direct analysis.
3. For each candidate, assign:
   • Evidence label (from agent file taxonomy — mandatory)
   • Confidence (0–100%)
   • Tier: Tier-1 if confidence ≥ TIER_1_CONFIDENCE_FLOOR; Tier-2 if ≥ TIER_2_CONFIDENCE_FLOOR; dropped otherwise
4. Apply Cluster Rule (from D8): 3+ Tier-2 objections converging on the same assumption → synthesize into a Tier-1 cluster objection.
5. Apply Hard Cap: maximum TIER_1_HARD_CAP Tier-1 objections total (including S6 promotions). If exceeded, collapse by theme and surface strongest representatives.
6. Drop sub-threshold findings; count them for the footer.

Objection block format:

**O<N> — <Short title>** [LABEL]
Tier: <1 or 2>
Confidence: <range or value>%
Consequence if true: <1 sentence — what breaks if this objection holds>
<Detail paragraph with evidence>

Footer (after all objections):

N sub-threshold concerns considered and dropped (confidence < TIER_2_CONFIDENCE_FLOOR%).

Output composition:

• Section 3 (Objections) — Tier-1 followed by Tier-2, each labeled by tier
• Section 8 (Unfalsifiable / Out of Scope) — items declined, with reason

---

S8 — Forced Verdict

Purpose: Produce the final verdict. Forced — pick exactly one tier. No hedging.

How:

1. Review all Tier-1 objections (including auto-Tier-1 shepardize findings).
2. Apply the 4-tier verdict scale from adversarial-reviewer.md:
   • No material Tier-1 objections → SURVIVES
   • Tier-1 objections present but none falsify the central claim → SURVIVES WITH CAVEATS
   • At least one load-bearing Tier-1 objection; central claim weakened → WOUNDED
   • Central claim falsified or wholly unsupported → DOES NOT SURVIVE
3. Write the Bottom Line: 2–3 sentences, references only Tier-1 objections, no hedge.
4. Write "What Would Change the Verdict" — mandatory: explicit conditions under which the verdict would move up or down a tier.

Anti-weasel check: Read the verdict sentence. If it contains "it depends", "further analysis needed", or any hedge — rewrite it.

Output composition:

• Section 1 (Header) — verdict as first line, plus: target ref, mode (research), scope, timestamp, reviewer model
• Section 2 (Bottom Line) — verdict rationale, Tier-1 refs only
• Section 9 (What Would Change the Verdict) — mandatory

---

Output Artifact Composition

After all pipeline steps complete, assemble the artifact at:

~/.pi/agent/reviews/<slug>-<timestamp>.md

Sections in order (per D6, research-mode subset):

Section	Name	Written by
1	Header (verdict, target, mode, scope, timestamp)	S8
2	Bottom Line	S8
3	Objections (Tier-1 then Tier-2)	S6 promotions + S7
4	Citation Shepardizing table	S6 (adversarial-shepardize output)
5	Key Assumptions Surfaced	S2
6	(Pre-mortem — change mode only; omit in research mode)	—
7	Competing Hypotheses / ACH matrix	S5
8	Unfalsifiable / Out of Scope	S7
9	What Would Change the Verdict	S8

Print the output file path at the end of every session.