# Prompt: adversarial comparison of research documents (Fable vs Codex)

> Usage: paste the following block into a third agent (ideally a different model from the two authors).
> Replace `<PATH_CODEX>` with the actual path to Codex's research before executing.

---

You are an adversarial research arbiter. You have two research documents answering the same question, written independently by two distinct models:

- Document A (Fable): `docs/research-fable-agentes-curva-u.md`
- Document B (Codex): `<PATH_CODEX>`

Original question both answer: *"Why does an AI agent work well in small repos, collapse at an indeterminate point as the repo grows, want to rewrite everything, and regain bearings when software is structured with sealed compartments and extensive tests? Is it studied? Scientific/mathematical profile."*

Your job is NOT to summarize or praise: it is to attack both documents with equal rigor and produce a verdict per claim and synthesis better than either one.

## Impartiality rules

1. Ignore authorship, style, length and format. Only count: verifiable accuracy, calibration and explanatory power.
2. If you're a model from the same family as either author, declare it in your output's first line and apply extra skepticism to that family's claims (self-preference bias).
3. No points awarded for rhetoric. Each point you assign must cite the evidence justifying it.
4. Verbosity is not coverage: a claim repeated in three ways counts once.

## Process (in order, don't skip steps)

### Step 1 — Normalized extraction
Extract from each document its list of factual claims and formal models, numbered (A1..An, B1..Bn). A claim = a falsifiable assertion. Discard filler.

### Step 2 — Alignment
Build the alignment matrix: for each claim, classify:
- **COINCIDE**: both assert it (strong signal: independent convergence).
- **CONTRADICT**: assert incompatible things → must go to primary source.
- **UNIQUE-A / UNIQUE-B**: only one has it → decide if *omission* by the other (point for the one who has it, if it survives verification) or *correct dismissal* (point against the one who included it).

### Step 3 — Source verification (the part you cannot skip)
For each citation (paper, benchmark, figure):
- Try to access the primary source (fetch from arXiv/URL). Verify that (a) the paper exists, (b) it says what the document attributes to it, (c) the figures match.
- Nonexistent or invented citation = GRAVE failure (−3 to author's accuracy, noted explicitly).
- Real citation but misrepresented = serious failure (−2).
- Real citation, imprecise attribution but directionally correct = minor failure (−1).
- If you can't verify, mark **UNVERIFIED** — doesn't count for or against, but list it.

### Step 4 — Attack
For each CONFIRMED claim try actively to refute it: counterexamples, papers saying the opposite, logic errors in mathematical derivation. A formal model (e.g. K vs C_eff, percolation, O(εT²)) is evaluated by: is it falsifiable? is it consistent with cited evidence? is it operationalized or just metaphor? Metaphor presented as metaphor isn't penalized; metaphor presented as result is.

### Step 5 — Calibration
If a document labels confidence (HIGH/MEDIUM/SPECULATIVE) or pre-registers weaknesses, evaluate *calibration quality*: did HIGH claims survive better than MEDIUM ones? Honest labels matching your verification results add to calibration; a document without confidence tags is treated as asserting everything with HIGH confidence (and is penalized equally as a failed HIGH). Pre-registered weaknesses don't exempt from error, but add to intellectual honesty.

### Step 6 — Score and synthesis

Score each document 0-10 on:

| Dimension | What it measures |
| --- | --- |
| Verified accuracy | claims that survived steps 3-4, weighted by centrality |
| Calibration | declared confidence vs verification result |
| Coverage | relevant claims the other lacks and survived |
| Explanatory power | does the mathematical model explain the full U-curve, including "abruptness" and rewrite impulse? |
| Actionability | concrete design implications for this repo's tools (prd-vigia.md, prd-objective-guard.md) |
| Intellectual honesty | declared provenance, recognized limits, absence of overclaiming |

## Output format (mandatory)

1. Conflict-of-interest declaration (rule 2).
2. Complete alignment matrix (table: claim | A | B | relation | verdict | evidence).
3. List of citations that failed verification, by author and severity.
4. Scores per dimension with one-line justification each.
5. Global verdict: which document is more reliable and WHY in ≤5 lines.
6. **Merged synthesis**: the best possible document using only claims that survived, with attribution [A]/[B]/[both] per claim. This synthesis is the principal deliverable.
7. "What both ignored": shared gaps — the list of questions neither attacked.

Constraints: don't invent verifications you didn't perform — if you didn't fetch a source, it's UNVERIFIED. Suggested budget: deepen on the 5-8 most central citations rather than shallowly verify all.