docs: research paper + presenter pack (written from recorded results only)

[ksdisch]

What this is

Two Markdown deliverables under docs/paper/, produced entirely from numbers already recorded in the repo — no new measurements, no model/API calls, no figure regeneration. This PR is for your review; per the mission it is not merged.

• docs/paper/forge-gap-paper.md (~4,300 words) — a professional empirical-methods paper: abstract, introduction, background/related work, methods, experimental setup, results (with the six committed figures embedded and a consolidated results table), discussion, limitations, conclusion, provenance note, references.
• docs/paper/forge-gap-presenter-pack.md (~1,500 words) — the defense companion: one-minute story, claim-by-claim defenses, a provenance table (claim → number → source file), anticipated hard questions with crisp answers, and how to trace any figure live.

Headline results, as written in the paper

Experiment (stage)	Result	Verdict
Error-recovery on injected transient faults, GLM-4.6, N=40	67.5% → 100%, +32.5 pp, Newcombe 95% [+17.3, +48.0]	real
Retry-nudge (and error-recovery) on injected malformed calls, GLM-4.6, N=20	all arms 100%, +0.0 pp [−16.1, +16.1]	null — reported as one
Submit-nudge on mistral-nemo's natural no-submit gap, N=20	0% → 75%, +75.0 pp [+47.8, +88.8]; retry-nudge null (0 fires) in the same run	real
Validation stacked on submit-nudge, mistral-nemo, N=40	75% → 100%, +25.0 pp [+11.1, +40.2]	real
Validation un-stacked on llama-3.1-8b's hallucination gap, N=40	0% → 45%, +45.0 pp [+28.2, +60.2]; 55% residual decomposed (35% / 10% / 7.5% / 2.5%)	real

Honesty framing held throughout: injected gaps labeled injected on every mention; both nulls reported at full prominence; nemo's capstone +100.0 pp [+81.7, +100.0] explicitly labeled cross-run; GLM-4.6 given no guardrail bar (none was ever measured on its clean task); validation presented as a self-consistency check with its fooled-152 proof and measured blind spot.

Sources of truth used

All arm rates, CIs, and mechanism-fire counts: docs/figures/*-data.json. Pilots, sizing analyses, the S10 residual decomposition, and robustness probes: docs/DECISIONS.md (D12–D24) with ROADMAP/README cross-checks. Design facts: scenario.py, agent.py, faults.py, stats.py. The 77-test count was verified by direct grep. No numbers were taken from runs/.

Citations

The repo names no publications for the guardrail primitives, so the paper frames them as reproduced common practice (stated explicitly in References) and cites only the canonical statistics references the repo names by method: Wilson (1927) and Newcombe (1998, method 10).

Flagged gaps (numbers that don't exist in the repo)

• Per-run token/dollar costs for the main ablations were never recorded (only the S7 pilot's token counts exist, in D20). Both documents therefore make no absolute cost claims; this is disclosed in the paper's Limitations.

Scope

Only the two new files — no project code, tests, figures, or existing docs/ spine files were touched.

🤖 Generated with Claude Code