# Pre-registration — Study 2 Phase A: the contact factor — 2026-07-02

Filed before the run. Graduates the Phase A portion of
`DRAFT_study2_independence_x_contact.md`; Phase B (independence) remains
drafted, gated on multi-provider access. Inherits Amendment 1 and every
Section 6 stop-rule of the 2026-06-25 registration.

## 1. Question

Does **contact** — the ability to check a claim against something outside the
conversation — discipline the confidence that debate manufactures? Study 1's
corrected finding and its replicate showed debate converting disagreement into
confidence without accuracy (7-agent arm: +0.043 overconfidence, exactly zero
accuracy change). Phase A asks whether giving the same debaters a tool changes
that.

## 2. Design (pinned)

Runner: `scripts/run_study2_contact_v2.py` at the commit that files this
document. Model `claude-haiku-4-5-20251001`, temperature 0.7 everywhere,
**7 agents, 2 rounds** (the replicate's confirmed configuration). Per item, two
cells, equal votes by construction:

- **NT (no tools):** debate exactly as in the replicate.
- **T (tools):** identical, except each agent turn may run ONE sandboxed Python
  computation (`instrument.contact.run_python`: no network, resource-limited)
  and sees its printed output before committing Answer/Confidence. Tool calls
  are logged (TOOL_CALLED/TOOL_RETURNED), so per-claim grounding is metered.

Round-0 answers and per-turn tool use are checkpointed. Cell order per item:
NT then T (fixed; no answer-state crosses cells). The T cell spends more
tokens by design; the within-cell pre/post contrast is compute-neutral, and
NT-vs-T token totals are reported.

## 3. Battery (frozen)

n = **122**: 110 items selected from the completed difficulty-calibration pass
(797 fresh items, 5-resample error rates; per-item CSV published) by the rule
pinned in the runner —

- all 12 empirically-medium items;
- 48 empirically-hard: all non-TriviaQA hard, plus TriviaQA hard by spread
  (desc, id tiebreak) to cap the TriviaQA skew;
- 36 easy-with-disagreement (spread > 0.2, by spread desc);
- 24 clean easy (seeded hash order, seed 20260702)

— plus the **12 human-truth-locked fresh controls** (7 post-cutoff, 5 private
reasoning) frozen at commit `028643f`
(`preregistration/2026-07-02_fresh-items-locked.md`; one drafted answer was
caught wrong by human verification and dropped). Composition: 42 GSM8K /
36 TriviaQA / 32 MMLU / 12 locked. No item was used in Study 1 or the
replicate.

## 4. Registered predictions (stated so they can fail)

**P-A (primary).** The within-swarm debate-inflation delta (post-debate
overconfidence minus round-0 overconfidence) is **smaller in T than in NT**:
the paired bootstrap 95% CI of (Δ_NT − Δ_T) lies above 0.
*Falsified if* the CI includes or lies below 0 — i.e., contact does not
discipline debate (or worsens it). Reported either way, same prominence.

**P-B (mechanism).** Within the T cell, final answers whose agent traced a tool
call (grounded) are better calibrated than ungrounded ones: overconfidence
(grounded) < overconfidence (ungrounded), CI excluding 0.
*Falsified if* grounded and ungrounded claims calibrate the same or grounding
is worse. If tool uptake is below 10% of turns, P-B is reported as
**underpowered-by-uptake** (a finding about tool neglect, not calibration).

**P-C (secondary, estimation).** Accuracy effect of tools on the
empirically-hard stratum, with CI — the "contact is the dominant term"
sub-claim's first quantitative estimate. No significance commitment.

**Registered descriptives:** per-cell accuracy/confidence/overconfidence/ECE/
Brier/AUC (overall and by stratum/source/locked-slice); round-0 unanimity;
diversity by round; tool-use rate by round and by stratum; NT-vs-T token
totals.

## 5. Analysis conventions (pinned)

10-bin equal-width ECE; tie-corrected Mann-Whitney AUC; paired item-level
percentile bootstrap, B = 10,000, seed 20260702; `instrument/metrics.py`
unchanged. Analysis script committed before the run completes; published with
the run. No slicing beyond the registered descriptives.

## 6. Cost, gates, honesty

Worst-case ≈ 5,124 calls (~$15–20 at Haiku pricing; realistic ~$10–15), under
the key-level cap; checkpointed, resumable, mock-validated before spend. Gates
at filing: calibration pass complete (✓, published), ≥10 locked controls
committed (✓, 12 at `028643f`), this document filed (✓). The Anthropic API
accepts no seed parameter; runs are not seed-reproducible — model id,
temperature, prompts, item ids, and analysis seed are pinned instead. All
Section 6 stop-rules of the 2026-06-25 registration apply; additionally, the
run does not start until the tool-sandbox self-test passes on the run host.