Download analysis/preregistration.md from siddharthmb/2026.RA.Game-Structure-Generality: direct link, hf CLI and curl.
- Browser
- Download file 17.9 kB
-
https://huggingface.co/datasets/siddharthmb/2026.RA.Game-Structure-Generality/resolve/main/analysis/preregistration.md
- Command line
-
hf download hf://datasets/siddharthmb/2026.RA.Game-Structure-Generality/analysis/preregistration.md
-
curl -L -o preregistration.md https://huggingface.co/datasets/siddharthmb/2026.RA.Game-Structure-Generality/resolve/main/analysis/preregistration.md
Preregistration — does the talk channel's value generalize across game structures?
[Results → research note 0086 · control dir game_structure_generality_v1/ · run dirs under /nlp/scr/siddharth/ii_mats/rational_agents/gsg_v1__*]
Extends: 0044 — moves-only matched control · 0049 — 4-of-5 majority arms · 0058 — quorum × rounds sweep · 0069 — generality bank, free arms · 0081 — discovery vs coordination
Why
Every headline number in the talk-channel paper comes from one game family: five parties, four issues of four options (256 packages), unanimity, private score sheets and thresholds, bank instances_five_seat_private_v1. There, five talking Opus 5 seats close 0.958, the same five seats with every public message disabled close 0.758 (paired −0.200 [−0.325, −0.092], note 0044), and five Bayesian seats close 0.225. A reviewer asks whether this holds "with different numbers of participants, other task structures". The paper's own account — unanimity turns closing into a coordination problem, and talk supplies agreement on a package that has already been found (note 0081: −0.125 of the −0.200 is coordination) — makes a directional prediction: the channel should matter more when more parties must be aligned under unanimity, and less when the agreement rule no longer needs everyone.
Structures
All banks come from the one generator, generate_five_seat_bank.py, now parameterized by --n-parties (default 5 reproduces every frozen bank). The generator calibrates thresholds so the same FRACTION of the 256 packages clears every threshold at each difficulty level (ladder 0.16 → 0.015, levels cycling over 24 instances), so a party-count change moves the number of parties that must agree while holding the zone-of-agreement size fixed. The screening matches bank v1 exactly: 24 instances, same ladder, same feasibility guarantee (every instance admits an all-IR package), same integerization and exact re-solve, no infeasible stratum.
| structure | bank | rule | Opus arms (new unless noted) | Bayesian arm |
|---|---|---|---|---|
| S3: three parties | instances_n3_private_v1 (24, first seed 53000) |
unanimity (3 of 3) | all_llm, all_llm_moves_only, seeds 0–2 (72 episodes each) |
all_rational, seeds 0–4 |
| S5: headline | instances_five_seat_private_v1 (24) |
unanimity (5 of 5) | frozen all_llm (note 0030 campaign) and all_llm_moves_only (note 0044), seeds 0–4 |
all_rational, seeds 0–4, re-run on this engine (frozen arm kept as a vintage check) |
| S7: seven parties | instances_n7_private_v1 (24, first seed 57000) |
unanimity (7 of 7) | all_llm, all_llm_moves_only, seeds 0–2 (72 each) |
all_rational, seeds 0–4 |
| Q4: agreement rule | bank v1 re-materialized at min_accept=4 (veto seat retained, the semantics of notes 0049/0058) |
4 of 5 | frozen all_llm_quorum_4 (note 0049); NEW all_llm_moves_only_quorum_4, seeds 0–4 (120) |
all_rational_quorum_4, seeds 0–4 |
| I (only if budget remains after S3/S7/Q4) | a five-party bank with more packages (5 issues × 4 options = 1,024), else cut | unanimity | all_llm, all_llm_moves_only, seeds 0–2 |
all_rational, seeds 0–4 |
Seven parties is the largest table worth running: the engine names at most eight seats, the generator and exact solver handle seven in seconds, and cost grows roughly with the square of the party count (seats × context).
Outcome-blind feasible-set statistics of each bank (bank_structure_stats.py, game_structure_generality_v1/bank_stats.md), computed before any episode:
| bank | parties | feasible / knife-edge / infeasible | NBS≠KS | zone size (packages clearing the rule) | IR margin | pivotal-seat burden | difficulty scalar |
|---|---|---|---|---|---|---|---|
| S3 | 3 | 23 / 1 / 0 | 0.29 | 16.5 [8.5, 27.8] | 0.265 [0.233, 0.31] | 0.75 [0.668, 0.829] | 0.693 [0.664, 0.721] |
| S5 (bank v1) | 5 | 22 / 2 / 0 | 0.25 | 17 [8, 27.5] | 0.292 [0.159, 0.332] | 0.52 [0.413, 0.65] | 0.641 [0.61, 0.685] |
| Q4 (bank v1 at 4 of 5) | 5 | 22 / 2 / 0 | 0.25 | 60 [33.2, 80] | 0.292 [0.159, 0.332] | 0.52 [0.413, 0.65] | 0.641 [0.61, 0.685] |
| S7 | 7 | 24 / 0 / 0 | 0.21 | 15 [10, 27] | 0.305 [0.231, 0.338] | 0.405 [0.291, 0.485] | 0.622 [0.596, 0.645] |
Median [interquartile range]. The zone of agreement is matched across party counts (≈16 of 256 packages); under 4-of-5 it is 3.5× larger.
Held fixed (identical to the frozen campaign)
Model anthropic:claude-opus-5, adaptive thinking on every seat, 16,384-token turn floor; private sheets and whole-number thresholds; datacenter framing, canonical scaffold; four rounds with rotating proposer offset by the episode seed, then a forced final vote; per-round discount 0.95. Muting is the protocol arm moves_only: every seat gets the moves-only action contract and no message is published (note 0044). All cells run through run_five_seat_cell.py (generalized to read the party count from the bank), on Slurm partition john, one host per structure where the scheduler allows.
Endpoints
Primary, per structure (deal rate = fraction of episodes ending in an agreement):
- Muting cost E1 = deal(
all_llm_moves_only) − deal(all_llm), paired on (instance, seed). - Opus-minus-Bayesian gap E2 = deal(
all_llm) − deal(all_rational), paired on (instance, seed).
Intervals: 95% percentile bootstrap over instances (each instance's seeds resampled together), 10,000 draws. Cross-structure comparisons (P1) are differences between two independently bootstrapped banks (each bank resampled separately), and are labelled cross-bank. The Q4 comparison (P2) is a paired difference-in-differences on bank v1's identical (instance, seed) keys.
Secondary: the discovery/coordination census of note 0081 (analyze_zopa_discovery.py: zone tabled, closure given tabled, the product-rule split of E1 into a discovery and a coordination term; the zone is the unanimity all-IR set, so under Q4 a deal outside it shows up as a remainder, which is reported); normalized score; parties left below threshold by a closed deal; walks; the ICC of the deal outcome per arm (between-instance share of variance), which says how much the 3-seed cells lose against 5 seeds.
Predictions (directional, with falsifiers)
- P1 — muting cost grows with party count under unanimity. Predicted order E1(S3) > E1(S5) > E1(S7) (less negative to more negative). Confirmed if E1(S7) − E1(S3) < 0 with a cross-bank interval excluding zero; direction-only if the three point estimates are in the predicted order but that interval includes zero; rejected if E1(S7) − E1(S3) ≥ 0 at the point estimate.
- P2 — muting cost shrinks under 4-of-5. Predicted E1(Q4) − E1(S5) > 0. Confirmed if the paired difference-in-differences interval excludes zero in that direction; direction-only if positive with an interval including zero; rejected if ≤ 0.
- P3 — talking Opus tables out-close Bayesian tables in every structure. E2 > 0 with an interval excluding zero in S3, S7 and Q4. A Q4 Bayesian table closed 0.492 in note 0058, so the Q4 gap is predicted smaller than S5's.
- P4 — the muting cost is mostly coordination. In every structure where E1 is resolved, the coordination term exceeds the discovery term. Rejected for a structure if the discovery term is the larger.
A ceiling caveat is registered now: if a talking table closes every episode in some structure, E1 there is bounded by the muted table's shortfall alone, and a small cost is then partly a ceiling effect. Reported, not corrected.
Power
On bank v1 the 5-seed muting cost had interval half-width ≈0.12. At 3 seeds × 24 instances the half-width is expected near 0.13–0.15 (between-instance variance dominates the deal outcome for these arms; the ICC is measured and reported). A cross-bank difference therefore resolves only if it exceeds roughly 0.25–0.30. P1 is a test of a large effect; a smaller true gradient will read as direction-only. That is a property of the $800 budget, stated in advance.
Validity gates (checked before any result is read)
- G1 every cell has its planned episode count with status done; episodes lost to HTTP 529 or other API errors are re-run with
fill_five_seat_cell_tail.pyon the same node and merged before analysis. - G2
audit_empty_turns/gen_failures= 0 fabricated turns per cell, reported by phase. - G3 protocol check: zero published messages in every moves-only cell, non-zero in every talking cell.
- G4 banks: every instance feasible under its rule, party count and min_accept as declared, SHA-verified by the runner.
- G5 voided turns (
stop_reason == "refusal") counted per cell and by phase; the primary estimates include every episode, and a sensitivity analysis drops episodes in which a seat was voided at the forced final vote (the mechanism shown in notes 0074/0073 to manufacture no-deals). - G6 vintage check: the re-run S5
all_rationalagainst the frozen one (statistical equivalence only; 3–6% of Bayesian episodes flip across hosts).
Budget
Hard cap $800 including smokes. Measured references on bank v1: all_llm $1.31/episode, all_llm_moves_only $0.96, all_llm_quorum_4 $1.15. Estimates: S3 ≈ $0.5/$0.35 per episode (≈$65 for both cells), S7 ≈ $2.5/$1.9 (≈$320), Q4 muted ≈ $115. A two-episode smoke per party count fixes the per-episode cost before the full cells launch, and each cell gets a mechanical --budget-usd. Priority if money runs short: S7 and S3 first, then Q4, then I.
What would make a number unsafe to quote
The S5 Opus arms are frozen cells from an earlier engine vintage paired against this lane's cells only within S5; S3 and S7 cells are all from this lane. Cross-structure contrasts compare different banks and are labelled cross-bank. Any cell short of its plan after fills is reported as partial and not interpreted.
Amendment A1 — 2026-11-12, before any scored Opus episode (protocol change forced by the API)
What happened. The first two smokes (3- and 7-party all_llm) returned stop_reason = "refusal" with zero output tokens on 16/16 calls. Direct API tests show claude-opus-5 (and claude-fable-5) now refuse the canonical negotiation prompt deterministically, including the stored first-turn views of the frozen bank-v1 all_llm cell that played normally in August, on both API keys; claude-sonnet-5 and claude-opus-4-8 do not. The engine stores each refusal as the placeholder abstention "(ran out of time this turn and says nothing substantive)" with parse_ok = True, so a cell would have completed as an all-abstain table. Leave-one-out bisection of the system prompt isolates one line, the scratchpad field help (private notes to yourself. NEVER shown to any other party. Reason here freely — writing here is not speaking and costs you nothing.). Replacing it with working notes for your own planning; other parties do not see it. makes 8/8 stored views complete (bank v1, 3-party, 7-party; first and last turn; talking and moves-only).
Change. Every Opus cell of this family runs the scaffold canonical_neutral_scratchpad (prompts.VARIANTS; run_five_seat_cell.py --scaffold), identical to canonical except that one line. This is a new protocol version, so the frozen canonical Opus cells (bank-v1 all_llm, all_llm_moves_only, all_llm_quorum_4) can no longer serve as paired references. Design changes, all decided before any scored Opus episode:
| structure | Opus cells under the new scaffold | seeds |
|---|---|---|
| S3 | all_llm, all_llm_moves_only on instances_n3_private_v1 |
0–2 |
| S5 | all_llm, all_llm_moves_only on bank v1 (NEW re-runs; the frozen canonical cells are reported only as context) |
0–2 |
| S7 | all_llm, all_llm_moves_only on instances_n7_private_v1 |
0–2 |
| Q4 | all_llm_quorum_4 (NEW re-run), all_llm_moves_only_quorum_4 on bank v1 at 4 of 5 |
0–2 |
The Bayesian cells are model-free, see no prompt, and stand as run (all four 120/120). P2's difference-in-differences now pairs the re-run Q4 and S5 cells on bank v1's 72 shared keys. The issue-structure arm (I) is cut to fit the eight Opus cells under the $800 cap. Predictions, endpoints, gates and power reasoning are unchanged. A further registered check: the new S5 talking and muted deal rates are compared with the frozen canonical 0.958 / 0.758 as a cross-protocol context row (not a paired contrast); a large shift would mean the scratchpad wording itself matters and would be reported as such.
Amendment A2 — 2026-11-12, supersedes A1, before any scored Opus episode
Decision (team lead). No reworded prompt is used for any result of this family. Claude Opus 5 and Claude Fable 5 decline the canonical negotiation prompt at the provider level (a provider safeguard); the block was localised to the scratchpad instruction. A1's canonical_neutral_scratchpad scaffold variant and the --scaffold flag are removed from the harness. The two gsg_v1_smoke2__* runs made under that scaffold were stopped unread by the team lead; their outcomes are excluded from every analysis and are not opened or quoted. No further probing of the classifier is done.
Design. Every hosted cell runs anthropic:claude-opus-4-8 on the unmodified canonical scaffold (the model the other revision lanes moved to; in the cell allowlist). Thinking, turn floor, framing and protocol are otherwise as registered above. The structures, all self-contained (talk, mute and Bayesian cells on the same bank and keys):
| structure | Opus 4.8 cells (canonical scaffold) | seeds | Bayesian cell |
|---|---|---|---|
| S3 | all_llm, all_llm_moves_only on instances_n3_private_v1 |
0–2 | gsg_v1__n3__all_rational, 0–4 (as run) |
| S7 | all_llm, all_llm_moves_only on instances_n7_private_v1 |
0–2 | gsg_v1__n7__all_rational, 0–4 (as run) |
| Q4 | all_llm_quorum_4, all_llm_moves_only_quorum_4 on bank v1 at 4 of 5 |
0–2 | gsg_v1__n5q4__all_rational_quorum_4, 0–4 (as run) |
| S5 (reference, not run here) | the model-breadth lane's Opus 4.8 all_llm / all_llm_moves_only pair on bank v2 (note 0084), subsampled to seeds 0–2 for the like-for-like comparison |
0–2 | an all_rational cell on bank v2 (the bank-v1 cell gsg_v1__n5__all_rational stays as the Q4 unanimity anchor) |
Consequences for the predictions. P1 compares E1 across the 3-party, five-party (bank v2) and 7-party banks, all cross-bank. P2's "unanimity" side is no longer on bank v1 for Opus 4.8, so the paired difference-in-differences on identical keys is not available; P2 is tested as the cross-bank difference E1(Q4, bank v1) − E1(S5, bank v2), labelled as such, with the bank-v1 Bayesian quorum-vs-unanimity contrast reported as the paired companion. The issue-structure arm stays cut.
Gate added. Every hosted cell is gated on refusals: at most 5% of seat turns with stop_reason = "refusal", and in the smoke none at a final vote; a cell that fails is reported, not interpreted.
Amendment A3 — 2026-11-12, before any scored Opus episode: the 4-of-5 structure moves to bank v2
Under A2 the five-party unanimity reference is the model-breadth lane's Opus 4.8 talk/mute pair on bank v2 (instances_five_seat_private_v2). Running Q4 on bank v1 would have made P2 a cross-bank comparison. Q4 therefore runs on bank v2 re-materialized at 4 of 5 (veto seat retained), so P2 is again the registered paired difference-in-differences on identical (instance, seed) keys: E1(Q4, v2) − E1(S5, v2), on seeds 0–2 (the reference cells subsampled to those seeds). Bank v2 is the bank note 0058's Bayesian quorum sweep used, which is why that note's 4-of-5 closure (0.492) differs from this lane's bank-v1 Bayesian quorum cell (0.617): different banks, not a discrepancy. Added free Bayesian cells: gsg_v1__n5v2__all_rational and gsg_v1__n5v2q4__all_rational_quorum_4 (seeds 0–4). The bank-v1 Bayesian cells already run are kept as context. Outcome-blind statistics of bank v2 and its 4-of-5 re-materialization: bank_stats_v2.md (zone 17 [7, 28.5] packages at unanimity, 59.5 [32, 77.8] at 4 of 5; NBS≠KS 0.38).
Amendment A4 — 2026-11-12, before any full-cell result is read: leftover budget buys seeds 3–4
If money remains under the $800 cap after the six 72-episode Opus 4.8 cells, it buys seeds 3–4 (48 more episodes per cell) for the 3-party and 7-party talk and mute cells, not the issue-structure arm (which stays cut). Seeds 3–4 run as separate cells under identical settings and are merged with seeds 0–2 on (instance, seed) keys. Allocation rule: if the projected cost of all four extensions (at each cell's measured $/episode) fits, run all four; otherwise the party count whose talk and mute cells have the LOWER mean ICC of the deal outcome gets its seeds first (a lower ICC means extra seeds shrink the interval more), and the other follows only if money remains. The primary analysis then uses every seed run; the seeds 0–2 estimate is reported beside it so the extension cannot be chosen after seeing its result.