siddharthmb/2026.IC.Logical-Copy-Coordination

Isolated Logical-Copy Coordination

Confirmatory v3 closure battery

The latest release contains 11,584 actually paired trials (23,168 isolated root tasks; 26,478 successful provider calls) across five preregistered families: actual-by-stated model identity, decision thresholds, randomized label conventions, complementary serializer/parser code, and shared/private/no randomness at two temperatures. Behavioral failures remain intention-to-treat zeros. Every accepted family has a complete hash inventory and accepted manifest; the public audits/v3-final-isolation-integrity-audit.json passes all hard isolation/integrity gates.

The sole primary actual-by-stated identity interaction was -20.90 percentage points (95% item bootstrap interval -26.76 to -15.04; two-sided sign-flip sensitivity p approximately 1e-5), opposite the hypothesized positive logical-copy effect. Its product-marginal component was -21.19 points, while designated-minus-off-diagonal excess was only +0.39 points (Holm p=.807). None of 13 designated-pair residuals triggered replication. The decision-theory briefing changed individual policies, and self-calls carried predictive information in the codec task, but neither produced evidence of pair-specific logical coupling.

Accepted raw archives are under accepted/; final analysis archives are under analysis/; excluded operational windows and outcome-blind ledgers are under quarantine/ and exclusion-ledgers/; preregistration, frozen replication contingencies, and audit code are under preregistration/. analysis/v3-e5-balance-gate-correction.json records a deterministic post-analysis correction that recomputes E5 contrasts with the preregistered balance-gated endpoint and uses no new data. The corrected release is analysis/confirmatory-v3-r9-balance-corrected-analysis.tar.gz (SHA-256 3a8551ad39e92d7f6b721538a0f393a21ac68a9ba5abb863cd96f1e85cbc8f29); its embedded completion.json records Slurm job 16731868 (COMPLETED, exit 0:0) and every output hash. The r8 archive is preserved but superseded for E5 inference. documentation/main.pdf is the five-page V3 report (SHA-256 ef583ea78b9896a927ca13461fc608e9f11bf7726357e048420d3835fef26ed1). The exact corrected analyzer and regression test are public under analysis/v3-e5-correction/ (SHA-256 f8f860fc236d0799ef7d80f41f4e708721805b7f217fd7b7ef70a6d4d9c4b3a6 and 4906fbc60c9f5d6f5817e2b119b34d9f72f888149eecae03a3fd9353ad0ee13f) at immutable commits e0abe2789a96a89eed81a9fa73fa2a9893d440c5 and 24fffc37ce3190025d68574a8b625aa9518cc2cc. The corrected literature review is SHA-256 09ac68c0c4d28ae2478664994868e36c8e5956073a87ed4b8a207840ad9bdcf8; immutable synchronized source/data snapshot 68088cdaff39d87d0ed9d84ca0e6d49db70a570d contains it alongside r9, the completion ledger, correction sources, and the preceding verified V3 PDF.

“Same model” means the same returned endpoint ID, not verified weights, RNG, routing, hardware, or opaque provider state. The audit verifies coordinator-side requests and absence of experiment- provided communication channels; it is not provider attestation or OS isolation of tool-enabled agents. Pair errors remain intention-to-treat zeros. A non-gating pacing diagnostic flagged 94 dispatch gaps below the nominal spacing threshold; all hard isolation and integrity gates passed.

Initial v1 study (legacy)

The initial release contains 780 actually paired, independently sampled language-model trials testing coordination with no communication, behavioral evidence, or shared agent-visible files. It includes 420 Claude Fable 5 pairs across seven prompt conditions, 120 Fable 5 pairs with independent private context, and 240 Claude Haiku 4.5 pairs split between temperature 0 and temperature 1.

Each pair plays either an arbitrarily labelled one-shot copy Prisoner's Dilemma or produces sealed serializer/parser Python artifacts that are composed only after both calls terminate. Full prompts, visible reasoning, tool events, generated code, token usage, API errors, and pair-level evaluations are included. The optional self-copy tool launches a fresh isolated emulation and never observes the actual partner.

Experiment names

experiment-name	Description
isolated-logical-copy-v1	Paired LLM trials, including main, private-context, and temperature runs.
formal-epsilon-grounded-tft	Exact epsilon-grounded program-controller self-play and AlwaysDefect controls.

Run names:

• fable5-main-v1: Fable 5, seven conditions, 30 pairs per task/condition.
• fable5-private-context-v1: Fable 5, three conditions, 20 pairs per cell, independent irrelevant notes.
• haiku45-temp0-v1: Haiku 4.5, three conditions, 20 pairs per cell, temperature 0.
• haiku45-temp1-v1: Haiku 4.5, three conditions, 20 pairs per cell, temperature 1.

Files

• data/pair_outcomes.jsonl: compact pair-level table used for analysis. Every row has an experiment-name column. It reports both the original strict artifact validator and the corrected validator that admits harmless startswith/endswith/isspace string methods.
• data/full_records/*.jsonl: complete nested pair records, including prompts, reasoning, code, tool-call histories, usage, errors, and hidden-test details.
• data/formal/*.jsonl: exact epsilon-grounded-controller trials.
• metadata/*/{config,manifest}.json: resolved parameters, API usage/cost, SDK version, and record hash.
• analysis/: cell summaries, preregistered and cross-run contrasts, and a vector PDF figure.

Regeneration

From repository root at /juice2/u/siddharth/ii_mats, with the Anthropic key loaded by source set-env.sh, run:

uv run python experiments/logical_copy_coordination/run_experiment.py \
  --run-name fable5-main-v1 --model fable5 --tasks pd codec \
  --conditions baseline_copy fdt_copy grounded_copy grounded_copy_tool cdt_copy generic_copy grounded_unrelated \
  --pairs-per-cell 30 --concurrency 8 --max-tokens 4096 --effort medium \
  --formal-trials 1000 --epsilon 0.25 --max-simulations 2

uv run python experiments/logical_copy_coordination/run_experiment.py \
  --run-name fable5-private-context-v1 --model fable5 --tasks pd codec \
  --conditions baseline_copy grounded_copy grounded_copy_tool \
  --pairs-per-cell 20 --private-context --concurrency 8 --max-tokens 4096 \
  --effort medium --formal-trials 1000 --epsilon 0.25 --max-simulations 2

uv run python experiments/logical_copy_coordination/run_experiment.py \
  --run-name haiku45-temp0-v1 --model haiku45 --tasks pd codec \
  --conditions baseline_copy grounded_copy grounded_copy_tool \
  --pairs-per-cell 20 --temperature 0 --concurrency 8 --max-tokens 4096 \
  --formal-trials 1000 --epsilon 0.25 --max-simulations 2

uv run python experiments/logical_copy_coordination/run_experiment.py \
  --run-name haiku45-temp1-v1 --model haiku45 --tasks pd codec \
  --conditions baseline_copy grounded_copy grounded_copy_tool \
  --pairs-per-cell 20 --temperature 1 --seed 20260812 --concurrency 8 --max-tokens 4096 \
  --formal-trials 1000 --epsilon 0.25 --max-simulations 2

uv run python experiments/logical_copy_coordination/analyze.py \
  --runs experiments/logical_copy_coordination/results/fable5-main-v1 \
         experiments/logical_copy_coordination/results/fable5-private-context-v1 \
         experiments/logical_copy_coordination/results/haiku45-temp0-v1 \
         experiments/logical_copy_coordination/results/haiku45-temp1-v1 \
  --output-dir experiments/logical_copy_coordination/analysis/final-v3 \
  --bootstrap-samples 10000

Local logs and complete per-agent workspace artifacts live under experiments/logical_copy_coordination/results/<run-name>/. No Weights & Biases runs were used.

Related work

• Oesterheld, Robust program equilibrium: https://doi.org/10.1007/s11238-018-9679-3
• Meulemans et al., A game theory for foundation models: https://arxiv.org/abs/2608.03958
• Motwani et al., Secret Collusion among AI Agents: https://proceedings.neurips.cc/paper_files/paper/2024/hash/861f7dad098aec1c3560fb7add468d41-Abstract-Conference.html
• Oesterheld et al., Newcomb-like decision-theory dataset: https://arxiv.org/abs/2411.10588