lewtun/gitswarm-ablation-replication

GitSwarm Appendix C replication — Hidden Rulebook (persistence)

Partial replication of GitSwarm: Decentralized Compounding Inference (arXiv 2610.04862), specifically the Appendix C.1 "Hidden Rulebook" controlled test of the persistence substrate property. The Maze (branchability, C.2) and Retracted Evidence (attribution, C.3) modules are implemented and offline-validated in code/ but were not run (the sweeps were scoped out by the task owner).

No official code was released with the paper; the harness was built from the verbatim protocol in Appendix C.

What the test measures

A rulebook of records with opaque IDs: facts hold integers, rules apply add/multiply mod 997 to one earlier record. Each of P targets heads a chain of D rules ending in one fact, so Q = P·(D+1) records are relevant; the rest are disconnected distractors. Workers see only target IDs and must discover chains one record at a time (inspect reveals a rule's operation and its predecessor's ID, never its contents).

Treatments (aggregate budgets: shared ≈ Q, independents ≈ 4Q/3, so the independents have more total queries):

• shared — sequential workers, each reads previous workers' directories (read-only), k=10 queries/worker, B ≈ Q/10
• indep_B2 — B=2 isolated workers, k=⌈2Q/3⌉
• indep_k4 — k=4 isolated workers, B=⌈Q/3⌉

Metric: relevant-record coverage = distinct inspected records on target chains ÷ Q, oracle-computed from the union of all workers' observations (submission correctness not required, as in the paper).

Results (moonshotai/Kimi-K3 via HF Inference Providers, 3 seeds)

Coverage, mean over seeds (n in parens; error runs excluded):

Condition	shared	indep_B2	indep_k4
P=2, D=4 (Q=10)	1.00 (3)	0.80 (2)	0.60 (3)
P=4, D=4 (Q=20)	1.00 (3)	0.95 (2)	0.33 (3)
P=8, D=4 (Q=40)	0.75 (3)	0.68 (2)	0.06 (2)
P=16, D=4 (Q=80)	0.82 (2)	0.78 (2)	0.11 (2)
P=4, D=2 (Q=12)	0.78 (3)	0.92 (3)	0.17 (3)
P=4, D=8 (Q=36)	1.00 (3)	0.75 (3)	0.18 (2)

Sweep points are our choice; the paper's Table 4 body is not recoverable from its public rendering. Paper prose targets: independents flat at ~70–75% (B=2) and ~20–23% (k=4), dropping with D, while shared memory stays high.

Findings vs the paper

1. The directional claim replicates. Shared persistent memory achieves 0.75–1.0 coverage on an aggregate budget of only ~Q, beating both independent baselines that command ~4Q/3 across the P sweep.
2. indep_B2 ≈ paper level but noisier: 0.68–0.95 vs the paper's flat 70–75%. At small P it exceeds the paper's number.
3. indep_k4 declines instead of staying flat: 0.60 at P=2 down to 0.06–0.11 at P=8–16, below the paper's flat ~20–23%. Kimi-K3 workers with a 4-query budget scale worse than the paper's Gemini-3.1-Pro workers.
4. Shared's advantage grows with depth (D=8: shared 1.00 vs B2 0.75), consistent with the paper's qualitative claim.

These are 3-seed results with one model; they support the qualitative claim, not the exact magnitudes. 8 of 54 condition-runs errored on a harness bug (fixed in code/rulebook.py); coverage for those cells is reported over the remaining seeds.

Cost

6,558 LLM calls, 155.4M prompt + 1.25M completion tokens (conversation history is resent every call). 2h29m on cpu-basic for orchestration.

Repo layout

• code/agent.py — worker loop: JSON tool-call protocol over HF Inference Providers (moonshotai/Kimi-K3)
• code/rulebook.py — C.1 generator, treatments, oracle (validated)
• code/maze.py — C.2 generator, three memory treatments, restore validation (offline-validated, sweeps not run)
• code/evidence.py — C.3 archive generator with unique-true-parent invariant, attribution treatments (offline-validated, sweeps not run)
• code/run_rulebook.py, code/run_maze.py, code/run_evidence.py — sweep drivers
• results/rulebook/rulebook.jsonl — all 54 condition-runs, raw
• results/rulebook/summary.json — aggregates above
• runs/rulebook/ — per-run artifacts (instances, observations.jsonl)

Run it

pip install huggingface_hub==2.2.0
export HF_TOKEN=...            # HF Inference Providers token
python code/run_rulebook.py --out results/rulebook.jsonl --concurrency 4

Citation

@article{gitswarm2026,
  title={GitSwarm: Decentralized Compounding Inference},
  author={Shah, Vedant and Goyal, Anirudh and Samanta, Ankur and Dahal, Paras
          and Plekhanov, Mikhail and Wu, Carole-Jean and Yih, Scott and
          Munos, Remi and Fergus, Rob and Foerster, Jakob and
          Salakhutdinov, Ruslan and Arora, Sanjeev and Weston, Jason and
          Courville, Aaron},
  journal={arXiv preprint arXiv:2610.04862},
  year={2026}
}