siddharthmb/2026.RA.Auction-InstructedRing-Opus48

<!-- generated by experiments/rationalagents/auction/packagehfauction.py — every number below is derived from analysis/summary.json and campaign/campaignmanifest.json at build time, never hardcoded. -->

2026.RA.Auction-InstructedRing-Opus48

Rollouts of the INSTRUCTED-RING auction campaign (interlens arena): five bidders, four of them told in as many words that they have agreed to coordinate bidding and divide the lots. The instruction scripts no division, no price and no punishment scheme, because each is a quantity being measured. Cells form a three-rung side-payment ladder under one unchanged instruction — no transfer field, an UNCONDITIONAL transfer, then an ESCROWED transfer contingent on the recipient taking no lot — against a matched no-ring control, plus an all-inclusive ring and a Dutch-clock pair. The question is whether an enforceable side payment gets used, and whether anything moves if it does. This population is quarantined from the coordination-neutral campaign and may never be pooled with it.

419 episodes · 1,676 stages · 22,568 turns · 15 cells · model `anthropic:claude-opus-4-8` · $688.52 of API spend. (Summed from each packaged run's own usage.json. The launcher's spend_ledger reads $336.33; the $352.19 difference is spend from runs that executed outside the launcher and so never entered its budgeted attempts.)

Design and preregistration: https://github.com/Sid-MB/iimats/blob/main/experiments/rationalagents/auction/docs/ring-campaign-design.md. Research note: https://github.com/Sid-MB/iimats/blob/main/experiments/rationalagents/research-notes/0066-instructed-ring-campaign.md.

Cells

all_llm seats every chair with the model; all_rational best-responds from its own information only; all_oracle best-responds knowing every seat's private values; one_* arms seat one computable agent among LLMs. The computable arms are model-free and are the reference every LLM number is stated against.

Validity gates

This campaign did not run the persona controls (G2/X2), which belong to the committed matrix; its gated quantities are protocol validity and computable-seat integrity.

The collusion battery (bid suppression against each mechanism's own equilibrium benchmark)

Positive = bidding below the benchmark (the direction a ring would move); negative = bidding above it.

Mutual information on the message channel and Porter–Zona losing-bid rationality are in analysis/summary.json under battery. The transcript classifier (precedence 4) was not run: not computed by this analyzer, by design: measure 4 needs paid judge calls and this pass is free. Run auction/classifytranscripts.py (calibrate, then run, then report) and read its dissociation table, which carries each cell's suppression beside each cell's talk rate. For pkgbopus48v1 that lane is already run: see auction/docs/transcript-classifier.md and campaigns/pkgbopus48v1/analysistranscript_measure4/.

Provenance

1 code fingerprint(s), 0 transition(s); analysis/summary.json → code_vintage carries each transition's reason and the behavioral-equivalence verification. Per-run vintage.json sidecars are shipped under runs/. Excluded episodes: 0 (recorded in runs/<run>/exclusions.json with reason and replacement; excluded rows are not in the tables here).

Human-readable per-episode transcripts are kept on the cluster rather than published here: /juice2/scr2/siddharth/ii_mats/rational_agents/auction/pkgb_opus48_v1__*__{free,llm}/transcripts/. Per-turn rendered prompts (view) are likewise omitted; they are reproducible from the shipped banks.

Layout

episodes.jsonl        # `episodes` config — one row per episode
stages.jsonl          # `stages` config   — one row per played stage
turns.jsonl           # `turns` config    — one row per turn
analysis/             # the campaign analysis verbatim (summary.json, analysis.md, row CSVs)
campaign/             # campaign_manifest.json (allocations, invocations, vintages, spend ledger)
runs/<run>/           # each run's manifest.json, usage.json, vintage.json, exclusions.json
banks/<bank>/         # the frozen instance banks the episodes were played on
artifact_manifest.json  # sha256 of every shipped file + the build's own counts

Every row carries experiment_name, cell_id, arm, run_name and in_campaign_analysis, so a later campaign with the same columns appends to this dataset rather than forking a new one.

Method & citations

• —[vickrey1961] Vickrey, W. (1961). Counterspeculation, Auctions, and Competitive Sealed Tenders. The Journal of Finance 16(1):8-37. <https://www.jstor.org/stable/2977633>
• —[riley_samuelson1981] Riley, J. G. & Samuelson, W. F. (1981). Optimal Auctions. The American Economic Review 71(3):381-392. <https://www.jstor.org/stable/1802786>
• —[milgrom_weber1982] Milgrom, P. R. & Weber, R. J. (1982). A Theory of Auctions and Competitive Bidding. Econometrica 50(5):1089-1122. <https://doi.org/10.2307/1911865>
• —[robinson1985] Robinson, M. S. (1985). Collusion and the Choice of Auction. The RAND Journal of Economics 16(1):141-145. <https://www.jstor.org/stable/2555596>
• —[porter_zona1993] Porter, R. H. & Zona, J. D. (1993). Detection of Bid Rigging in Procurement Auctions. Journal of Political Economy 101(3):518-538. <https://doi.org/10.1086/261885>
• —[kagel1995] Kagel, J. H. (1995). Auctions: A Survey of Experimental Research. In Kagel, J. H. & Roth, A. E. (eds.), The Handbook of Experimental Economics, pp. 501-585. Princeton University Press. <https://press.princeton.edu/books/hardcover/9780691042909/the-handbook-of-experimental-economics>
• —[calvano2020] Calvano, E., Calzolari, G., Denicolo, V. & Pastorello, S. (2020). Artificial Intelligence, Algorithmic Pricing, and Collusion. The American Economic Review 110(10):3267-3297. <https://doi.org/10.1257/aer.20190623>
• —[fish2024] Fish, S., Gonczarowski, Y. A. & Shorrer, R. I. (2024). Algorithmic Collusion by Large Language Models. arXiv:2404.00806. <https://arxiv.org/abs/2404.00806>