01 / Watch, don't imagine
Watch the learning. Then inspect the result.
Two learners negotiate. Eight bystanders keep their roles. Play actual recorded training, or inspect the settled evaluation policy. Changing the rules selects another completed run; it does not retrain the agents.
12,000 training rounds each
Latest recorded receipts
| Round | Offer | Decision | Delivered | Bags P / R |
|---|
Statistics are exact expectations of the frozen policies, averaged over seeds. Intervals resample seeds, not individual nights. Eight bystanders are fixed in this bargaining stage; the separate baseline has ten active learners.
02 / The result that looked too good
The generous deal has a loose thread.
Veto + three candies elsewhere + unenforced promises, with personal rewards. The formerly disadvantaged agent looks better off—until we inspect unilateral incentives.
Keep the receiver fixed. Change only the proposer.
These are exact all-seed counterfactual means, not a prediction of the next thirty nights. The receiver would have opportunities to relearn later.
What the probe actually establishes
Inspect the policy, not its reputation. Enumerate every legal offer and hidden honor/renege action against the saved receiver.
Keep the comparison honest. Replace one policy only. Use the same game, outside option, and receiver responses.
Report the contradiction. 58 of 64 saved proposer policies can gain by deviating. The average gain is 3.47 candies.
A pleasant distribution can be an unstable learning snapshot. It is not evidence that the goblins invented a durable institution.
03 / Your contract, not an agent trace
Try the witch’s accounting.
Use the currently selected institution and outside option. Your choices below are counterfactual inputs, clearly separate from the saved runs.
Escrow overrides an attempt to renege. The no-refusal institution overrides a rejection. Outside candy is counted separately.
Why exit alone cannot force a transfer here
With independent porches, an offer of zero leaves the proposer with eight whether the receiver stays or leaves. A positive transfer can only lower that payoff.
The veto changes this: rejection now costs the proposer its own payout. A credible alternative gives the receiver a reason to reject. Escrow makes an accepted promise executable.
The learners choose numerical actions within this game. They did not invent the veto, the outside source, escrow, or the available promise-breaking action.
04 / No selected winners
All 24 conditions.
Each row includes all 64 primary seeds. “Deviation gain” is the proposer’s possible reward increase against the frozen receiver, in that row’s reward units.
| Refusal | Contract | Reward | Outside | Proposer | Receiver | Refusal | Deviation gain |
|---|
05 / Back to the original street
The unequal roles really reappeared.
A fresh ten-learner baseline, now with all actions, rewards, seeds and checkpoints preserved. This is not a reconstruction of the earlier missing training program.
Held-out group candy, out of 43
64 seeds per reward. In every shared-reward and marginal-reward seed, the group reached 43, while individual role means still ranged from two to eight.
Three distinctions that survived the run
Production is not distribution. The group can fill every porch while keeping one individual in the two-candy role.
A high score is not equilibrium. The new personal-reward baseline averages 40.77, but only 4/64 greedy argmax profiles pass the auxiliary Nash check.
A better opportunity is not compensation. The outside-three exit outcome is 41 neighborhood candies plus three from elsewhere—not a transfer from the proposer.
06 / Take the receipts with you
Methods, source, and complete records.
Read the plain-language report, keep an offline copy, or inspect the full run. The browser view is only a window into the evidence.
Study CANDY-REVOLT-2026-10-31-001 · Science v0.1.0 · Public presentation v0.1.2. This release incorporates the replay repair and adds the summary and navigation. It does not introduce a new research treatment or change the findings.
Start with the original trick-or-treating street ↗ · Release receipt · Checksums · All lab experiments ↗
07 / Open the notebook
What this does—and does not—show.
Receipts, not a simulated success story
The protocol was hashed locally before the primary run. That is not external preregistration. An author-written, separate audit reconstructs every training bag and final action value; it is not external peer review.
Primary: 24 conditions × 64 seeds × 12,000 rounds. Sensitivity: learning rates 0.02 and 0.10, 32 seeds per condition each. All rates and every seed are retained.
No API credits were spent. No external model or human was tested. This is the public v0.1.2 exhibit. The scientific run is unchanged; independent inspection is pending.
What the agents did not invent
There is no language negotiation, coalition formation, reputation ledger, time-horizon planning, or new social institution. The bargaining stage has two independent tabular learners and eight fixed bystanders. Mechanisms are experimentally supplied.
What the thirty-night shock did not show
A ten-night increase in the outside option did not increase refusal rates in the private-reward escrow exit or veto cells relative to the paired control. The extra candy came from the same exploratory refusals receiving outside pay. Rapid, durable adaptation is not established.
Why a zero deviation gain is not a universal stability certificate
The probe checks unilateral changes with the other policy held fixed. It does not prove subgame perfection, reliable learning at unvisited offers, long-run mutual adaptation, or robustness to a new partner.
Relation to existing research
Learning and asymmetric equilibria in bargaining are established subjects. This is a particular executed toy study, not the discovery of bargaining theory. Its tabular learner differs from the FTRL analysis in Kamp, Liebman & Fish (2025). It also has no inequity-aversion reward such as those studied by Hughes et al. (2018).
What changed in this replay repair?
The original player displayed only settled evaluation rounds. Every one of its 1,536 thirty-round sequences repeated a single action/decision/bag record, so Play looked inert. This repair adds three fixed, consecutive training windows for every condition and seed, labels the data source, shows accumulated candy within the selected window, and displays explicit playback progress and receipts. The scientific results, frozen protocol and original evaluation records are unchanged. Training variability includes exploration, not just learning progress. No new agents were run.
Full reproduction bundle
The downloadable research ZIP includes run_experiment.py, audit_records.py, oracle_controls.py, protocol.json, README.md, METHODS.md, all compressed raw arrays, seed tables, test receipts, and SHA-256 checksums. This HTML alone is self-contained for exploration, not the full evidence archive.