The porch that ate Halloween.
One porch advertises eight candies. It can serve one goblin. The sign-rewarded crowd moves in anyway. Reward the whole neighborhood, and the bags fill—but somebody may spend every Halloween at the two-candy house.
Evidence boundary: this instrument independently checks the finite six-porch model. The learning averages quoted later came from an earlier in-chat run report; its training traces and source are not included in this release. The workbench is a new, transparent reconstruction—not a playback of those runs.
The problem isn't always dishonesty.
Each goblin visits exactly one porch. The advertised candy number is real. Capacity is real too. Overcrowded porches use a fair lottery: one winner receives the whole listed amount, while the others leave empty-handed.
When individuals seek their own candy, some stable distributions leave useful capacity unused. When a group maximizes all candy, it can split across porches perfectly. But total productivity and fair individual experiences are different measurements.
Possible crowd allocations. A unique 43-candy optimum, and exactly two strict-unilateral-stable personal-reward allocations: 38 and 35.
Set the reward counter. Watch where they go.
Try a crowded night, change the objective, then let the goblins make only individually improving switches. That is best-response search, not the earlier reinforcement-learning experiment. You can also assign each goblin a porch yourself and sample the lottery.
The six porches
The bar shows successful visits in green and visitors beyond capacity in red. Every winning visitor receives that porch's full payout.
No lottery sampled yet. Group candy is exact for the current crowd allocation; only the winners' identities vary.
Ten named goblins
Move a goblin manually and watch the real consequences. Expected candy already averages over the fair lottery; a sampled night marks the actual winner or empty bag.
What is rewarded?
The sign counter sees ten claims of eight candies, even though capacity admits one visitor.
Latest decision trace
- No changes yet. Select a preset or let the goblins respond.
Different question: who gets the two-candy job?
The rotation is an engineered follow-up, not an emergent goblin discovery. On the optimal 43-candy allocation it equalizes expected returns without changing capacity use. It does not guarantee equal bags on any one night.
The training results that started this Halloween mess
The earlier conversation reported ten independent small learners per group, 64 training seeds per reward condition, 12,000 training rounds, and 2,000 evaluation rounds. These means are quoted from that report; this release does not contain its learning source, raw traces or a matching rerun. Do not equate them with the best-response results in the interactive workbench above.
| Reward signal | Reported candy / round | Reported empty bags | Interpretation |
|---|---|---|---|
| Sign value | 8.21 / 43 | 89.54% | All want the 8; too few winners |
| Individual candy | 32.62 / 43 | 37.44% | Honest reward; crowd still competes |
| Shared group candy | 42.68 / 43 | 0.78% | High output; unequal fixed roles can emerge |
| Marginal contribution | 42.41 / 43 | 1.55% | Near-optimal but not better than shared here |
Reproduced separately: exhaustive counting of 3,003 visitor-count arrangements in this exact six-porch fixture gives a 43-candy maximum at 1 / 1 / 2 / 2 / 3 / 1. The only personal-candy unilateral equilibria total 38 and 35. Inspect the public reference checker and methods. Those discrete results are reproducible; the preceding learning averages remain unverified in this release.
What survives when the rules change?
- Try abundant supplies: the ten goblins can all collect eight candies at one porch. Concentration now produces 80 candies; a detector that always penalizes crowding would be wrong.
- Try personal reward: best-response switches can stop at 35 or 38 even though 43 is attainable. A stable choice is not necessarily an efficient outcome.
- Try shared reward: the total aligns with the 43-candy target, while the fixed-role payoff distribution remains unequal.
- Try rotating identities: expected fairness can improve without reducing output here, but this toy does not model travel, specialization costs, preferences, communication, or real agents.
Every number generated by the instrument follows published deterministic formulas; the sampled lottery uses a local seed. No outside LLMs, browsers, shoppers, purchases, or personal information are involved.
Methods, full limits & evidence receipt ↗