Local-rule attention ecology

A deliberately small hypothetical multi-period agent model. Agents are scripted resource/skill/quality profiles, not LLMs or sampled humans. It studies whether bounded credit and exploration are enough to prevent attention concentration and repair overload. The answer is conditional and often negative.

From papers/07-release-packet:

python3 -m unittest discover -s models/ecology/tests -v
python3 -m models.ecology.run
python3 -m models.ecology.correction_game

Python 3.10+ standard library only. Nine perturbation regimes × three policies × eight paired seeds × 80 periods produce 216 rows in results/sweep.csv. Every policy receives the same exogenous offers, defect states and common blind periods for a given seed/regime; policy-dependent choices and observations diverge. summary.json records source/test hashes, Python version, sweep hash, means and seed ranges, and paired differences against uniform group routing. These ranges are descriptive simulation variation, not confidence intervals. Full default bounded-policy history is retained in example-trajectory.json.

Only a few local rules

Each group emits synthetic jobs according to fixed arrival/volume, skill and defect parameters. One large team has more volume and capacity, but remains one accountable group. The first group also starts with a visibility advantage.
Route an oldest queued offer within a selected active group. Prestige uses cumulative completed attention plus pending-volume multiplication. Bounded uses decaying capped verified-check/correction credit, with an exploration mixture. Uniform chooses among active groups uniformly.
Select an independent qualified reviewer proportional to available capacity. Intake costs two tokens, check ten, independent audit five. Review quality is a stipulated diligence parameter; blind periods create common-mode errors.
Flagged defects or audit-discovered misses create correction jobs. Corrections consume twenty tokens (scaled in stress tests), oldest first, from the same independent qualified experts before new checks. Completed corrections and favorable audits supply capped credit.

Budgets reset each period; queue and correction debts persist. The simulator records every offered job, completed job, unresolved offer and correction debt, all stage labor, refused-capacity attempts and audits unavailable for lack of capacity. Invitation/refusal counts are attempts, so a queued offer may count again next period. An inability to find even two intake tokens is counted as a capacity refusal without a paid invitation. Routing CPU cost and author labor are excluded, not claimed free in a real institution.

The capacity/skill map is heterogeneous. Low-resource actors can perform one check but cannot individually absorb a twenty-token repair. Tasks are indivisible within a period; no fractional progress or pooled repair team is modeled. Therefore aggregate spare capacity does not imply feasible specialized repair. Reviewer and auditor control groups differ from the author, and an auditor differs from the reviewer; actual independence is a trusted fixture assertion.

What the metrics mean

correct_checks_per_offer: correct toy judgments divided by all offers, including work not reached. It is not scientific value.
accuracy_among_completed: conditional judgment accuracy; can be high while coverage and defect discovery are poor.
defect_detection_fraction: defects found by a check or independent audit divided by all actual synthetic offered defects, including unresolved work.
attention_hhi: sum of squared shares of completed author attention. Unequal offered volume makes this a descriptive concentration measure, not a fairness verdict. Compare per-group completion coverage too.
minimum_group_coverage: the lowest completion/offered ratio, not minimum attention share. Reducing concentration can worsen this metric.
repair_backlog: known pending correction jobs. It excludes undiscovered defects; a quiet correction queue can be a blindness signal rather than success.
labor: intake + checks + audits + repairs over all agents, constrained by the same per-period individual capacities. No token is spent twice.

Deliberate limitations and negative controls

Audit truth is a fixture oracle, conditional on access and capacity; it is not a real scientific verification mechanism. A biased-verification regime gives large team offers much greater audit access. Blind checks always return support, so high common-mode blindness can produce low observed repair debt while missing most defects. A unit test sets all claims defective, full blindness and no audits: completed accuracy is zero while the repair backlog is also zero.

The model has no endogenous participation, strategic effort, lying about control, staged repair credit, institutional enforcement or equilibrium. “Cooperation” here means accomplished check/repair tasks under scripted willingness; it is not an emergent motive. For strategic honest/shallow/refuse comparisons use the separate models/coupled study. Recognition and dominance are hypotheses represented by allocation feedback, not asserted explanations of human scientific conduct.

There is no claim that bounded routing beats uniform routing. The overload and expensive-repair regimes expose tradeoffs and failures. Run-level differences are exploratory, not optimized or preregistered. All policies and all regimes are kept. See ../../manuscripts/local-rules.md for exact equations and bounded conclusions.

Correction-credit incentive bridge

The separate correction_game.py compares clean work with deliberately manufactured and then repaired work using the same bounded score/attention update. Three credit policies, six parameter cases and eight additional tests preserve a unilateral and coalition counterexample, a no-allocation-credit baseline, cost internalization, credible intent-detection assumptions and scarce independent repair capacity. results/correction-game.json records exact payoffs and source/test hashes. This is a two-action one-step game, not an equilibrium claim about the ecology or humans. The ordinary ecology’s original no-audit credit vulnerability remains unchanged. The full tests command now runs eighteen methods.