Coupled allocation, toy behavior, capacity and repair

Read coupled-agents.md for the model, exact special cases and conditional interpretation.

python3 -m unittest discover -s papers/07-release-packet/models/coupled -p 'test_*.py' -v
python3 papers/07-release-packet/models/coupled/run_experiments.py

Standard library only. 14 tests; 74 scenarios x 100 replicate seeds; 88,800 synthetic requests; 888 metric rows with replicate-mean nominal intervals. Five full event traces show scarce capacity, plentiful capacity, naive multiplicity, hidden control and selective reliance. No claim of real efficacy or protocol conformance is made.

The shared person ledger covers intake, triage, invitation/check work, scoped reliance and repair. It accounts for holds and actual debits atomically. Skills and actual-person mappings are fixed toy facts; the operator does not thereby learn real identity. Repair checks do not automatically renew reliance. Missing dependencies remain invisible to the local reliance view and visible to the simulation evaluator’s separate oracle.

Behavior injection: run(Config(...), seed=1, trace=True, behavior=fn) accepts a function (person, request_index, prior_used_hours, rng) returning (action, reported_completed, actual_hours, utility_dict). It must honor actual_hours <= 1 per invitation and use honest, shallow or refuse actions. The default model gives actors a perceived audit probability; it does not deliver or fund audits. Use the separate math model for audit budget experiments.

Returned simulation receipts are plain records, not signed protocol messages. The adjacent runtime has stronger target and authority semantics. Connecting the behavior model to that runtime requires an explicit adapter and new validation.