One queue, one ledger: coupled toy agents for protocol allocation and repair

Research experiment, 25 September 2026. All identities, utilities, scientific labels and dependency graphs are artificial. This is an inspectable allocation and accounting simulation. It is not a deployed review service or an empirical behavioral model.

Why join the models

A clone-invariant lottery can feed a clone-biased completion population. A locally valid invitation can consume the time required to repair yesterday’s work. A completed repair check can still lack authority to renew a reliance decision. The first dossier showed these failures separately. This experiment makes them compete for the same finite person-hour ledger so their interaction is visible. It responds directly to the red team’s request for a coupled pipeline and retains negative outcomes as results rather than forcing every request into success.

The interface remains offer, check, rely, amend. Internally, a synthetic request passes through selection, atomic reservation, invitations, utility-driven responses, reported completion, a separate scoped reliance decision and a later repair obligation. No new participant-facing verb is required. These are simulation records; they are not authenticated protocol messages and do not constitute conformance to the adjacent toy_agents runtime.

State, people and decisions

Each independent replicate contains 12 requests in one epoch. Three actual people A, B and C can perform the toy review skill; one steward performs intake, reservation triage, invitations and reliance administration. Every person has a fixed capacity of 3, 8 or 20 model hours. Repair requires A’s calibration skill, C’s inference skill and steward time. Qualification is a fixed fixture predicate; there is no learned qualification model. Money, utility and hours are different units.

A normally has one declared group with 1, 10 or 100 interchangeable labels. Cloning never adds a person or an hour. A deliberately incorrect-control scenario presents A1 and A2 as different groups while debiting both to actual person A. The simulated evaluator knows that mapping; a real operator generally does not. This distinction lets the negative control measure false independence without pretending that the router has solved identity discovery.

For each request, selection draws a two-group panel using either:

• group first: a uniform pair of declared groups;
• naive multiplicity: a uniform representative panel, whose group-pair weight is the product of its groups’ representative counts.

In fixed-panel retry, the first selected pair is retained for all attempts. In reroll retry, each attempt draws again. Both have a maximum of three attempts. A new selection can fail capacity reservation; it still consumes triage effort. No finite-capacity feasibility claim is made by the initial lottery.

Atomic capacity and actual work

The ledger tracks capacity $C_p$, spent work $U_p$ and held work $H_p$ for each person. A demand vector $d$ is reserved only if

$$U_p+H_p+d_p\le C_p\quad\text{for every person }p.$$

The operation is all-or-none: failure changes no hold. A token can be settled once, charging actual work $0\le x_p\le d_p$ and releasing the unused reservation. Lanes share the same capacity, so repair cannot spend an hour already consumed by checking. This is a sequential in-memory atomicity model, not a concurrent database transaction or a guarantee about unrecorded external commitments.

Every admitted request charges $.05$ steward hours for intake. Each attempted reservation charges $.02$ for triage. A reserved invitation holds one hour per selected actor and $.1$ steward hours. Actual actor work is one hour for honest work, $.25$ for shallow work and $.05$ for refusal/response. A reliance decision costs another $.1$ steward hours. Each repair requires $.5$ A hours, $.5$ C hours and $.1$ steward hours. If hidden labels share person A, their demands are summed before admission. Costs are arbitrary declared constants, not measured labor.

A reserve variant holds one A hour and one C hour during initial checking, then releases the holds immediately before repair. It can reduce new completions to preserve some repair opportunity. It does not guarantee repair: steward capacity and the number of affected uses can still bind. It also cannot repair an undiscovered dependency.

Utility-policy toy agents

Actors choose among honest, shallow and refuse using the largest hypothetical utility, with ties preferring refusal, then honest work. A perceived audit chance $q=.2$, loss $F=2$, false detection $\alpha=.02$, detection $\beta=.8$ and reward $R=.8$ give

$$U_H=R-c_p(j,U_p)-q\alpha F,\quad U_S=R-q\beta F,\quad U_R=u_p+\xi.$$

The outside-option shock $\xi$ is uniform on $[0,.25]$. Base effort costs are $.12,.44,.23$ for A, B and C, with request-type increments $.04(j\bmod3)$ and fatigue increment $.012U_p$. Base outside utilities are $.05,.4,.12$. These quantities are synthetic utility units; they do not equal hours or money. The monitoring probability is an exogenous belief. No audit is delivered or funded by this coupled experiment; the separate hard-budget audit model examines that missing commitment assumption.

An honest actor reports completion with probability $.95$; shallow effort reports completion with probability $.98$. Refusal never completes. A panel completes only when both actors report completion. This deliberately distinguishes a finished assertion from adequate actual effort: the evaluator records whether both worked honestly, but the decision actor does not receive that oracle. There is no claim that a shallow check is scientifically valid or that a protocol record discovers the lie. Actual scientific correctness is not simulated.

Independent invitation shocks permit a refusal to change on retry. Fatigue and capacity make repeated outcomes endogenous, so the IID geometric cost formulas from the simpler mathematical model do not apply here. The executable event trace supplies observed cost and unresolved counts directly.

Completion, scoped reliance and amendment remain different

After a reported completion, the steward either relies on every completion or, in a negative-control condition, only on panels including A. Reliance records name an exact toy target such as claim-5:v1, the scope toy numerical decision only, the synthetic decision actor and the declared groups. These labels are simulation identifiers, not cryptographic content addresses or human approvals. Completion by itself does not issue a reliance record.

All 12 toy claims lie in a true dependency chain. The declared graph omits every fourth edge. A material amendment to claim 0 therefore truly affects the chain but is discoverable only through the initial declared segment. This is a known simulation oracle for measuring missed awareness, not a proposed completeness algorithm. A real deployment needs declared evidence and admits that some edges may remain unknown.

Noticed affected receipts become unresolved_material_change. Repair can be completed if the shared ledger has capacity. Repair never renews reliance automatically: the original receipt stays unresolved until a separately authorized new decision, which this experiment intentionally does not implement. Undiscovered affected receipts stay locally as_recorded; the evaluator counts them as unnoticed affected uses. The contrast exposes the gap between a quiet local view and scientific currentness.

A finite exact benchmark before stochastic policy sweeps

For a one-request, nonbinding-capacity special case, set actor completion probabilities to $P_A=1,P_B=1/4,P_C=1$, independent across attempts. A uniform group-pair draw gives panel completion probabilities $1/4,1,1/4$. With three attempts, exact enumeration yields:

For fixed panels, average each panel’s finite geometric paths; for rerolls, first mix the per-attempt probabilities. The policies are different even with no strategic learning and no resource bottleneck. This oracle is exact Fraction arithmetic and has tests against the listed rational values. An all-complete special case is also checked against the full runtime. These special cases do not establish exact expectations for the fatigue-dependent multi-request model.

Results from the coupled sweep

The experiment runs 74 scenarios with 100 independent replicate seeds each: 7,400 runs and 88,800 offered toy requests. The main factorial varies lottery, retry rule, label multiplicity, capacity and protected repair reserve. Two extra scenarios use hidden control and selective reliance. The following values are means or pooled descriptive shares from the specified synthetic fixture, not forecasts of human behavior. Each row here uses 100 labels, no repair reserve.

Group-first runs with 1 and 100 labels are identical event for event under a shared seed, including resource debits and repair outcomes. That stronger fixture property follows because multiplicity changes neither the lottery nor the agent behavior or person map. Thirty seeds test this full-trace invariance. Changing actual capacity, qualification or true control would change the premises.

The naive rule’s lower work expenditure in scarce conditions is not efficiency: it overloads the commonly selected A and leaves other capacity unusable for those panels. It completes fewer requests and leaves more unresolved. In plentiful conditions, naive selection can complete more reported panels because A more often accepts; that is still a heavily concentrated population. Completion counts alone would reward a different objective than representative opportunity.

Rerolling increases completion in these examples, but it also shifts the completed population. With plentiful capacity, group-first offers put A in roughly two thirds of initial panels while reroll completions include A about $.814$ of the time. A separate reliance filter can raise the share to one without changing any completed check. Neither outcome violates the fixed-set group theorem; each is a new selection stage.

Repair creates another stark tradeoff. In the plentiful group-first/reroll case, an average 3.48 known affected uses receive repair checks, while 7.07 affected uses remain unnoticed because of missing declared dependencies. In the scarce case, no repairs complete and about two noticed uses per run await capacity. Even completed repair checks do not clear the reliance state. More available hours cannot reconstruct the hidden chain edge by itself.

The full output contains 888 metric rows with pointwise normal 95% intervals on replicate means. Requests inside one replicate compete for the same capacity, so they are not treated as independent Bernoulli samples. Pooled stage shares are reported descriptively without a misleading binomial interval. Common random seeds across policies permit paired future analyses, but no paired confidence claim is made here. Intervals address finite simulation variation, not utility misspecification, behavioral uncertainty or human adoption.

Accounting identities and falsifiable boundaries

Every trace separates request-level stages from attempt and actor counts:

$$\mathrm{offered\ requests}=\mathrm{completed\ requests}+\mathrm{unresolved\ requests},$$

$$\mathrm{invited\ actors}=\mathrm{accepted\ actors}+\mathrm{refusing\ actors},$$

$$\mathrm{affected\ relied\ uses}=\mathrm{unnoticed}+\mathrm{pending\ repair}+\mathrm{repair\ checked}.$$

Intake or capacity rejection remains an unresolved request in this experiment; there is no terminal withdrawal state. A request may receive several refused or failed attempts, so refusal counts are not substituted for unresolved counts. Completed-but-not-relied requests are explicit. Actual hours and holds are exposed for each person and lane. At termination, every hold is settled or released.

Fourteen tests check atomic failure, settlement bounds, double-settlement rejection, shared lanes, independent panel enumeration, clone-invariant complete traces, policy/capacity/control accounting, fixed-panel preservation, selective reliance, hidden control sharing actual capacity and the absence of automatic renewal after repair. Every positive behavioral interpretation remains conditional on synthetic utilities and the chosen order of operations.

What each community can change

A physics/astronomy group can replace the toy chain with a calibration-to-inference graph and expose how scarce instrument experts constrain correction. A biology group can make completion delay and evidence accessibility depend on replication or data-use authority without treating refusal as a negative scientific verdict. Economics and social-science groups can replace the synthetic utility function, compare exclusion and welfare objectives, and investigate whether repeated invitations impose unequal costs. Legal scholars can add deadline and recusal constraints; the current hour ledger makes no claim of legal scheduling adequacy.

The human entry point is to inspect one trace, identify one assumption that would be wrong in the participant’s field, and propose a counterexample. The agent entry point is to substitute a behavior function or graph and run the accounting tests immediately. Neither path requires accepting a universal trust score.

Reproduce

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 Python is sufficient. results/summary.json contains settings, seeds, source/table hashes and aggregate results. Five full JSON traces expose individual actions and ledger balances. The CSV is a plotting surface, not an empirical dataset. The model uses no credentials, network access, real research participants or private data. Primary mathematical context and source verification are in math-foundations.md and its models/math/source-map.json; the coupled simulation is a new local fixture responding to reviews/red-math/report.md.