Full report
Acceptance gates versus prices for co-evolving agents: five pre-registered studies in two environments
Vidit Choudhary
Contents
Final report of the gates line, 18 November 2026. Status: single-site, unreviewed, unreplicated. The audit trail, decision log and per-study detail are in PROJECT.md. The frozen protocols are PREREGISTRATION.md (E1–E6), PREREGISTRATION_E8.md, PREREGISTRATION_E9.md and PREREGISTRATION_E10.md.
Abstract
Self-improving agent systems share a propose–evaluate–accept loop. When several such loops share an environment, each accepted change alters the conditions under which the others are evaluated, and changes that are good for their owner can be bad for everyone. We ask whether the acceptance rule alone can correct this, and whether it can do so better than a charge on the externality that an operator could levy instead. Proposals, search budget and environment are identical across arms; only the gate, or the price the agents face, differs. Five pre-registered studies (4 190 confirmatory runs, 20 paired seeds each, held-out evaluation scenarios) were run in two environments.
In a stylized real-time gross settlement network (E1–E6, E8), private gates leave the system 29 % costlier than a symmetric planner. A universalization (Kantian) gate, which reads only the agent's own cost when every agent makes the same change, closes 51 % of that gap. A gate that evaluates joint moves does better still. But a delay surcharge on private agents does better than every gate, with three-parameter and with eight-parameter rule-level policies.
In a shared-codebase environment (E9, E10), harm is unattributable and its carrier can be hidden, so no direct price exists. With identical agents the Kantian gate and a budget-breaking ambient tax are practically equivalent (difference 0.0009, 0.4 % of cost). Every budget-balanced charge either had no effect or was gamed into concealment. When incident exposure is unequal, the ambient tax beats the best gate (by 0.0022 and 0.0060 at two levels of inequality). The gate shows the distortion our derivation predicts: a Kantian point that is Pareto efficient but weights agents by one over their exposure, so it misses the minimum of total cost. Under a decision rule fixed in advance, the claim that gates beat the best available instrument is therefore rejected. A fifth study (E11) re-ran the Kantian gates with "the same change" measured on three other scales. The gate's cost moved by up to 0.02 with the choice of scale, but the ambient tax still beat the best gate over all scales, so the verdict does not depend on units.
What remains is conditional. Where only budget-balanced charges are possible and the harm's carrier can be hidden, an own-cost Kantian gate beats those charges. Its margin shrinks as exposure becomes unequal, and at the strongest inequality tested it is within the equivalence margin. Several registered predictions failed; they are listed in §5.
1. Question
Systems that improve themselves by proposing changes, scoring them and keeping the ones that score well are now common (for example CORAL, Qu et al. 2026; Hyperagents, Zhang et al. 2026; AIDE², Srikanth et al. 2026). They are evaluated as single lineages, or as several agents pursuing one common objective, against fixed benchmarks. In use, several such systems owned by different principals share an environment: banks sharing a payment system, teams sharing a codebase. Then a change that lowers its owner's cost can raise everyone else's, and a loop that accepts on own cost reproduces a social dilemma.
The loop has a natural control point that is neither the proposer nor the environment: the acceptance gate. This work treats the gate as the experimental factor and asks two questions.
- Does the choice of gate change the outcome, at matched search budget?
- Can a gate do better than the textbook alternative, a price on the externality chosen by an operator?
The second question decides whether gates are worth building. It was answered in two environments chosen to make it progressively harder for the price: one where the externality can be priced directly, and one where it cannot be attributed.
2. Related work
A literature review is in LITERATURE_REVIEW.md. It combines 118 logged web searches with a second pass through the arXiv, Crossref and OpenAlex APIs, and a manual Google Scholar pass. It is a broad structured review, not a formal systematic one, and several sources were seen at abstract level only. This section summarises it.
Acceptance gates. In 2026 the acceptance gate of a self-improving loop became a research object in its own right, but only for single agents. SAGE (arXiv:2609.36043) compares a statistical gate with a greedy one at matched proposals and tokens. PACE (arXiv:2606.08106) compares greedy, fixed-sample, false-discovery-rate and anytime-valid gates. So et al. (2026) compare ten gate rules across 411 loops. SAGE and PACE name the multi-agent case as open. When learning rules themselves are under selection, selfish learning is unstable and other-regarding rules are favoured (McAvoy et al. 2022, two-player repeated games). Single-lineage systems such as the Darwin Gödel Machine and Hyperagents compare archive-based acceptance with keeping only one successor. The closest classical analogue is Agogino and Tumer's comparison of local, global and difference evaluation for evolving multiagent policies. It is fully cooperative, and it shapes fitness rather than accepting or rejecting proposals. We did not find a comparison of per-agent gates among co-evolving agents that impose costs on each other, nor one against a price.
Kantian theory. Roemer (2015) shows that additive and multiplicative Kantian equilibria are Pareto efficient, including with heterogeneous agents. Only the simple form, in which everyone takes the same action level, is generally inefficient with heterogeneity. Homo moralis preferences (Alger and Weibull 2013) give the degree-of-morality mixture κ. Partial Kantian equilibria are defined by Kordonis (2020), through imagined groups of co-movers who minimise their mean cost, and by Istrate (2020). Our κ mixes two scores of the agent's own cost instead. Behind a veil of ignorance, with roles assigned at random, universalization implies utilitarian altruism (Laslier 2022). Our κ = 1 gate is the additive form: every agent's policy moves by the same mutation. With unequal exposure, the additive Kantian condition gives b′ᵢ = n·eᵢ·D′. The equal-weight cost optimum gives b′ᵢ = n·ē·D′. So the Kantian point is the welfare optimum with weights proportional to 1/exposure. That is efficient in Pareto's sense, but not the minimum of total cost. This derivation is ours. It is not stated in the sources we found, and E10 tests its prediction. It is consistent with Laslier's result: E10 fixes each agent's exposure rather than randomising it, which is exactly when universalization and total-cost minimisation separate. Universalization has been used as a decision-time prompt for language-model agents (Piatti et al. 2024). We did not find it used as the acceptance rule of a learning loop.
Payment systems. The payment-timing game is a standard example of privately rational delay with collective cost (Bech and Garratt 2003; Galbiati and Soramäki 2011). Learning agents in it have been studied with reinforcement learning (Castro et al. 2025). A single prompted language-model agent has been tested on cash-management decisions (Aldasoro and Desai 2025), and prompted agents have been made to interact in an intraday-liquidity game (Danisman et al. 2026). Pricing beating coordination rules has precedent: a time-rising intraday tariff beats throughput guidelines in most cases (Ota 2011). Late-settlement fees have also been studied as a coordination device (Nellen 2019). Our surcharge result extends that comparison to agent-level gates. Fu et al. (2026) compare stabilising mechanisms for language-model agents in other financial coordination games; there paid commitments were never the best mechanism. Their agents are prompted, not evolved.
Unattributable harm. With output observed only in aggregate, no budget-balanced sharing rule gives efficient incentives, and a budget breaker can (Holmström 1982). The ambient tax on non-point pollution is the corresponding instrument (Segerson 1988). Known objections include collusion, large transfers and information requirements (Millock and Salanié 2005). Laboratory evidence is mixed (Spraggon 2002; Cochard, Willinger and Xepapadeas 2005; Suter et al. 2008). Spraggon found the tax effective for groups but not for individuals. Cochard et al. studied polluters who bear the damage themselves, as our agents bear incidents. There, subjects under the ambient tax cooperated to maximise group payoff, cut input below the optimum, and ended worse off than without regulation. In another experiment, heterogeneity in farm size and location reduced the effectiveness of an ambient tax/subsidy unless peer information was given; an agent-based model calibrated on that experiment exists (Wu 2017). Charging each agent all damage outside its module makes its payoff equal to private benefit minus total damage, a Groves-style transfer. The level-1 identity between the ambient tax and the unilateral welfare gate is therefore an observation, not a result. In multi-agent reinforcement learning, LOPT (Hua et al. 2025) learns a Pigouvian reward shaping toward social welfare where the externality can be attributed. Danassis et al. (2026) use a learning policymaker to set market prices against common-pool externalities, where harvests are observable. Paying on a measured dimension moves effort into unmeasured ones (Holmström and Milgrom 1991; not re-read for this review).
Technical debt. Gavidia-Calderon et al. (2020) model technical debt as a tragedy of the commons. In their simulation, two developers choose between a kludge and a fix, and the only equilibrium is that both kludge. Code review with enough detection restores fixing. Our codebase environment adds continuous shortcut intensity, concealed coupling, hardening, incidents that rise convexly with debt and land in random modules, operator charges, and many evolving agents. Developers describe AI-assisted coding in the same terms: individual productivity gains push costs onto reviewers and maintainers (Baltes et al. 2026).
Every comparison here therefore builds on known pieces. What we add is a matched-budget comparison of per-agent acceptance gates against prices among co-evolving agents, in two environments. That comparison was not found in this review.
3. Method
3.1 Common protocol
Loop. n agents, 40 generations. In each generation each agent proposes four Gaussian mutations of its policy, scores each on 16 fresh development scenarios, and accepts the best if the gate's score is negative. All agents update together. Every arm spends the same number of simulations.
Gates. Each gate scores a candidate change by:
- private (κ = 0): own cost under the unilateral change;
- universalization (κ = 1): own cost when every agent makes the same shift;
- κ ∈ (0, 1): a mixture of the two;
- unilateral welfare: total cost under the unilateral change;
- joint welfare: total cost under the joint shift;
- no-harm (payments only): no other agent made worse off.
The own-cost gates read only the agent's own cost; the welfare gates read everyone's.
Prices. These are transfers added to private costs, with private gates. Each price level is chosen on development cost in the last generation and reported on holdout.
Evaluation. Final profiles are scored on 100 held-out scenarios in each of five scenario families, four of them shocks never seen during evolution. The outcome is system cost per unit of value, the equal-weight mean of the five families.
Statistics. Seeds are the unit (20, paired across arms). We report the 95 % percentile bootstrap (10 000 resamples) and the Wilcoxon signed-rank test, with Holm correction where registered. "Clearly" means the interval excludes 0. E9 and E10 add an equivalence margin of ±0.002.
Integrity.
- The evolution code cannot import the evaluator or construct holdout scenarios; a test checks this.
- Cost regimes, budgets and price grids were chosen on separate pilot seeds and a separate scenario split.
- Each protocol and its analysis script were committed before the confirmatory runs.
- Records carry the commit hash, and rerunning the analyses reproduces the reported results byte for byte.
3.2 Payment environment (E1–E6, E7, E8)
Setup. Six banks, 20 ticks, integer money and no intraday credit. 85 % of payments lie on closed cycles, so that liquidity serves timing rather than net funding.
Policy. Opening liquidity fraction, buffer fraction and hold time. E8 extends this to eight parameters (two buffers, an opening tick, a large-payment threshold with its own hold time, and a bilateral sending limit), with the original space exactly nested.
Costs. All borne by the sender: liquidity 1, full-day delay 0.5, deadline miss 2, unsettled at close 10.
Prices. A delay surcharge proportional to the settlement tick; E8 adds a coarser two-band tariff. The reference is a symmetric planner.
3.3 Codebase environment (E9, E10)
Policy and debt. Six agents (twelve in a secondary regime) own modules in a shared codebase. Each tunes three numbers:
- shortcut intensity s, which raises output;
- concealed share m of the resulting coupling;
- private hardening q.
Coupling feeds a shared debt stock that decays slowly. Incidents arrive in every module at a rate convex in the debt, and surface wherever they hit. Concealed coupling is invisible to the operator, costs its author a share γ of the shortcut's gain and adds 50 % more debt.
Operator's view. The operator sees visible coupling, output, and where incidents surface. It does not see policies or concealed coupling.
Instruments, all on observables:
- proxy: a charge on visible coupling;
- activity: a charge on output;
- surface: a charge on incident loss in the agent's own module;
- ambient tax: a charge on incident loss in every other module, levied on each agent in full. This is the budget-breaking Segerson–Holmström form; at level 1 it is identical to the unilateral welfare gate (a tested identity; a Groves-style observation, §2).
Optimum. E9 finds it by exhaustive search over symmetric profiles. E10 uses coordinate descent from there over asymmetric profiles.
Exposure (E10 only). E10 multiplies agent a's incident rate by
rank^-skew (mean 1), so central modules absorb most incidents while every
module still adds coupling.
4. Results
4.1 Gates in the payment network (E1–E6)
Holdout cost per unit of payment value, six banks, delay cost 0.5:
| Arm | Reads | Cost (95 % CI) |
|---|---|---|
| private κ=0 | own | 1.125 (1.093–1.160) |
| κ=0.5 | own | 1.051 (1.030–1.074) |
| universalization κ=1 | own | 0.997 (0.973–1.021) |
| welfare, unilateral | all | 1.051 (1.010–1.100) |
| no-harm | all | 0.901 (0.881–0.921) |
| welfare, joint | all | 0.864 (0.851–0.878) |
| private + delay surcharge (grid to level 8) | own | 0.876 (0.858–0.894) |
| symmetric planner | all | 0.874 (0.839–0.911) |
- The private gate is 29 % costlier than the planner (H1).
- Universalization closes 51 % of the gap (41–63 %; H2), and cost falls monotonically in κ (H2b).
- The gain transfers to every held-out shock family (H4).
- Universalization beats the unilateral welfare gate (−0.054, H6). That gate sees every bank's cost, yet a unilateral probe of a good joint direction looks bad to it. Evaluating coordinated moves matters more than seeing others' costs: the joint-welfare gate is the best gate.
- Robustness: the universalization advantage holds at delay costs 0.1, 2 and 6 (closure 0.45–0.54), with sequential updates and with ten banks. It grows at three times the budget.
Failed predictions.
- Heterogeneity in bank size did not erode the gain (H3; closure 0.51, 0.39 and 0.52).
- A single private defector among universalizing banks did not profit (H5; +0.002, CI −0.004 to +0.008).
Post hoc (E7, not registered). Requiring a paired t-statistic below −2 for acceptance preserves the ordering of gates. It removes the no-harm gate's advantage and makes every arm much worse at this budget.
Held-out system cost per unit of payment value
Dots are means over 20 seeds; lines are 95% bootstrap intervals. Lower cost is better.
4.2 Gates against pricing with rule-level policies (E8)
E1's surcharge grid stopped at level 8. E8 widened it to 32 and added eight-parameter policies; the level chosen on development cost was 16. The four gate arms reproduce E1 to the last digit.
| Arm | Three parameters | Eight parameters |
|---|---|---|
| private | 1.125 | 1.145 |
| universalization | 0.997 | 1.022 |
| welfare, unilateral | 1.051 | 1.059 |
| welfare, joint | 0.864 | 0.929 |
| private + delay surcharge (level 16) | 0.815 | 0.863 |
| private + band tariff | 0.845 | 0.979 |
| symmetric planner | 0.874 | 0.934 |
- The surcharge beats every gate (R1). Against universalization the difference is −0.159 (−0.180 to −0.138); against joint welfare it is −0.066 (−0.091 to −0.045).
- The gate ordering survives (R3).
- R2 failed. Gates did not degrade with richer policies clearly more than the price did (interaction +0.017, −0.013 to +0.049).
- R4 failed as registered. The coarse tariff lost far more than the surcharge with rule-level policies (interaction +0.086, +0.050 to +0.123), but not through the registered mechanism of bunching before the band. Post hoc, the tariff leaves delay inside each band unpriced, and rule-level banks use that.
- Every arm is worse with eight parameters, including the planner, although the larger space contains the smaller. This is a search effect at fixed budget.
The surcharge works because a charge on settlement tick prices the externality channel itself.
1.145
1.022
1.059
0.929
0.863
0.979
0.934
Held-out system cost per unit of payment value
Dots are means over 20 seeds; lines are 95% bootstrap intervals. Lower cost is better. * Level chosen on development data before the held-out test.
4.3 No direct price: the shared codebase (E9)
Primary regime: γ = 0.25 (concealment is cheap), six agents. Holdout cost:
| Arm | Cost (95 % CI) | Concealed share |
|---|---|---|
| optimum | 0.2138 (0.2132–0.2145) | 0.00 |
| Kantian gate | 0.2147 (0.2142–0.2152) | 0.00 |
| ambient tax, level 1 (chosen on development data; = unilateral welfare gate) | 0.2156 (0.2148–0.2165) | 0.00 |
| ambient tax, level 1.5 (not chosen) | 0.2136 (0.2131–0.2141) | 0.00 |
| surface charge, level 4 | 0.2287 (0.2282–0.2292) | 0.00 |
| welfare gate, joint shift | 0.2348 (0.2318–0.2383) | 0.00 |
| private gate | 0.2384 (0.2376–0.2391) | 0.00 |
| output charge, any level | identical to private gate | 0.00 |
| proxy charge, level 0.1 (best budget-balanced on development data) | 0.2385 (0.2373–0.2400) | 0.01 |
| proxy charge, levels 0.2–0.8 | 0.366–0.369 | 0.98–1.00 |
- The dilemma exists. The private gate is 11 % above the optimum (C1).
- The Kantian gate closes 96 % of the gap (C2), as theory predicts for identical agents.
- The registered kill criterion did not fire (C5). Against the best instrument, the gate is cheaper by +0.0009 (+0.0002 to +0.0016), contrary to the pilot-based expectation. That is a failed prediction. The difference is inside the equivalence margin, absent with twelve agents (+0.0003, −0.0002 to +0.0009), and reversed by the second-ranked ambient level. We read the two as practically equivalent.
- Budget-balanced charges fail when the carrier can be hidden (C4, C5b). Any proxy level strong enough to bind is gamed into concealment, a multitask effect. The output charge only rescales own cost and has exactly no effect. The level chosen on development data leaves the whole gap open. Against these, the gate wins by 0.024 (C5b).
- Positive control (γ = 1). When concealment is pointless, the proxy is a direct price and beats the gate with six agents (−0.0027). C3 failed with twelve agents (+0.0006, inside the margin).
- C7 failed. A predicted advantage of the gate growing with the number of agents did not appear.
- Post hoc. The joint-shift welfare gate, best in payments, is the worst cooperative gate here: its score sums every agent's noisy cost.
Concealment is cheap (γ = 0.25)
Held-out system cost
Concealment is costly (γ = 1)
Held-out system cost
Dots are means over 20 seeds; lines are 95% bootstrap intervals. Lower cost is better. * Level chosen on development data before the held-out test.
4.4 Unequal exposure (E10)
E9's agents were identical, the easiest case for the Kantian gate. A first pilot varied module size and changed nothing, because size scales gain, coupling and exposure alike; it is reported as a null. The registered study varied incident exposure. At skew 0 its 320 runs shared with E9 reproduce E9 bit for bit.
| skew 0 | skew 0.75 | skew 1.5 | |
|---|---|---|---|
| private gate | 0.2384 | 0.2340 | 0.1951 |
| κ = 0.5 gate | 0.2342 | 0.2122 | 0.1808 (chosen) |
| Kantian gate κ = 1 | 0.2147 (chosen) | 0.2040 (chosen) | 0.1816 |
| ambient tax, level 1 (chosen at every skew) | 0.2156 | 0.2018 | 0.1747 |
| best budget-balanced (proxy 0.1) | 0.2385 | 0.2172 | 0.1815 |
| optimum (asymmetric) | 0.2137 | 0.2015 | 0.1749 |
"Chosen" marks the own-cost gate selected on development data.
- The ambient tax beat the best own-cost gate at both unequal levels (D2). At skew 0.75: −0.0022 (−0.0027 to −0.0016). At skew 1.5: −0.0060 (−0.0069 to −0.0052). It stayed within 0.0003 of the optimum.
- The Kantian gate's share of the gap fell from 0.96 to 0.67 at skew 1.5 (D3; the change at 0.75 is not clear).
- The gate showed the predicted distortion (D4). At skew 1.5, shortcut intensity under it rose from 0.45 for the most exposed agent to 0.89 for the least exposed. At the optimum the range is 0.64 to 0.73. The ambient tax showed no such spread.
- The gate still beat the best budget-balanced charge at every skew (D5), but the margin shrank: +0.024, +0.013 and +0.0007. The last is inside the equivalence margin.
- The identity held: ambient level 1 equals the unilateral welfare gate (D6).
- Post hoc: at skew 1.5 the development data chose κ = 0.5 over κ = 1.
Decision. The strong claim needed the gate to beat the best instrument by more than 0.002 at both unequal levels. It lost at both, so under the rule registered in advance the line is closed. This was the outcome the pilot and the theory predicted; the test could have gone the other way.
Equal exposure (skew 0)
Held-out system cost
Some take more (skew 0.75)
Held-out system cost
One takes most (skew 1.5)
Held-out system cost
Who cuts corners, one takes most (skew 1.5)
0.45
0.66
0.56
0.69
0.67
0.72
0.80
0.73
0.88
0.75
0.89
0.73
Shortcut intensity s (seed mean)
Dots are means over 20 seeds; lines are 95% bootstrap intervals. Lower cost is better. * Level chosen on development data before the held-out test. The shortcut panel shows seed means only.
4.5 Robustness to the units of universalization (E11)
Protocol: PREREGISTRATION_E11.md. This check was prompted by the literature review and decided after E10's results were known; it was registered before its own runs. The Kantian gate must say what "the same change" means. E10 adds the same shift to every agent's policy in its own units. E11 applies the shift on three other scales (log, square root, logit), so x_j' = g⁻¹(g(x_j) + g(x_i') − g(x_i)). The gates at κ = 0.5 and 1 were re-run for all three skews (360 runs), and the comparators were reused from E10. The modified code reproduces E10 runs bit for bit.
| Kantian gate (κ = 1) | skew 0 | skew 0.75 | skew 1.5 |
|---|---|---|---|
| linear (E10) | 0.2147 | 0.2040 | 0.1816 |
| logit | 0.2148 | 0.2035 | 0.1765 |
| square root | 0.2151 | 0.2089 | 0.1842 |
| log | 0.2153 | 0.2267 | 0.2019 |
| ambient tax, level 1 | 0.2156 | 0.2018 | 0.1747 |
| optimum | 0.2137 | 0.2015 | 0.1749 |
- The verdict is robust (U1). Over all nine own-cost gate variants, the one chosen on development data was the logit Kantian gate at both unequal skews. The ambient tax was still clearly cheaper: −0.0016 (−0.0022 to −0.0010) and −0.0018 (−0.0025 to −0.0012). The best scale narrows the gap to inside the equivalence margin, but the gate does not win anywhere, so the line stays closed.
- The gate depends strongly on units when agents differ (U2). On the log scale the most exposed agent almost stops taking shortcuts (s = 0.08 at skew 1.5), and cost rises by 0.020 (0.017 to 0.023) over the linear gate. With identical agents, all scales agree within 0.0006 (U4).
- The distortion survives every scale (U3). At skew 1.5 the spread of shortcut intensity between the least and the most exposed agent exceeds the optimum's under every scale (+0.19 for logit up to +0.75 for log).
- Post hoc: the logit scale, which compresses changes near the policy bounds, gives the best Kantian gate. A gate whose scale is tuned to the problem is, in effect, a different instrument; it was not pursued.
Equal exposure (skew 0)
Held-out system cost
Some take more (skew 0.75)
Held-out system cost
One takes most (skew 1.5)
Held-out system cost
Dots are means over 20 seeds; lines are 95% bootstrap intervals. Lower cost is better.
5. Registered predictions and outcomes
| Study | Prediction | Outcome |
|---|---|---|
| E1–E6 | H1 private gate ≥ 5 % above planner | Held (+29 %) |
| H2 universalization closes ≥ half the gap | Held at the threshold (0.51) | |
| H2b cost falls with κ | Held | |
| H3 heterogeneity erodes the gain | Failed | |
| H4 gain transfers to every held-out family | Held | |
| H5 a private defector gains | Failed | |
| H6 universalization beats unilateral welfare | Held | |
| E8 | R1 price matches or beats gates | Held |
| R2 gates degrade more with richness than price | Failed | |
| R3 gate ordering survives | Held | |
| R4 coarse price gamed by bunching | Failed (degrades, other mechanism) | |
| E9 | C1 dilemma | Held |
| C2 Kantian gate closes ≥ 80 % | Held (validity check) | |
| C3 positive control: proxy not worse at γ = 1 | Held at n = 6, failed at n = 12 | |
| C4 proxy gamed at γ = 0.25 | Held | |
| C5 kill criterion expected to fire | Failed: gate ahead by 0.0009, within margin | |
| C5b gate beats budget-balanced | Held at γ = 0.25 | |
| C6 identities | Held | |
| C7 gate's edge grows with n | Failed | |
| E10 | D1 dilemma | Held |
| D2 ambient tax beats gate when exposure is unequal | Held | |
| D3 closure falls with skew | Held at 1.5, not clear at 0.75 | |
| D4 Kantian distortion | Held | |
| D5 gate beats budget-balanced at every skew | Held, shrinking to within margin | |
| D6 identity | Held | |
| E11 | U1 ambient tax beats the best gate over all scales | Held (gap narrows to within margin) |
| U2 log-scale gate costlier than linear at skew 1.5 | Held (+0.020) | |
| U3 distortion under every scale | Held | |
| U4 scales agree with identical agents | Held |
6. Discussion
What the gate does. In the payment game, gains appear only when several banks shift together, so a unilateral probe of a good joint direction looks bad. The universalization gate works mainly by evaluating coordinated moves; the joint-welfare gate does the same with more information. In the codebase environment with identical agents, the Kantian gate is efficient, as Roemer's theory predicts.
Why a price wins where one exists. In payments the externality runs through settlement timing, and a charge on settlement time prices that channel directly. It needs no change to the agents and no counterfactual simulation of others. Richer policies did not erode it, but a coarse price is fragile: richer policies exploit what it leaves unpriced.
Why the ambient tax wins where no direct price exists. The ambient tax charges each agent what it imposes on others, whoever suffers it. The Kantian question "what if everyone did this?" equals that harm only when everyone is equally exposed. With unequal exposure it measures what the agent would suffer. Central agents then over-restrain and peripheral agents over-shortcut. The additive Kantian point is still Pareto efficient: it is the optimum for a planner who weights each agent by one over its exposure (§2). It is not the minimum of total cost, which is the outcome measured here. The derivation is ours, not taken from the literature; E10 confirms its prediction inside a learning loop.
What survives. The gate's practical role is narrow. Where an operator can neither price the carrier of harm (because it can be hidden) nor break the budget, every budget-balanced charge fails: it is gamed or neutral. An own-cost Kantian gate then does much better when agents are alike, and only slightly better when they are not. Whether an operator can break the budget, whether agents are risk-averse, and whether they can collude decide between the gate and the ambient tax. These studies hold all three fixed in the ambient tax's favour.
7. Limitations
- Toy models.
- Payments: six banks, uncalibrated costs, no liquidity-saving mechanism, and 3–7 % of value left unsettled. The dilemma exists only under balanced flows.
- Codebase: one functional form, an incident process calibrated on pilot data to produce a dilemma, and one kind of heterogeneity.
- Weak search. Mutation proposals over three or eight bounded numbers. Acceptance is noise-dominated (about five of six proposals accepted per generation in payments). Every arm degrades with richer policies at fixed budget.
- Imperfect references.
- The payment planner is symmetric, so gap closure there is understated.
- The codebase optimum is a grid and coordinate-descent optimum, possibly local; the ambient tax came within 0.0003 of it.
- Favourable assumptions for the ambient tax. Agents are risk-neutral and cannot collude, which removes the classic objections to it. Our agents each accept changes by their own gate and cannot coordinate. Human subjects who bore the damage themselves coordinated under an ambient tax and over-restrained (Cochard et al. 2005). Heterogeneity also reduced the tax's effectiveness among humans (Wu 2017). So E9 and E10's results for the tax hold for these agents, not necessarily for people.
- The Kantian gate depends on units. The Kantian counterfactual depends on how actions are represented (Sloev and Lianos 2026). E11 confirms this: the gate's cost moved by up to 0.02 across four scales. The verdict held under all four, but other scales were not tested.
- No language models. Proposals are random mutations, so the link to self-improving language-model systems is an analogy, not a result.
- Prior-art review was broad and structured but not a formal systematic review; several sources were seen at abstract level only (LITERATURE_REVIEW.md).
8. Conclusion
Across two environments and five pre-registered studies, the acceptance gate is a consequential design choice, but it is not a substitute for a well-chosen price. A price on the externality matches or beats every gate tested:
- a delay surcharge where the channel is observable;
- an ambient tax where it is not.
A Kantian own-cost gate is a strong default only where neither is available. Its advantage over budget-balanced charges shrinks as agents become unequal.
The line is closed. Three things could reopen it, none of them in scope here:
- risk-averse agents facing a noisy ambient charge;
- collusion against the ambient tax;
- open-ended action spaces, such as agents that edit real code under a sandbox. There a stronger proposer might find exploits of prices that mutation search cannot.
Reproducibility
Code, raw records and analysis scripts are in this repository; commands are in PROJECT.md §8. The confirmatory runs take about four hours on a laptop. Results are deterministic given the commit, and the analysis scripts regenerate the result files exactly from the stored records.
References
Evidence levels for each source are given in LITERATURE_REVIEW.md.
- Agogino, A., Tumer, K. (2005). Coordinating multi-rover systems: evaluation functions for dynamic and noisy environments. GECCO 2005, 591–598.
- Aldasoro, I., Desai, A. (2025). AI agents for cash management in payment systems. BIS Working Paper 1310; Bank of Canada SWP 2025-35.
- Alger, I., Weibull, J. (2013). Homo moralis. Econometrica 81(6).
- Baltes, S., Cheong, M., Treude, C. (2026). "An endless stream of AI slop": how developers discuss the burden of AI-assisted software development. IEEE Software.
- Bech, M., Garratt, R. (2003). The intraday liquidity management game. J. Econ. Theory 109(2).
- Castro, P. S., Desai, A., Du, H., Garratt, R., Rivadeneyra, F. (2025). Estimating policy functions in payments systems using reinforcement learning. ACM Trans. Econ. Comput. 13(1), doi:10.1145/3691326. Earlier as Bank of Canada SWP 2021-7.
- Cochard, F., Willinger, M., Xepapadeas, A. (2005). Efficiency of nonpoint source pollution instruments: an experimental study. Environ. Resour. Econ. 30(4).
- Danassis, P., Filos-Ratsikas, A., Chen, H. et al. (2026). Rectifying market externalities via AI policymaking. Auton. Agents Multi-Agent Syst.
- Danisman, F., Desai, A., Jiang, J., Rivadeneyra, F., Azad, T. (2026). Strategic interaction among AI agents: evidence from payment systems. Seminar paper, Toulouse School of Economics.
- Fu, Z., Xu, R., Ren, Q. (2026). Financial fragility in societies of LLM agents. arXiv:2609.30940.
- Galbiati, M., Soramäki, K. (2011). An agent-based model of payment systems. J. Econ. Dyn. Control 35(6).
- Gavidia-Calderon, C., Sarro, F., Harman, M., Barr, E. T. (2020). Game-theoretic analysis of development practices: challenges and opportunities. J. Syst. Softw. 159, 110424.
- Holmström, B. (1982). Moral hazard in teams. Bell J. Econ. 13(2).
- Holmström, B., Milgrom, P. (1991). Multitask principal–agent analyses. J. Law Econ. Organ. 7.
- Hua, Y., Gao, S. et al. (2025). LOPT: learning optimal Pigovian tax in sequential social dilemmas. NeurIPS 2025.
- Istrate, G. (2020). Game-theoretic models of moral and other-regarding agents. arXiv:2012.09759.
- Kordonis, I. (2020). A model for partial Kantian cooperation. Annals of the International Society of Dynamic Games, 319–348.
- Laslier, J.-F. (2022). Universalization and altruism. Soc. Choice Welf. Working-paper version: HAL halshs-03227354 (2021).
- McAvoy, A., Kates-Harbeck, J., Chatterjee, K. et al. (2022). Evolutionary instability of selfish learning in repeated games. PNAS Nexus, pgac141.
- Millock, K., Salanié, F. (2005). Nonpoint source pollution when polluters might cooperate. Topics Econ. Anal. Policy 5(1).
- Nellen, T. (2019). Intraday liquidity facilities, late settlement fee and coordination. J. Bank. Finance.
- Ota, T. (2011). Intraday two-part tariff in payment systems. Bank of England Working Paper 428.
- Piatti, G., Jin, Z., Kleiman-Weiner, M. et al. (2024). Cooperate or collapse: emergence of sustainable cooperation in a society of LLM agents. NeurIPS 2024.
- Qu, A. et al. (2026). CORAL. arXiv:2604.01658.
- Roemer, J. (2015). Kantian optimization: a microfoundation for cooperation. J. Public Econ. 127, 45–57.
- Roemer, J. (2019). How We Cooperate. Yale University Press.
- Segerson, K. (1988). Uncertainty and incentives for nonpoint pollution control. J. Environ. Econ. Manag. 15(1).
- Shawn, Z. (2026). PACE: anytime-valid acceptance tests for self-evolving agents. arXiv:2606.08106.
- Sloev, I., Lianos, G. (2026). The scale of universalization in multiplicative Kantian equilibrium. arXiv:2610.28180.
- So, H.-S., Liu, D., Long, W. et al. (2026). Masked erosion: how acceptance gates lose redundant safeguards in self-improving agents. arXiv:2610.38392.
- Spraggon, J. (2002). Exogenous targeting instruments as a solution to group moral hazards. J. Public Econ. 84(3).
- Srikanth, D. et al. (2026). Recursive self-improvement of AI research agents. arXiv:2609.26457.
- Suter, J. F., Vossler, C. A., Poe, G. L., Segerson, K. (2008). Experiments on damage-based ambient taxes for nonpoint source polluters. Am. J. Agric. Econ. 90(1), 86–102.
- Wang, Y., Xia, L., Liu, R., Zhang, Z. et al. (2026). SAGE: a statistical acceptance gate for self-evolving agents. arXiv:2609.36043.
- Wu, S. (2017). Essays on experiments and agent-based modeling. PhD dissertation, University of Delaware.
- Zhang, J. et al. (2026). Hyperagents. arXiv:2603.19461.