Pith. sign in

REVIEW 3 major objections 5 minor 31 references

Local LTL safety contracts let multi-agent learners coordinate safely under decentralised execution without discarding team-optimal joint actions.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 07:37 UTC pith:5XBELJ5A

load-bearing objection Solid formal fix for the coordination gap in multi-agent shields: circular contract products give decentralised safety without the Cartesian conservatism, with library-dependent optimality. the 3 major comments →

arxiv 2606.14130 v2 pith:5XBELJ5A submitted 2026-06-12 cs.LG cs.MA

Contract-Based Compositional Shielding for Safe Multi-Agent Reinforcement Learning

Omar Adalat , Edwin Hamel-De le Court , Francesco Belardinelli This is my paper
classification cs.LG cs.MA
keywords safe multi-agent reinforcement learningcompositional shieldingassume-guarantee contractsLTL safetydecentralised executioncircular fixed-point certificationcontract selection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

In cooperative multi-agent reinforcement learning, one agent's action is often safe only if teammates keep certain promises. Purely unilateral action masks therefore throw away coordinated behaviours that a central shield would allow. This paper shows how to recover those behaviours while still enforcing a single shared global safety objective under fully decentralised execution. Agents choose among finite tuples of local safety obligations whose conjunction entails the global formula; the whole tuple is certified at once by a circular assume-guarantee fixed point, then projected into local action masks. During training a discounted bandit simply selects which pre-certified contract is active, so reward can be optimised without ever admitting an uncertified shield. Across six cooperative benchmarks the method keeps zero safety violations while closing the reward gap to centralised shielded baselines.

Core claim

If a whole tuple of local LTL_safe obligations is certified together—entailing the global safety formula and keeping the initial product state inside the rectangular winning region—then the projected local action masks enforce the global property on every execution, even when each agent relies on the others' obligations as assumptions. When the certified library also contains a contract that admits a team-optimal safe deterministic policy, outer-loop selection over that library can recover the optimal safe team return without centralised runtime control.

What carries the argument

Contract-product rectangular winning region: the greatest fixed point of a predecessor operator that admits only non-empty Cartesian action rectangles closed under all induced successors; its projection yields the local masks of a certified contract.

Load-bearing premise

The finite library of candidate local obligations must contain at least one certified contract that still allows the team-optimal safe joint policy; if that contract is missing, safety is kept but the optimum cannot be recovered.

What would settle it

On any of the six benchmarks, construct or search a certified contract library that admits the known optimal safe joint policy and check whether Contract-IPPO or Contract-IQL reach that policy's team return while still recording zero safety violations; failure on a representable optimum would refute the recovery claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes contract-based compositional shielding for cooperative multi-agent RL under decentralised execution. Agents share a global LTL_safe specification Φ_safe and select among finite tuples of local LTL_safe obligations whose conjunction entails Φ_safe. Whole contract tuples are certified simultaneously via a contract-product rectangular winning region (greatest fixed point of an assume-guarantee predecessor), then projected into local action masks. A discounted-UCB bandit selects among pre-certified contracts at episode/dwell boundaries to optimise team return while preserving deterministic safety. Theorems 1–3 establish circular compositional soundness, conditional optimal-safe recovery under library representability, and safety of certified selection. Evaluation covers six cooperative environments and fifteen algorithmic variants, with structural zero-violation safety and reward gains where factorised shields are unrealisable.

Significance. Safe coordination is a genuine gap between centralised multi-agent shields and conservative factorised masks. The contribution is a clean compositional alternative: circular assume-guarantee contracts certified by a rectangular fixed point, with offline library construction and online bandit selection that never leaves the certified set. Strengths include explicit appendix proofs for Theorems 1–3, open-source code, structural (not penalty-based) safety, and experiments that isolate environments where Cartesian factorisation fails (RWARE, Connector, Pressure Plate). The optimality claim is correctly conditioned on library representability rather than overstated. If the approach scales, it is a useful bridge between formal multi-agent shielding and practical decentralised MARL.

major comments (3)
  1. Theorem 2 and §5.2–5.3: optimal safe recovery is conditional on some certified C* in the finite depth-bounded library admitting the deterministic team-optimal safe policy. The manuscript states this hypothesis clearly, but the main empirical narrative (abstract, §6, conclusion) still reads as if team-optimal safe behaviour is recovered in general. Please make the representability premise equally prominent in the abstract and experimental discussion, and report, for each environment, whether the selected contract is known to admit a near-optimal safe policy or only a safe high-return one among the enumerated library.
  2. §5.1 / Algorithm 1 and Table 1: library construction is a bounded enumeration (up to 12,001 profiles; D=2) with a temporal-form heuristic. Scalability and completeness of Form≤D are load-bearing for any claim beyond safety. The paper should quantify certification cost (time/memory per environment), discuss how the candidate family is generated in practice, and either provide a constructive method for useful local obligations or clearly limit the contribution to settings where a small hand- or heuristically-generated library is available.
  3. Definition 12 (local mask identifiability) and Appendix E: decentralised execution without runtime communication is central to the claim. The manuscript assumes that each agent can recover its projected mask from the active contract identifier and local observation history, but does not systematically verify or measure this for all six environments (especially when product state includes teammate-dependent automaton components). Please state explicitly, per environment, what local information suffices and whether any residual synchronisation is required at runtime.
minor comments (5)
  1. Figure 1 is helpful but the three panels and the Safe(x) definitions are dense; a short caption walk-through of which actions are enabled under each regime would improve accessibility.
  2. Notation for lifted entailment C |=↑ Φ_safe and for local alphabets Api is introduced carefully, but a short summary table of symbols (X_C, Win□_C, Adm_C, Lcert) would help readers navigating §4–5.
  3. Appendix D learning curves are informative; the main-text Figure 4 could note more clearly that ordinary shielded baselines are omitted or replaced by conservative runner-only masks precisely where factorisation is unrealisable.
  4. Related work correctly positions Melcer et al. and Brorholt et al.; a sentence on how the circular fixed point differs from ordered assume-guarantee in terms of training procedure (no cascading order) is already present and could be slightly expanded for readers less familiar with compositional verification.
  5. Minor typos and formatting: “safe coordinationproblems” (missing space) in the abstract; occasional line-break artefacts in formula displays (e.g. Always(p 2 = 0)); ensure consistent use of LTL_safe vs LTLsafe.

Circularity Check

0 steps flagged

No significant circularity: safety follows from simultaneous fixed-point certification and entailment; optimality is explicitly conditional on library representability.

full rationale

The derivation chain is self-contained and non-circular. Theorem 1 proves that any certified contract (entailment C |=↑ Φ_safe plus initial state in the rectangular winning region Win□_C) keeps product executions inside Win□_C under projected masks Adm_C, so local obligations and global Φ_safe hold; the proof is ordinary greatest-fixed-point induction on the monotone operator T_C and does not redefine the target in terms of itself. Theorem 2 is an explicit conditional recovery statement: if some C* in the finite certified library admits the deterministic team-optimal safe policy, then the max over library-compatible policies equals that optimum; the hypothesis is stated in Definition 4, Algorithm 1 and the theorem statement, so the claim is not forced by construction. Theorem 3 simply reapplies Theorem 1 episode-wise under pre-certified arms. The phrase “circular assume-guarantee” denotes simultaneous certification of a whole obligation tuple via a contract-product fixed point (standard compositional-verification technique), not a paper-level circular derivation. No parameters are fitted and then re-presented as predictions; the bandit only selects among already-certified contracts. No load-bearing self-citation of uniqueness theorems or smuggled ansätze appears. Empirical reward curves are measured under structurally enforced zero-violation masks and do not enter the safety proofs. Library representability is an explicit limitation, not a hidden circular premise.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

Central claims rest on standard LTL safety and Markov-game foundations, plus paper-specific contract-product constructions and a finite hand-bounded contract language. Free parameters are search/selection/RL knobs that affect which contracts are found and how fast reward improves, not the safety theorems themselves. Invented entities are definitional engineering objects with proofs, not unfalsifiable physical posits.

free parameters (4)
  • formula depth bound D
    Set to D=2 in experiments; controls expressiveness of Form≤D and thus whether optimal contracts exist in the library.
  • max profiles / candidate family size F
    Bounded enumeration limits (e.g. up to 12,001 profiles) chosen per environment; determines certified library coverage.
  • discounted UCB η, β, warmup H0, dwell H
    Bandit hyperparameters (η=0.95, β often 1.0, dwell/warmup in episodes) chosen by hand; affect which certified contract is selected during learning.
  • RL training hyperparameters (lr, γ, network widths, etc.)
    Standard learner knobs in Appendix F; affect empirical reward curves but not the safety certificates.
axioms (5)
  • standard math Safety LTL fragment with bad-prefix automata characterisation (Kupferman–Vardi style) is adequate for the global and local obligations.
    Used throughout §3–4 and Definition 6 to compile obligations into monitors.
  • domain assumption Cooperative concurrent stochastic games with finite state/action spaces and a finite safety abstraction exact for the shielding problem (e.g. via bisimulation).
    Stated in §3 and §4; synthesis and fixed points require finiteness.
  • domain assumption Local alphabets and projections correctly capture the propositions each agent may promise; shared propositions agree across agents.
    Definition 3 and implementation notes in §6; needed for lifted entailment C |=↑ Φ_safe.
  • domain assumption Local mask identifiability: each agent can compute its projected mask from the active contract id and its observation history without runtime communication.
    Definition 12; required for true decentralised execution of certified rectangles.
  • ad hoc to paper Finite candidate obligation languages Form≤D_i are rich enough that a useful (ideally optimal-preserving) certified contract exists when recovery is claimed.
    Definition 4, Algorithm 1, Theorem 2 hypothesis; the paper’s optimality story depends on this library design choice.
invented entities (3)
  • LTL_safe contract tuple and circular assume-guarantee obligations A_i^C no independent evidence
    purpose: Decompose global safety into local promises agents may assume of each other.
    Definitions 2–5; engineering abstraction with semantic entailment check, not an external physical entity.
  • Contract product X_C and rectangular winning region Win□_C = νW.T_C(W) no independent evidence
    purpose: Certify non-empty Cartesian local action interfaces closed under successors while avoiding bad monitor states.
    Definitions 7–10; standard safety-game fixed point specialised to rectangles for decentralised masks.
  • Certified profile library Lcert(F) with discounted-UCB contract selector no independent evidence
    purpose: Separate offline safety certification from online reward-driven choice among safe shields.
    §5.1–5.3, Algorithms 1–2; methodological construct validated empirically, not independently measured outside the paper.

pith-pipeline@v1.1.0-grok45 · 27396 in / 3359 out tokens · 37050 ms · 2026-07-13T07:37:07.792338+00:00 · methodology

0 comments
read the original abstract

Safe coordination problems surface in multi-agent reinforcement learning when global safety cannot be enforced by any agent unilaterally: the admissibility of one agent's action may depend on the dynamics of other agents. Decentralised shields can enforce safety at runtime, but purely factorised permissions often exclude optimal team behaviour that is safe only through coordination. We study deterministic safety guarantees for agents trained and deployed under decentralised execution, recovering team-optimal safe behaviour without centralised runtime control. Agents have a shared global specification $\phi$ in the safety fragment of Linear Temporal Logic ($\mathsf{LTL}_{\mathsf{safe}}$ ), and select among tuples of local $\mathsf{LTL}_{\mathsf{safe}}$ obligations whose conjunction implies the global specification $\phi$. Each agent may rely on the other agents' local obligations as assumptions because the whole contract tuple is certified simultaneously and allows projection into local action masks. At learning time, a non-stationary multi-armed bandit chooses among a library of local $\mathsf{LTL}_{\mathsf{safe}}$ obligations to select the tuple that optimises team reward, all without forgoing end-to-end safety. We evaluate the approach across 6 environments and 15 algorithmic variants.

Figures

Figures reproduced from arXiv: 2606.14130 by Edwin Hamel-De le Court, Francesco Belardinelli, Omar Adalat.

Figure 1
Figure 1. Figure 1: A 2 × 2 safe-coordination instance in which unilateral local masks lose the safe optimal action, while a refined obligation recovers it through teammate commitment. which makes the cooperative setting a natural target for safe MARL. However, standard reward penalties are insufficient to verify that behaviour is safe once a policy is deployed [14]. From the formal methods standpoint, shielding [16] is a pop… view at source ↗
Figure 2
Figure 2. Figure 2: summarises our approach. The labelled game and global safety spec￾ification are compiled into contract-specific safety products with local obligations providing the assume-guarantee interface of Section 4.1; resulting contract prod￾ucts are solved and projected to decentralised masks in Section 4.2. The certified contracts then form the library used for learning-time selection in Section 5. Global safety Φ… view at source ↗
Figure 3
Figure 3. Figure 3: Safety certification and reward-aware contract selection are distinct objectives. 5.2 Permissiveness and Optimality Definition 14 gives a statewise mask-size metric, obtained simply by counting the sizes of local action masks relative to the protocol-available action sets. Definition 14 (Cardinality Permissiveness). At a shield state x C = (s, q¯), the cardinality permissiveness of agent i’s local mask is … view at source ↗
Figure 4
Figure 4. Figure 4: Team-mean reward curves for all six benchmarks using shielded algorithms. Shaded regions are ± 95% confidence intervals across runs. actions. Where ordinary factorised shields are deployable, contract shielding improves reward and converges toward Shielded-Joint-PPO. For RWARE, Pressure Plate, and Connector, the selected global specifications admit no factorised shield because masks must be conservative ov… view at source ↗
Figure 5
Figure 5. Figure 5: Renders of the six benchmark environments used in the evaluation. occupies it, and door k is open iff that plate is pressed. The latch C ′ contains door k exactly when the runner entered a cell of door k, left such a cell toward the far side, or crossed the door boundary during the transition; it is empty on non-crossing transitions. The episode terminates when some agent reaches the goal or the horizon is… view at source ↗
Figure 6
Figure 6. Figure 6: Learning curves for Pressure Plate and RWARE. Shaded regions are 95% confidence intervals across runs [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Learning curves for Connector and Flatland. Shaded regions are 95% confidence intervals across runs [PITH_FULL_IMAGE:figures/full_fig_p028_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Learning curves for Level-Based Foraging and Car Platoon. Shaded regions are 95% confidence intervals across runs [PITH_FULL_IMAGE:figures/full_fig_p029_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 6 linked inside Pith

  1. [1]

    In: International conference on machine learning

    Achiam, J., Held, D., Tamar, A., Abbeel, P.: Constrained policy optimization. In: International conference on machine learning. pp. 22–31. Pmlr (2017)

    2017

  2. [2]

    Ai Communications35(4), 357–368 (2022)

    Ahmed, I.H., Brewitt, C., Carlucho, I., Christianos, F., Dunion, M., Fosong, E., Garcin, S., Guo, S., Gyevnar, B., McInroe, T., et al.: Deep reinforcement learning for multi-agent interaction. Ai Communications35(4), 357–368 (2022)

    2022

  3. [3]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Alshiekh, M., Bloem, R., Ehlers, R., Könighofer, B., Niekum, S., Topcu, U.: Safe reinforcement learning via shielding. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 32 (2018)

    2018

  4. [4]

    MIT press (2008)

    Baier, C., Katoen, J.P.: Principles of model checking. MIT press (2008)

    2008

  5. [5]

    In: International Conference on Learning Representations

    Bonnet, C., Luo, D., Byrne, D., Surana, S., Abramowitz, S., Duckworth, P., Coyette, V., Midgley, L., Tegegn, E., Kalloniatis, T., et al.: Jumanji: a diverse suite of scalable reinforcement learning environments in jax. In: International Conference on Learning Representations. vol. 2024, pp. 49264–49293 (2024)

    2024

  6. [6]

    arXiv preprint arXiv:2410.10460 (2024)

    Brorholt, A.H., Larsen, K.G., Schilling, C.: Compositional shielding and rein- forcement learning for multi-agent systems. arXiv preprint arXiv:2410.10460 (2024)

    Pith/arXiv arXiv 2024

  7. [7]

    Advances in neural information processing systems33, 10707–10717 (2020)

    Christianos, F., Schäfer, L., Albrecht, S.: Shared experience actor-critic for multi-agent reinforcement learning. Advances in neural information processing systems33, 10707–10717 (2020)

    2020

  8. [8]

    De Witt, C.S., Gupta, T., Makoviichuk, D., Makoviychuk, V., Torr, P.H., Sun, M., Whiteson, S.: Is independent learning all you need in the starcraft multi-agent challenge? arXiv preprint arXiv:2011.09533 (2020)

    Pith/arXiv arXiv 2011

  9. [9]

    Current Robotics Reports2(2), 189–200 (2021)

    Drew, D.S.: Multi-agent systems for search and rescue applications. Current Robotics Reports2(2), 189–200 (2021)

    2021

  10. [10]

    In: Proceedings of the 20th International Conference on Autonomous Agents and MultiAgent Systems

    El Mqirmi, P., Belardinelli, F., León, B.G.: An abstraction-based method to check multi-agent deep reinforcement-learning behaviors. In: Proceedings of the 20th International Conference on Autonomous Agents and MultiAgent Systems. pp. 474–482 (2021)

    2021

  11. [11]

    In: Proceedings of the 20th International Conference on Autonomous Agents and MultiAgent Systems

    ElSayed-Aly, I., Bharadwaj, S., Amato, C., Ehlers, R., Topcu, U., Feng, L.: Safe multi-agent reinforcement learning via shielding. In: Proceedings of the 20th International Conference on Autonomous Agents and MultiAgent Systems. p. 483–491. AAMAS ’21, International Foundation for Autonomous Agents and Multiagent Systems, Richland, SC (2021)

    2021

  12. [12]

    In: International Conference on Learning Representations

    Gallici, M., Fellows, M., Ellis, B., Pou, B., Masmitja, I., Foerster, J., Mar- tin, M.: Simplifying deep temporal difference learning. In: International Conference on Learning Representations. vol. 2025, pp. 78148–78190 (2025)

    2025

  13. [13]

    In: International conference on algorithmic learning theory

    Garivier, A., Moulines, E.: On upper-confidence bound policies for switching bandit problems. In: International conference on algorithmic learning theory. pp. 174–188. Springer (2011)

    2011

  14. [14]

    Advances in Neural Information Processing Systems36, 18964– 18993 (2023)

    Ji, J., Zhang, B., Zhou, J., Pan, X., Huang, W., Sun, R., Geng, Y., Zhong, Y., Dai, J., Yang, Y.: Safety gymnasium: A unified safe reinforcement learning Contract-Based Compositional Shielding 17 benchmark. Advances in Neural Information Processing Systems36, 18964– 18993 (2023)

    2023

  15. [15]

    In: 2nd PASCAL Challenges Workshop

    Kocsis, L., Szepesvári, C.: Discounted ucb. In: 2nd PASCAL Challenges Workshop. vol. 2, pp. 51–134 (2006)

    2006

  16. [16]

    Formal Methods in System Design 51(2), 332–361 (2017)

    Könighofer, B., Alshiekh, M., Bloem, R., Humphrey, L., Könighofer, R., Topcu, U., Wang, C.: Shield synthesis. Formal Methods in System Design 51(2), 332–361 (2017)

    2017

  17. [17]

    Formal methods in system design19(3), 291–314 (2001)

    Kupferman, O., Vardi, M.Y.: Model checking of safety properties. Formal methods in system design19(3), 291–314 (2001)

    2001

  18. [18]

    In: Machine learning proceedings 1994, pp

    Littman, M.L.: Markov games as a framework for multi-agent reinforcement learning. In: Machine learning proceedings 1994, pp. 157–163. Elsevier (1994)

    1994

  19. [19]

    Advances in Neural Information Processing Systems35, 13367–13379 (2022)

    Melcer, D., Amato, C., Tripakis, S.: Shield decentralization for safe multi- agent reinforcement learning. Advances in Neural Information Processing Systems35, 13367–13379 (2022)

    2022

  20. [20]

    arXiv preprint arXiv:2012.05893 (2020)

    Mohanty, S., Nygren, E., Laurent, F., Schneider, M., Scheller, C., Bhat- tacharya, N., Watson, J., Egli, A., Eichenberger, C., Baumberger, C., et al.: Flatland-rl: Multi-agent reinforcement learning on trains. arXiv preprint arXiv:2012.05893 (2020)

    Pith/arXiv arXiv 2012

  21. [21]

    arXiv preprint arXiv:2006.07869 (2020)

    Papoudakis, G., Christianos, F., Schäfer, L., Albrecht, S.V.: Benchmarking multi-agent deep reinforcement learning algorithms in cooperative tasks. arXiv preprint arXiv:2006.07869 (2020)

    Pith/arXiv arXiv 2006

  22. [22]

    John Wiley & Sons (2014)

    Puterman, M.L.: Markov decision processes: discrete stochastic dynamic programming. John Wiley & Sons (2014)

    2014

  23. [23]

    In: Safety and Security in Multiagent Systems: Research Results from 2004-2006, pp

    Raja, A., Barley, M., Zhang, X.S.: Towards safe coordination in multi-agent systems. In: Safety and Security in Multiagent Systems: Research Results from 2004-2006, pp. 1–7. Springer (2009)

    2004

  24. [24]

    Formal Aspects of Computing6(5), 495–511 (1994)

    Sistla, A.P.: Safety, liveness and fairness in temporal logic. Formal Aspects of Computing6(5), 495–511 (1994)

    1994

  25. [25]

    arXiv preprint arXiv:1706.05296 (2017)

    Sunehag, P., Lever, G., Gruslys, A., Czarnecki, W.M., Zambaldi, V., Jader- berg, M., Lanctot, M., Sonnerat, N., Leibo, J.Z., Tuyls, K., et al.: Value- decomposition networks for cooperative multi-agent learning. arXiv preprint arXiv:1706.05296 (2017)

    Pith/arXiv arXiv 2017

  26. [26]

    MIT Press, Cambridge, MA (1998)

    Sutton, R.S., Barto, A.G.: Reinforcement Learning: An Introduction. MIT Press, Cambridge, MA (1998)

    1998

  27. [27]

    coopera- tive agents

    Tan, M., et al.: Multi-agent reinforcement learning: Independent vs. coopera- tive agents. In: Proceedings of the tenth international conference on machine learning. pp. 330–337 (1993)

    1993

  28. [28]

    AI magazine29(1), 9–9 (2008)

    Wurman, P.R., D’Andrea, R., Mountz, M.: Coordinating hundreds of co- operative, autonomous vehicles in warehouses. AI magazine29(1), 9–9 (2008)

    2008

  29. [29]

    arXiv preprint arXiv:2304.06281 (2023)

    Xiao, W., Lyu, Y., Dolan, J.: Model-based dynamic shielding for safe and efficient multi-agent reinforcement learning. arXiv preprint arXiv:2304.06281 (2023)

    Pith/arXiv arXiv 2023

  30. [30]

    avoid the owner’s reserved route while the owner is still unconnected

    Yu, C., Velu, A., Vinitsky, E., Gao, J., Wang, Y., Bayen, A., Wu, Y.: The surprising effectiveness of ppo in cooperative, multi-agent games. Advances in neural information processing systems35, 24611–24624 (2022) 18 O. Adalat et al. Appendix The appendix is organised as follows: Appendix A:LTL safe semantics are provided. Appendix B: Technical results and...

    2022

  31. [31]

    Warmup/dwell

    The benchmark-level discount factor and episode cap are: Benchmark Safety formulaγEpisode cap Pressure Pla te runner_requires_return_routes 0.99 150 R W ARE queue_yield_protocol 0.995 200 Car Pla toon maintain_safe_gap 0.99 60 Connector preserve_reserved_route_clearance 0.99 12 Fla tland avoid_deadlocks 0.99 12 Level-Based Foraging avoid_failed_loads 0.99...