P054 Inspection-game modelling of the regime
What do inspection-game models say the regime's optimal inspection rate and detectable-cheat parameters are, and how do they shift as thresholds fall over time?
Inspection-game models of the regime and their parameters.
Terms: inspection game: a mathematical model of an inspector versus a party that may cheat. In Avenhaus & Canty's notation, d is the inspected party's incentive to cheat and b the cost to it of being caught; the d/b (gain-to-sanction) ratio compares the two.
Status
Analyzed A game-theoretic model of a cloud provider detecting threshold evasion, with no empirical data: it locates a detection gap at a fixed FLOP band under FLOP-only reporting and says more models will fill it as software improves; it derives no inspection rate, no gain-to-sanction ratio and no treaty-regime equilibrium. Shavit (2023) section 3.2 gives a closed-form samples-per-year model (a sampling model, not a game). The inspection-game literature held in the collection (Avenhaus and co-authors) models nuclear safeguards, not this regime. The 'thresholds fall over time' clause has no model.
"We analyzed a simple model of evasion and showed that a detection gap exists in the model and contains AI trained on biological sequence data."
Raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring; Canary Institute research; Daley (2026), Proving the Negative: Three Warnings from Arms Control for the Age of AI (War on the Rocks); Moon, Vedula, Geneson & Bar-on (2025), Strategies and Detection Gaps in a Game-Theoretic Model of Compute Governance (RAND RRA3686-1); Amodo (2026), Example Schemes for Verifying High-Stakes AI Agreements.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P054.1 | Does the item bound change by more than a stated factor (a threshold to be agreed) between linear and log-concave operator gain in covert compute? | open |
| P054.2 | How much higher is a reactive inspector's value than fixed sampling in the small discrete game? | open |
| P054.3 | Is the optimal announced schedule front-loaded under decaying status quo? | a source suggests an answer (not checked) |
| P054.4 | Under a flat heartbeat schedule, what is the escape probability of a day-concentrated one-shot cheat? | a source suggests an answer (not checked) |
| P054.5 | In an inspection game where the cheating threshold declines over time, does the equilibrium inspection rate needed to deter rise over time? | open |
| P054.6 | How many background-monitoring samples per year are needed to catch a 10^25 FLOP run split across colluding provers with at least 90% probability before it completes, at current chip stock? | partly answered |
| P054.7 | For a stated threat (e.g. one covert run per year at a FLOP threshold to be agreed), what maximum unverified fraction of a state's compute keeps it below threshold? | open |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P054.8 | Can monitoring strategies beyond thresholds (proof of learning, KYC, outside account information) be analysed quantitatively in the compute-provider game? |
| P054.9 | What anchored value of the gain-to-sanction ratio d/b should the model use (currently an unanchored default of 3.0 with a prior of 1-30, per internal notes)? |
| P054.10 | Should audit capacity be provisioned as a fraction of fleet compute or as an absolute budget? |
| P054.11 | What audit/checkpoint rate should verifier and prover agree? |
| P054.12 | What undeclared-run size, detection probability, detection latency and distributed-run aggregation rule should the regime target? |
| P054.13 | How late are lies caught under cryptographic escrow, and is the delay acceptable? |
Possibly relevant work found by automated search
Found by AI agents searching for this problem's sub-questions, mostly by keyword and citation in the OpenAlex scholarly database and on the web; a second AI pass reviewed each one, and we spot-checked some by hand. They seemed reasonably related, as far as one can tell from outside each paper's field. Some come from neighbouring fields, such as arms control and financial audit, and bear on the problem by analogy. Not reviewed against the status above; a paper listed here does not change that status. 928 of 1052 quotes checked word for word against the paper's text.
For P054.2: How much higher is a reactive inspector's value than fixed sampling in the small discrete game?
For P054.3: Is the optimal announced schedule front-loaded under decaying status quo?
For P054.4: Under a flat heartbeat schedule, what is the escape probability of a day-concentrated one-shot cheat?
For P054.5: In an inspection game where the cheating threshold declines over time, does the equilibrium inspection rate needed to deter rise over time?
For P054.6: How many background-monitoring samples per year are needed to catch a 10^25 FLOP run split across colluding provers with at least 90% probability before it completes, at current chip stock?
For P054.7: For a stated threat (e.g. one covert run per year at a FLOP threshold to be agreed), what maximum unverified fraction of a state's compute keeps it below threshold?
For P054.9: What anchored value of the gain-to-sanction ratio d/b should the model use (currently an unanchored default of 3.0 with a prior of 1-30, per internal notes)?
For P054.10: Should audit capacity be provisioned as a fraction of fleet compute or as an absolute budget?
* Full text not reviewed; more likely to contain errors.
Data
- problems.csv — one row per problem
- problem_status.csv — status, scope and evidence per problem
- problem_sources.csv — the posed-by quote and other documents per problem
- leaves.csv — all sub-questions, including those without a checkable form
- leaf_sources.csv — every source row behind each sub-question