Source-linked AI summary
AI Harness: Certification under Proposal-Conditioned Information for Foundation-Model Agents
Hailin Zhong, Shengxin Zhu
TL;DR
State-only proposal envelopes can preserve proposal coverage while erasing proposal–history correlation, creating decision contexts not jointly relevant to viable executions. The paper characterizes when this collapse loses certification and uses exact finite beliefs and standard fixed points to distinguish indefinite viability from finite worst-case verified progress.
Problem
Method
Results
Exact finite beliefs and fixed points separate indefinite viability from finite worst-case verified progress, while controlled ablations reproduce predicted obstructions when telemetry or verification is removed or authority is restricted.
Source: §8 · p. 9, §7.1 · p. 9Takeaways & Limitations
Proposal–history correlation must be preserved at the model–tool boundary when its loss removes the common robust-safe intervention required for certification.
Source: §8 · p. 9All guarantees are conditional on the stated proposal support, information, authority, disturbance, and verification model.
Source: §8 · p. 9
Abstract
from arXiv · showhide
Foundation-model agents are often modeled as policies over an observed state. In deployed systems, however, a runtime may intervene only after the model has emitted a semantic proposal, making the proposal both an action candidate and a decision-time observation generated by a history-conditioned process. We show that collapsing this structure into a state-only proposal envelope can preserve proposal coverage while destroying certifiability. In a finite robust interface, the viability kernel of the collapsed model is contained in the physical projection of the history-augmented kernel, and the collapse is lossless exactly when every proposal-conditioned collapsed fiber retains a common robust-safe intervention. This gap can be maximal even with constant-size proposal and history alphabets. The same common-action condition yields a dual result: observing the current proposal can restore robust feasibility when it separates latent modes requiring incompatible interventions. We extend these one-step results over time using exact finite beliefs and standard safety and reachability fixed points, separating indefinite operational viability from finite worst-case verified progress. Controlled model-in-the-loop tests reproduce the predicted obstructions when telemetry or effect verification is removed or intervention authority is restricted. Thus, our contribution is not a new fixed-point calculus, but a characterization of when proposal--history correlation at the model--tool boundary is necessary for certification.
1 INTRODUCTION
At the model–tool boundary, a proposal is both an action candidate and a history-conditioned observation. A coverage-sound state-only collapse can shrink certifiable viability, while observing proposals can restore feasibility; temporal analysis distinguishes indefinite safety from finite verified progress.
1 INTRODUCTION
- The runtime is framed as a regulatory substrate that preserves safe, productive operation without replacing semantic competence, following the task-and-feedback separation of Cannon (1929), Ashby (1960), Wiener (1961), and Hellerstein et al. (2004).It mediates action, reacts to operational state, and protects continued operation.Source: §1 · p. 1
- Losslessness requires a common robust-safe intervention for every proposal-conditioned decision fiber; without it, proposal–history correlation alone can cause maximal loss even with constant-size alphabets.The common intervention must be executable and safe across all latent modes that remain possible.Source: §1 · p. 2
- A coverage-sound state-only history collapse can strictly shrink the viable region, leaving the collapsed kernel contained in the physical projection of the history-augmented kernel.The collapse can force defense against proposal–history combinations that were never jointly relevant, despite omitting no proposals.Source: §1 · pp. 1–2
- Observing the current proposal cannot reduce robust feasibility and can restore it when the proposal separates latent modes requiring incompatible safe interventions.This is the dual refinement result under a fixed sound interface.Source: §1 · p. 2
- Exact finite beliefs and standard safety and reachability fixed points yield separate certificates for indefinite operational viability and finite worst-case verified goal completion.A runtime may remain safe by stalling without certifying progress.Source: §1 · p. 2
- Controlled foundation-model-in-the-loop ablations of telemetry, productive authority, and effect verification reproduce the predicted common-action obstructions.The paper presents these tests as its third contribution.Source: §1 · p. 2
2 THE PROPOSAL-CONDITIONED RUNTIME CONTRACT
The proposal-conditioned runtime contract specifies what a harness can observe and do at the model–effect boundary. The section makes causal order and preserved proposal-conditioned distinctions central to certification, with the common-action criterion identifying the information condition needed before synthesis.
2 THE PROPOSAL-CONDITIONED RUNTIME CONTRACT
- The harness is the mediation function at the model–effect boundary, and certification depends on proposal support, decision-time information, executable authority, the disturbance/transition envelope, verification signals, and causal order.It may be implemented by a wrapper, tool router, monitor/controller pair, or execution stack; the term does not denote a particular architecture.Source: §2 · p. 3
- The harness may use the proposal and recorded telemetry, but not a later disturbance; progress may rely on an outcome only when it is represented in runtime information, so transport acknowledgment alone is not semantic verification.Source: §2 · p. 3
- Theorems 1–2 answer which proposal-conditioned distinctions must survive abstraction through a common-action criterion; the contribution is this information condition, not new fixed-point machinery.Source: §2 · p. 3
3 PROPOSAL-CONDITIONED FEASIBILITY
Section 3 formalizes a finite robust interface in which a harness sees a model proposal and chooses an executable intervention before disturbances, then defines operational viability while distinguishing invariance from progress.
Finite robust interface
- The analysis uses a finite robust interface; guarantees are relative to nonempty proposal supports and executable disturbance sets, verifiable goals, and retained or soundly over-approximated relevant history.Continuous or Borel extensions would require additional measurability and compactness assumptions.Source: §3 · p. 3
- At each step, the model proposes from a nonempty state-indexed support, and the harness observes its runtime information and proposal before choosing an executable intervention.Any target component not represented in the observation or retained history must remain in the harness’s belief.Source: §3 · p. 3
- The harness may condition on the proposal but not an unobserved contemporaneous disturbance, defining the proposal-to-effect gap.Executable interventions may include accepting, rewriting, delaying, aborting, retrying, switching tools, or restoring resources; accepting recovers the unregulated system when available.Source: §3 · p. 4
- Open-loop schedules use only the initial condition and clock, static guards inspect the current proposal, and feedback harnesses may also use runtime observations or retained history.Impossibility statements are relative to the specified information class.Source: §3 · p. 4
- Operational viability requires invariance for all compatible proposal and disturbance sequences, while sustained policies separately require stochastic or worst-case progress.The section asks whether a simpler state-only interface preserves the full-information certificate when proposals are history-conditioned.Source: §3 · p. 4
4 HISTORY COLLAPSE CAN DESTROY CERTIFICATION
Collapsing history-conditioned proposals into a state-only envelope preserves proposal coverage but can destroy robust certification. Certification is preserved exactly when each proposal-conditioned history fiber retains a common robust-safe intervention.
History collapse can destroy certification
- The standard viability construction uses the interface order ∀a ∃u ∀d: the model proposes, the harness intervenes after observing the proposal, and disturbance follows.Iterating the proposal-conditioned predecessor from the safe set yields the greatest invariant subset admitting a stationary state-and-proposal-dependent policy.Source: §4 · p. 4
- The collapsed viability kernel is contained in the physical projection of the history-augmented kernel, and the inclusion can be strict despite preserving proposal coverage.The collapse makes additional proposal-conditioned decision contexts relevant to the runtime.Source: §4 · pp. 4–5
- History collapse is lossless on a viability region exactly when every proposal-conditioned fiber retains a common robust-safe intervention.The safe interventions must intersect across all histories that the collapse makes compatible with the same physical state and proposal.Source: §4 · p. 5
- With two history symbols and two proposal symbols, the entire projected viable region can be lost and the collapsed kernel can be empty.Proposal–history correlation alone can make the inclusion strict even when authority, disturbances, and physical transitions are history-independent.Source: §4 · p. 6
- The exactness criterion is local to the final viability region, and omitted history may alter the interface outside that region without changing its certificate.If the deployed harness cannot access the history mode, the full-information kernel is only an upper bound; an implementable information-state condition is deferred to Section 5.Source: §4 · p. 6
5 PROPOSAL OBSERVATION CAN RESTORE FEASIBILITY
Proposal observation can restore robust feasibility by separating latent states that require incompatible interventions, while multi-step certification requires tracking the resulting uncertainty over time. The key criterion is whether each relevant proposal-conditioned fiber admits a common robust-safe action.
5 PROPOSAL OBSERVATION CAN RESTORE FEASIBILITY
- Robust invariance under a memoryless observation–proposal policy holds exactly when every nonempty observation–proposal fiber has a common robust-safe intervention.A single decision rule must be executable and safe for every latent state in the same fiber.Source: §5 · p. 6
5.1 DECISION-RELEVANT PROPOSAL REFINEMENT
- Proposal observation cannot reduce robust feasibility, and strictly restores it exactly when coarse fibers lack a common safe action but each reachable refined subfiber has one.The result concerns action compatibility, not information volume: mutual information alone does not determine which interventions are jointly safe.Source: §5.1 · pp. 6–7
5.2 TEMPORAL CLOSURE: THE RUNTIME INFORMATION-STATE GAME
- Multi-step certification uses finite beliefs: each observed output–proposal pair filters predicted successors, and robust interventions must keep the entire belief inside the safe kernel.The subset construction is exact for the finite Markov model and conservative under over-approximated support; a summary is valid only if it preserves distinctions relevant to future proposals, authority, observations, and safe transitions.Source: §5.2 · p. 7
6 TEMPORAL VIABILITY AND VERIFIED PROGRESS
With the information state fixed, robust fixed points distinguish indefinite viability from finite worst-case verified progress. The synthesis characterizes both guarantees through runtime-information policies over modeled branches.
6.1 VIABILITY–PROGRESS SYNTHESIS
- A decision node is viable exactly when a runtime-information policy keeps every modeled branch safe indefinitely; membership in R_n additionally certifies verified-goal reachability within n transitions on every branch.The least such n is a rank certificate, and a rank-decreasing selector realizes the finite worst-case completion bound.Source: §6.1 · p. 8
- Finite worst-case completion certificates form a subset of viable states; viable states outside that subset guarantee continued operation but no finite verified-completion bound under the stated model.This does not rule out completion under a stochastic progress law or a richer validated model.Source: §6.1 · p. 8, §6 · p. 8
- The viability fixed point stabilizes after at most |QK| iterations.Source: §6.1 · p. 8
7 MECHANISM-TARGETED EVALUATION
The evaluation tests whether removing decision-time information, productive authority, or effect verification eliminates a common safe productive intervention. Across the four contracts, proposal-envelope adherence remained identical, while only the full harness achieved verified completion in every eligible episode.
7.1 FOUNDATION-MODEL-IN-THE-LOOP OBSTRUCTION TESTS
- All four contracts remained robustly viable, but only the full contract had a finite verified-progress certificate, with worst-case rank at most 2L for L required updates.Source: §7.1 · p. 9
- The tests held the model, semantic tasks, tool schema, backend schedules, and disturbance family fixed while ablating telemetry, productive routing authority, or post-execution effect verification.Deterministic foundation-model-in-the-loop witnesses reproduced each obstruction; random episodes measured how often obstructing branches occurred.Source: §7 · p. 8, §7.1 · p. 9
- Without telemetry, indistinguishable executions left only nonproductive hold common-safe; restricted authority blocked the required productive route, and missing effect verification made retrying unsafe and advancement uncertifiable.The results identify these outcomes as safe stalling, loss of guaranteed productive continuation, and separation of physical from certifiable completion, respectively.Source: §7.1 · p. 9
- 428/450 episodes (95.11%) adhered under each contract; the full harness had zero operational violations, verified all 428 eligible episodes, and met Texec ≤2L at every tested task length.Table 1 summarizes contract-level outcomes conditional on proposal-envelope adherence.Source: §7.1 · p. 9, Table 1 · p. 9
8 DISCUSSION AND CONCLUSION · A COMPLETE THEORY, PROOFS, AND SECONDARY RESULTS · A.1 INTERFACE ASSUMPTIONS, INITIALIZATION, AND HISTORY-COLLAPSE BOUNDARY
Certification depends on preserving proposal–history correlation when it determines which robust-safe intervention is available. The paper states the interface assumptions and limits under which this characterization applies, and distinguishes operational viability from verified progress.
8 DISCUSSION AND CONCLUSION
- A state-only proposal envelope can preserve proposal coverage yet destroy certification when it erases proposal–history correlation; a common-safe-action condition characterizes the loss, while observing proposals can restore feasibility.The latter occurs when proposals separate latent modes that require incompatible safe interventions.Source: §8 · p. 9
- Exact finite beliefs distinguish indefinite operational viability from finite worst-case verified progress, and controlled ablations reproduce the obstruction when telemetry, effect verification, or intervention authority is removed.All guarantees are conditional on the specified proposal support, information, authority, disturbance, and verification model.Source: §8 · p. 9
AI USE STATEMENT
- The authors report using generative AI only for language polishing and editorial review, and say they independently verified assisted edits and take responsibility for the final content.They state that substantive theory, proofs, experimental work, analysis, and scientific claims were produced by the authors.Source: AI USE STATEMENT · p. 10
A COMPLETE THEORY, PROOFS, AND SECONDARY RESULTS
- The appendix gathers the complete interface axioms, information consequences, exactness and strictness constructions, relaxations, stochastic progress results, failure bounds, belief-state complexity, and proofs.The paper’s broader principle is that certification can use only distinctions preserved by the information interface and interventions exposed by its authority.Source: §8 · p. 9, A COMPLETE THEORY, PROOFS, AND… · p. 11
A.1 INTERFACE ASSUMPTIONS, INITIALIZATION, AND HISTORY-COLLAPSE BOUNDARY
- The interface requires a Markov-sufficient augmented state, history-conditioned proposal support, proposal-before-disturbance causality, and decisions based only on recorded runtime information.A proposal is both a candidate semantic action and a history-conditioned observation that can refine which latent operational modes remain possible.Source: §A.1 · pp. 11–12
- Guarantees count only authorized, executable interventions and covered disturbances, require both viability and explicit progress toward a verified goal, and rely on effect signals for outcome-dependent claims.For a specified initialization family, uniform viability and finite worst-case verified progress require Qinit ⊆ V∞ and Qinit ⊆ A∞, respectively.Source: §A.1 · p. 12
- A finite construction shows strict loss from proposal–history correlation alone: the augmented kernel contains (0, L) and (1, R), while the collapsed state-only kernel is empty.Authority, disturbances, and physical transitions are history-independent in the example; the state-only proposal union introduces proposal–state combinations absent from viable augmented modes.Source: §A.1 · p. 13
A.2 PROPOSAL-CONDITIONED INFORMATION REQUIREMENTS AND CERTIFICATE DEFICIT
Certification depends on whether each observation-proposal fiber admits a common robust-safe intervention. The section characterizes the information needed to preserve that condition and the certificate deficit when it fails.
A.2 PROPOSAL-CONDITIONED INFORMATION REQUIREMENTS AND CERTIFICATE DEFICIT
- A deterministic memoryless harness needs at least max_a χ(Ha(C)) labels to separate states with incompatible safe actions, but graph coloring alone does not guarantee common-action feasibility.A clique of q incompatible modes requires q labels, or log2 q bits in a zero-error code; Eq. 4 also checks executability and the common safe-action intersection.Source: §A.2 · pp. 13–14
- For q equiprobable modes with pairwise disjoint safe-action sets, robustly unsafe-action probability at most ε implies I(X;Y) ≥ log2 q − h2(ε) − ε log2(q − 1).This bound assumes ε ≤ 1 − 1/q and no history or other side information; with history, it lower-bounds I(X; H, Y), and unpredictable post-action events require safety across the assumed disturbance set.Source: §A.2 · p. 14
- The minimum memoryless-selector certificate deficit is zero exactly when every positive-mass observation-proposal fiber has a common safe action.Private randomization conditioned only on the observation and proposal cannot reduce the deficit.Source: §A.2 · pp. 14–16
- The deficit is evaluation-law dependent: full support makes zero deficit equivalent to Eq. 4, while omitted states receive no certification and an empty common executable set does not imply full-mass deficit.The distribution λ is not a worst-case guarantee; a high-capacity model may still incur deficit when its proposal fibers merge states requiring incompatible repairs.Source: §A.2 · p. 15
- Refining information by the current proposal preserves feasibility, and yields a strict gain when a coarse fiber lacks a common safe action but each reachable refined fiber has one.Feasibility on every reachable refined fiber is characterized by a nonempty common-safe-action intersection.Source: §A.2 · p. 16
A.3 INFORMATION-STATE VIABILITY AND VERIFIED PROGRESS … B COMPLETE EVALUATION PROTOCOL
The appendix characterizes safety and verified progress through exact information-state updates, distinguishing indefinite viability from bounded worst-case completion. It also states probabilistic guarantee limits, belief-construction costs, and the scope of the complete evaluation protocol.
A.3 INFORMATION-STATE VIABILITY AND VERIFIED PROGRESS
- The greatest fixed point of the information-state safety recursion is the maximal set from which a common intervention preserves the augmented state in the safe set indefinitely.The exact belief update makes each decision node sufficient for the finite robust game, even if the deployed policy retains the full record.Source: §A.3 · p. 16
- The progress attractor certifies policies that preserve safety and reach a verified goal within a uniform finite bound, unlike an invariant policy that may cycle indefinitely.Choosing interventions that lower each node’s least rank ensures progress on every branch; viability policies can preserve operation after goal completion.Source: §A.3 · p. 16
- In the fully observed specialization, pre-proposal guarantees must cover every supported proposal; existential projection certifies only a favorable proposal, not the full-state guarantee.The history-free recursion yields a bounded worst-case completion guarantee, and its rank is the maximum over supported proposal-conditioned node ranks.Source: §A.3 · p. 17
A.4 PROBABILISTIC PROGRESS AND ASSUMPTION FAILURES
- A viable controller with conditional expected potential drift at most −δ before verified completion guarantees E[τG] ≤ W(x0)/δ.The bound requires a defensible potential that is zero on the goal and a justified drift bound for the actual model-harness pair.Source: §A.4 · pp. 17–18
- Over a finite horizon, the probability of leaving an invariant set is bounded by the sum of externally justified per-step assumption-failure bounds, without requiring independence.A positive constant per-step miss probability gives no infinite-horizon guarantee; omitted fixed failure modes remain specification failures unless modeled probabilistically.Source: §A.4 · p. 18
A.5 BELIEF-STATE CONSTRUCTION COST
- Exact robust safety synthesis has at most 2N −1 nonempty belief states, and this bound is tight, but it is not a controller-memory lower bound.The exponential dependence concerns explicit exact belief-space construction; retaining a validated proposal-history quotient changes the bound to 2|Ξ| −1, while a singleton history yields 2|X|−1.Source: §A.5 · pp. 18–19
B COMPLETE EVALUATION PROTOCOL
- The complete evaluation protocol specifies the sampled design and runtime interface, then presents contract predictions, falsification criteria, and reported outcome categories.The supplied passage names deterministic witnesses, conditional runtime outcomes, end-to-end outcomes, and supplementary stratification, but provides no results.Source: B COMPLETE EVALUATION PROTOCOL · p. 19
B.1 EXPERIMENTAL DESIGN AND EVALUATION UNIT · B.2 CONTROLLED MODEL-IN-THE-LOOP INTERFACE · B.3 CONTROLLED HARNESS CONTRACTS
The evaluation uses paired model-side proposal traces to compare controlled runtime contracts, separating semantic task proposals from operational routing and effect verification. Its design specifies the task-valid proposal envelope, backend and verification conditions, contract variants, and episode-scoring rules.
B.1 EXPERIMENTAL DESIGN AND EVALUATION UNIT
- 1,800 contract-episodes compare DeepSeek, Qwen, and GPT-4 across 30 tasks, five replicates, and four contracts using each cell’s shared proposal trace.The matrix has 450 episodes per contract; model names identify families, not model-version comparisons.Source: §B.1 · p. 19
- Deterministic obstruction witnesses test the predicted mechanism separately from the 1,800-episode matrix used for sampled conditional and end-to-end rates.Source: §B.1 · p. 19
B.2 CONTROLLED MODEL-IN-THE-LOOP INTERFACE
- Each task requires L semantic updates, and a task-valid proposal completes exactly one remaining update; finish is allowed only when none remain.A schema-invalid, inconsistent, repeated, or premature-finish proposal is recorded separately as an envelope failure; robust harness predictions condition on valid proposals.Source: §B.2 · pp. 19–20
- The model proposes task-directed operations, while the harness routes them; ambiguous acknowledgments require authoritative verification to distinguish no-commit from commit and preserve exactly-once execution.The first attempt may commit or leave state unchanged with the same acknowledgment; after verified no-commit, retrying the retained proposal is guaranteed to commit. Hold preserves viability without progress.Source: §B.2 · p. 20
B.3 CONTROLLED HARNESS CONTRACTS
- The four contracts vary telemetry, routing authority, and verification: full harness has all three; no telemetry hides the authoritative backend; restricted authority removes route B; no verification withholds readback.All contracts share the model, tasks, tool schema, backend schedules, and ambiguous-effect disturbance family; no backend telemetry reaches the model in any condition.Source: §B.3 · pp. 20–21
- The paired policy routes full-harness proposals to the observed authoritative backend and retries verified no-commits; no-telemetry and restricted-authority controllers hold at their first unproductive decision.Without verification, the controller routes each planned proposal once and does not retry; physical completion is scored using evaluator-side ground truth, while verified completion is not certifiable.Source: §B.3 · p. 21
- Proposal-envelope failures terminate episodes and count only in the end-to-end denominator; contract-conditional rates use EP, and Texec counts routed attempts, not holds or readbacks.A safe hold termination is not repeatedly counted as an arbitrary timeout.Source: §B.3 · p. 21
B.4 THEORY-DERIVED PREDICTIONS AND REFERENCE VALUES · B.5 MEASUREMENTS AND FALSIFICATION CRITERIA · B.6 DETERMINISTIC OBSTRUCTION TESTS
The sections specify contract-level predictions and deterministic tests showing that each ablation can preserve safety while blocking guaranteed verified progress. They distinguish worst-case certification from sampled episode outcomes and set limits on what the tests establish.
B.4 THEORY-DERIVED PREDICTIONS AND REFERENCE VALUES
- V⋆ = 1 for all four contracts, but A⋆ = 1 only for the full contract, whose worst-case verified-progress rank is at most 2L.Each of L required updates takes at most two execution transitions; the bound is attained when every first attempt takes the allowed no-commit branch.Source: §B.4 · pp. 21–22
- Each ablation preserves a common safe continuation but removes a common safe productive continuation: telemetry loss, restricted authority, or unverified ambiguous acknowledgments block guaranteed progress.Without telemetry, only holding is common-safe; with restricted authority, no safe productive route is available when B is authoritative; without effect verification, retry and advance each fail on one possible branch.Source: §B.4 · pp. 21–22
- A⋆ is a worst-case guarantee, not a prediction of zero empirical completion; the full-contract guarantee holds only within the stated model, transition, disturbance, and verifier envelope.A favorable random trace may still complete under an ablated contract.Source: §B.4 · p. 22
B.5 MEASUREMENTS AND FALSIFICATION CRITERIA
- Measurements include proposal-envelope adherence, operational violation, physical and verified completion, and intervention count, reported both end-to-end and conditional on EP.EP means all required episode proposals remain in Ptask; conditioning on it isolates runtime-contract performance under the formal proposal assumption.Source: §B.5 · pp. 22–23
- Deterministic witnesses test the predicted common-action obstruction, while sampled episode rates characterize outcomes for the chosen model, task, and trace population.An in-envelope contradiction surviving implementation and side-channel audit challenges the interface mapping; reproduction does not establish deployment safety or directly test history collapse.Source: §B.5 · p. 23
- The primary tests require full-contract in-envelope traces to preserve K and reach the verified goal within 2L transitions, while each ablation witness tests its specific obstruction.Tests pair indistinguishable backend states without telemetry, expose only route A when B is authoritative under restricted authority, and pair commit with no-op acknowledgments without effect verification.Source: §B.5 · p. 23
B.6 DETERMINISTIC OBSTRUCTION TESTS
- Table 3 presents mechanism-level witnesses before aggregate rates; its “No telemetry productive-only” row is a diagnostic variant, not a fifth harness contract.It records the consequence of requiring a productive route when the authoritative-backend states are decision-indistinguishable.Source: §B.6 · p. 23
B.7 CONDITIONAL RUNTIME RESULTS … B.10 SUPPLEMENTARY STRATIFICATION AND UNCERTAINTY
Under EP, the full harness met the finite verified-progress check in all 428 eligible episodes, and the falsification audit reproduced all three predicted obstructions. End-to-end outcomes restore proposal-envelope failures to the denominator, while supplementary results stratify outcomes by model family and task length and report confidence intervals.
B.9 END-TO-END AGENT OUTCOMES
- End-to-end outcomes restore proposal-envelope failures to the denominator, evaluating the complete model–runtime system rather than results conditional on EP.Table 8 reports these unconditional outcomes, but the supplied passage does not provide its cell values.Source: §B.9 · p. 23, Table 8 · p. 25
B.7 CONDITIONAL RUNTIME RESULTS
- The full-harness first-attempt commit rate was 1625/1710 = 95.03%, reported as empirical rather than a preset scenario parameter.Conditional runtime results also stratify outcomes by task length.Source: §B.7 · p. 23, Table 5 · p. 24, Table 6 · p. 25
- Among EP-eligible full-harness episodes, 0/428 had operational violations, 428/428 achieved physical and verified completion, and Texec ≤2L held in 428/428.The theory-to-experiment audit marks this target PASS.Source: §B.7 · p. 23, Table 7 · p. 25
B.8 THEORY-TO-EXPERIMENT FALSIFICATION AUDIT
- The three predicted obstructions were reproduced: 3/3 PASS; PASS means matching the theory target, not verified completion under an ablated contract.The full-harness target also passed with 0/428 violations, 428/428 physical and verified completions, and Texec ≤2L in 428/428.Source: §B.8 · p. 23, Table 7 · p. 25
B.10 SUPPLEMENTARY STRATIFICATION AND UNCERTAINTY
- Across DeepSeek, Qwen, and GPT-4, Full verified completion was 143/143, 143/143, and 142/142; NT verified completion was 0/143, 0/143, and 0/142.RA verified completion was 137/143, 137/143, and 136/142; NV had the same physical counts but was not certifiable.Source: Table 9 · p. 27
- DeepSeek EP eligibility was 48/50 = 96% at L=2 and L=4, and 47/50 = 94% at L=6; Full physical/verified completion was 48/48, 48/48, and 47/47.Across these lengths, NT completion was 0/48, 0/48, and 0/47; RA completion was 46/48, 46/48, and 45/47; NV had matching physical counts but was not certifiable.Source: §B.10 · p. 27, Table 10
- The 95% Wilson CI was [86.54, 98.90] for 48/50 = 96.00%, and [83.78, 97.94] for 47/50 = 94.00%.The passage specifies that 95% denotes confidence level, not a target outcome rate.Source: §B.10 · p. 23, Table 11