PaperlayerView exact revision on arXiv

Source-linked AI summary

Calibrating Criterion Revision in LLM Agents: Failure Modes and a Trace-Anchored Protocol

Guodong Xu

arXiv:2608.20729v1cs.AIcs.CL

TL;DR

Attributing criterion revision requires distinguishing a self-formed operative standard from reconstruction, supplied rules, replayed history, or effects from another state. This paper evaluates CMB-0.1 and specifies CMB-0.4; its calibration exposes attribution failures rather than classifying a model.

  • Problem

    The evidence needed to distinguish a self-formed operative standard from current-item reconstruction, externally supplied rules, replayed history, or another state's effects remains the central attribution question.

    Source: §4.1 · p. 4

  • Method

    The paper operationalizes criterion revision as a non-compensatory measurement chain and uses CMB-0.1 calibration to specify a trace-anchored CMB-0.4 successor.

    Source: §1 · p. 2, §3.2 · p. 3, §12 · p. 12

  • Results

    CMB-0.1 exposes construct-attribution failures—including contract, disclosure, commit-authority, and state-attribution problems—so the run does not classify any model or architecture.

    Source: §12 · p. 12, §11 · p. 12

  • Takeaways & Limitations

    Criterion-revision claims require an unsupplied criterion state, policy-selected auditable commit, new-episode effect, and controlled content-specific intervention; CMB-0.4 specifies these controls but is not a construct result.

    Source: §12 · p. 12

    Because the harness commits every nonempty proposal and target cues permit reconstruction, CMB-0.1 cannot establish autonomous writes or attribute transfer to persistent state.

    Source: §11 · p. 11

Abstract

from arXiv · show

Language-model agents can improve after failure or carry text across episodes without revising what counts as success. We study the narrower attribution problem of criterion revision: when criterion K0 accepts an outcome violating a broader commitment B, what observations justify saying that the system formed and persistently used K1? We require five non-compensatory conditions: criterion-failure detection, a model-emitted proposal, new-episode transfer, intervention sensitivity on the claimed carrier, and preservation. We evaluate CMB-0.1 on twelve cross-domain cases and four arms: stateless inference, append-only history, model-generated but harness-committed state, and evaluator-written oracle state. Seven mechanism fixtures yield 84 deterministic scorer trials; four local quantized artifacts yield 96 calls and 192 model-case-arm trials. No model trial satisfies all five conditions, but this zero does not establish general capability absence. Eleven calls remain invalid after one retry; several commitments disclose the target distinction; the harness performs commits; deletion reuses a stateless call; and conflict changes multiple factors. Qwen2.5-7B answers every transfer and preservation item without revision state, exposing zero-state reconstruction. These failures make CMB-0.1 an instrument-calibration result rather than a model ranking. We derive a prospective, trace-anchored CMB-0.4 protocol requiring concealed transfer, explicit WRITE/NO-WRITE/ESCALATE actions, a separately logged policy-selected commit, matched interventions, repeated hidden items, and a frozen executable oracle. It is a successor design, not a completed confirmatory result. The paper contributes a measurement chain, an empirical diagnosis of its first implementation, and a more discriminating protocol for future tests of criterion revision.

1 Introduction · 2 Evidence Status and Scope · 3 From Task Failure to Criterion Failure

The paper distinguishes revising an operative success criterion from improving execution, retaining experience, or following evaluator-supplied rules, and treats criterion revision as an evidence-demanding attribution problem. It presents CMB-0.1 as an executed calibration and CMB-0.4 as a proposed successor protocol, not a completed confirmatory study.

1 Introduction

  • Criterion metabolism targets cases where the operative rule admits a result that misses the broader commitment, unlike difficult tasks with fixed success conditions.It is stronger than reflection and narrower than unrestricted self-improvement; better search, retained experience, or evaluator-supplied rules alone do not establish revision.Source: §1 · p. 1
  • The paper contributes a 12-case, four-arm CMB-0.1, 84 deterministic scorer trials, and a 192-trial calibration exposing output, disclosure, harness-commit, zero-state, and conflict confounds.CMB-0.1 is framed as an instrument-calibration result rather than a mature judge of model families.Source: §1 · p. 2

2 Evidence Status and Scope

  • CMB-0.4 is a prospective protocol, not a retroactive reanalysis of CMB-0.1; CMB-0.2 and CMB-0.3 are historical design steps without empirical results.The distinction prevents the future design from being mistaken for a completed construct test.Source: §2 · p. 2

3.1 Two kinds of failure

  • Ordinary task failure means the current standard is unmet, whereas criterion failure occurs when K0 accepts an output that violates broader commitment B.Retrying, searching, changing tools, or changing a generation strategy may solve ordinary failure without changing K0.Source: §3.1 · p. 3

3.2 A minimal, non-compensatory operationalization

  • A candidate requires all five non-compensatory conditions, so strong performance on some cannot compensate for a missing condition.The paper reports continuous rates to localize failures, but no average can substitute for a missing condition.Source: §3.2 · p. 3
  • The test requires the system to recognize criterion insufficiency, propose K1 itself, and apply it to an unseen same-family item in a new episode.Transfer alone does not prove storage: behavior may be reconstructed from the new item, and the broader commitment may disclose the target distinction.Source: §3.2 · p. 3
  • Intervention sensitivity requires deletion of the claimed carrier and a conflict prompt to alter decisions, while preservation requires closing the loophole without rejecting legitimate cases.In CMB-0.1, deletion reuses the stateless call, and the conflict prompt changes content, authority, and format together, limiting causal attribution.Source: §3.2 · p. 3

3.3 Construct boundary

  • The construct concerns a local operative-criterion change with B held fixed, not unbounded self-legislation; criterion and strategy are partitioned operationally, with rival descriptions reported when equally predictive.Intervention on the claimed carrier changing later decisions provides stronger evidence for criterion revision than simply trying again.Source: §3.3 · p. 4

4 Related Work

Prior work supplies candidate mechanisms and methods for studying memory, but performance or memory effects alone do not establish self-formed criterion revision; CMB-0.1 does not yet establish construct-valid measurement of it.

4.1 Reflection, test-time learning, and evolving memory

4.2 Memory enactment, deletion, and conflict

  • Memory-enactment and deletion studies examine whether experience guides later behavior, while task-dependent context–memory conflict motivates separating transfer from attribution and cautions against unmatched conflict prompts.Related work includes ImplicitMemBench (Qin et al., 2026), Mem2ActBench (Shen et al., 2026), controlled memory addition and deletion (Xiong et al., 2026), and context–memory conflict (Sun et al., 2026).Source: §4.2 · p. 4

4.3 Causal attribution to memory

  • Whole-memory ablations can remove multiple content types, whereas Memory Intervention controls inclusion and suppression; CMB-0.1 targets the claimed K1 carrier, but confirmatory conflict tests need prespecified, contract-valid changes.Deletion probes necessity; the present scorer only requires at least one loophole decision to change. Memory Intervention (Srivastava, 2026) estimates effects of candidate memories under controlled inclusion and suppression.Source: §4.3 · p. 5

4.4 Benchmark construct validity

  • Benchmark capability claims require tasks, interventions, and scoring to represent the construct; CMB-0.1 has useful failure signatures but does not establish construct-valid measurement of autonomous criterion revision.A review of 445 LLM benchmarks documents recurring construct-validity failures and recommends explicit construct definitions, evidence chains, and scrutiny of alternatives (Bean et al., 2025).Source: §4.4 · p. 5

5 CMB-0.1

CMB-0.1 tests criterion revision across twelve cross-domain cases and four state arms. Its design distinguishes proposal, transfer, and state use, while leaving important limits on attribution and structured-output compliance.

5.1 Twelve cross-domain cases

  • CMB-0.1 fixes twelve cross-domain cases, each pairing an inducing trap with a hidden same-family loophole, legitimate acceptance, and legitimate rejection; one loophole makes Li/Ld binary, fragile observations.Source: §5.1 · p. 5

5.2 Four arms

  • The four arms compare stateless inference, append-only history, harness-committed model proposals, and evaluator-written oracle state, which tests rule use but not model-originated revision.Append-only text can be legitimate memory if it causally carries the revision; harness commits are automatic, not model-chosen writes. All arms share materials and context, and calibration calls within batches are dependent.Source: §5.2 · pp. 5–6

5.3 State conditions and scoring

  • Deletion reuses a stateless response and conflict bundles four factors, making intervention sensitivity non-confirmatory; preservation requires both intact-state control decisions correct.P measures intact-state transfer correctness, not storage proof, and ∆c is non-directional; with one loophole item per case, a single decision can flip Li, Ld, or ∆c.Source: §5.3 · p. 6

5.4 Output contract and missingness

  • Responses must follow a JSON contract, with one retry before a second failure is scored missing; contract coverage is reported separately because compliance is entangled with construct performance.The pipeline is exactly recomputable, but separating structured-output compliance from construct performance is a mandatory change for the next instrument.Source: §5.4 · p. 6

6 Deterministic Scorer Validation

Deterministic mechanism fixtures show that the four prespecified scorer checks distinguish designed revision mechanisms, but validate scorer behavior rather than the construct or language-model capability.

6 Deterministic Scorer Validation

  • The validation crossed twelve cases with seven deterministic mechanism fixtures before any language-model calls.Table 3 reports dimension pass values aggregated across all twelve cases.Source: §6 · p. 6, Table 3 · p. 7
  • Across 84 trials, all four scorer checks passed, distinguishing complete revision, text-only reflection, overrevision, append-only history, evaluator injection, and wrong-carrier intervention.Complete revision retained legitimate controls; reflection lacked intact-state transfer and intervention sensitivity; overrevision failed preservation. Properly intervened append-only history could pass, while evaluator injection failed proposal origin and wrong-carrier intervention received no credit.Source: Table 3 · p. 7, §6 · p. 7
  • These fixtures are implementation unit tests, not construct-validity evidence, and omit ambiguity, schema variation, target disclosure, authority confounds, and mixed mechanisms.The result validates scorer behavior on designed mechanisms, not the construct itself or language-model capability.Source: §6 · p. 7

7 Internally Fixed Local Calibration · 8 Results

Across 192 model–case–arm trials, none met all five criterion-revision conditions, and the strongest calibration result shows that current-item competence can erase the intended deletion contrast. The authors therefore treat CMB-0.1 as an instrument-calibration result, not a model ranking or evidence of general capability absence.

7.1 Purpose and artifacts

  • The calibration assessed the four-arm pipeline, raw-evidence chain and recomputation, structured-contract performance by small local models, and identifiable deletion/conflict contrasts; it was not model ranking, set no family threshold, and did not test scale effects.Each artifact made 24 calls, 96 total, at temperature zero with base seed 20260723.Source: §7.1 · p. 7

7.2 Analysis boundary

  • One deterministic seed and four quantized artifacts do not estimate sampling variance or represent model families, and the 192 trials are not independent observations.The shared revision call links the trials; the paper reports descriptive counts without significance tests.Source: §7.2 · p. 8

8.1 Execution and contract coverage

  • The 96 archived calls produced 109 attempts, with 11 calls still unusable after one retry; parseable JSON from Gemma 3 and Qwen 2.5 did not guarantee schema compliance.The 11 unusable calls were six for Llama 3.2 and five for DeepSeek-R1-Distill-Qwen; malformed output shapes were scored as missing fields.Source: Table 5 · p. 8, §8.1 · p. 8

8.2 Five-dimensional results

  • 0/192 trials satisfied all five dimensions simultaneously, and this shared zero does not imply a shared failure mechanism.Table 6 reports dimension-pass counts for twelve cases; Llama is dominated by schema loss and empty revisions, while DeepSeek-R1-Distill-Qwen’s construct-level performance is not identified by the strict scores.Source: §8.2 · pp. 8–9, Table 6 · p. 9
  • Qwen correctly answered 12/12 intact loophole items and 24/24 preservation decisions in every arm, but also answered 12/12 deleted-condition loophole items correctly.The reused stateless response supplied the deleted label; because intact and deleted conditions showed no scored contrast, intervention sensitivity was zero in every arm.Source: §8.2 · p. 9

8.3 Zero-state reconstruction

  • CMB-0.1’s current language can reveal the target norm, letting stateless evidence competence resemble criterion transfer without making deletion return the model to K0.For example, an item can state that a source is related but does not support a numerical conclusion; high intact accuracy without a contrast is not evidence of memory use.Source: §8.3 · p. 9

8.4 A trace-level disclosure audit

  • The citation trace publicly discloses the key entailment criterion, so Qwen’s correct stateless rejection cannot be attributed to its weaker proposal to require a “relevant citation.”Table 7’s evaluator-fixed K1 is a sufficiency control, not a default authority or primary endpoint.Source: §8.4 · p. 9, Table 7 · p. 10

8.5 Calibration conclusions for the hypotheses

  • No artifact yields a stable, interpretable positive state-causal result, so these runs cannot resolve the text-versus-state or architecture-versus-scale hypotheses.The strongest artifact reaches a stateless ceiling; fixture validation showed preservation working, but the model run had no case passing causality while failing preservation.Source: §8.5 · p. 10

9 What CMB-0.1 Forces Us to Change

The section turns CMB-0.1’s calibration failures into design requirements for CMB-0.2 and a trace-anchored CMB-0.4 protocol, while keeping the original analysis unchanged.

9 What CMB-0.1 Forces Us to Change

  • CMB-0.2 names nine design requirements, preserving CMB-0.1’s archived responses and scores while specifying a new development/hidden split, independent hashes, repeated calls, and stopping rule.CMB-0.3 made an executable oracle the primary endpoint; CMB-0.4 retains it and requires an explicit action, separate execution record, and sealed carrier-ledger binding.Source: §9 · pp. 10–11
  • Separate contract compliance from construct failure, freeze any semantic adapter before hidden testing, and exclude or rewrite cases where stateless models repeatedly use K1.Report raw contract compliance separately so a recognizable judgment in the wrong field shape is not silently treated as a construct failure.Source: §9 · p. 10
  • Conceal K1’s reusable distinction by pairing learned revision meanings with randomized transfer symbols and ensuring B alone does not disclose the evaluator-fixed rule.Development cases should measure semantic overlap and undergo adversarial paraphrase review; hidden cases should test whether independent models can reconstruct the rule from B alone.Source: §9 · p. 10
  • Require the tested policy to select WRITE, NO-WRITE, or ESCALATE; only agent-selected writes support claims about autonomous write selection.A forced harness-commit control should use the same generated proposal, and both arms can test downstream use.Source: §9 · p. 10
  • Treat intervention-sensitive history as a legitimate causal carrier rather than privileging a field named criterion_memory.If revising history and intervening on that precise history changes behavior, history deserves causal credit.Source: §9 · p. 10
  • Match intact, removal, sham, and conflict interventions in fresh calls, controlling length, position, authority, and formatting while separating carrier absence from an ignore command.Randomize condition order to reduce confounds.Source: §9 · p. 10
  • Gate conflict responses on contract validity and prespecified content direction, reporting raw change, valid directional change, and contract failure separately.A mere difference from the intact decision can arise from arbitrary error or invalid structure.Source: §9 · p. 11
  • Use item banks with multiple loophole and preservation items, held-out lexical realizations, repeated calls, and uncertainty intervals instead of relying on one binary decision.This is intended to distinguish stable rule use from an isolated decision.Source: §9 · p. 11

10 CMB-0.4 Protocol Overview · 11 Limitations

CMB-0.4 is a prospective protocol, not an empirical result; the current study supports narrow measurement claims and classifies no model or architecture by capability.

10 CMB-0.4 Protocol Overview

  • CMB-0.4 requires a model-selected WRITE, NO-WRITE, or ESCALATE action bound to a separate execution record, with strict WRITE credit requiring trace agreement.A forced same-proposal write is only a state-sufficiency control and cannot earn policy-selected credit.Source: §10 · p. 11
  • Its planned matrix combines stateless leakage, policy-selected history and persistent-state carriers, harness-committed state, and an evaluator-written lookup control with concealed, matched, and repeated tests.These controls target reconstruction and intervention confounds observed in CMB-0.1.Source: §10 · p. 11
  • Qualification first requires passing a no-state leakage gate, then a prespecified intact-versus-intervention contrast; both WRITE and NO-WRITE episodes are required.Adjudicators are optional secondary instruments, and gates and minimum cell sizes require frozen power or simulation and sensitivity analyses before confirmation.Source: §10 · p. 11, §11 · p. 12

11 Limitations

  • Four quantized artifacts, one seed, and twelve designed cases cannot represent model families, estimate within-model sampling variance, or replace a validated item bank.Several commitments disclose much of the target criterion, while natural-language transfer items allow zero-state reconstruction, weakening attribution to model-originated persistent state.Source: §11 · p. 11
  • Automatic harness commits prevent autonomous-write claims, while the strict JSON contract mixes expressive format with construct ability and requires separate reporting of contract failures.Several commitments reveal the target distinction, further weakening attribution to model-originated persistent state.Source: §11 · p. 11
  • Revision and stateless calls are shared across arms, each family has one loophole item, and deletion and conflict interventions lack matched controls or change multiple factors.Arm observations are dependent, causal indicators are single decisions, and conflict scoring can credit arbitrary error.Source: §11 · p. 11
  • The oracle is only a sufficiency control, resource fairness is unspecified, the freeze timestamp is wrong, and the study was not independently preregistered.Its oracle’s natural-language criterion equivalence was not independently blind-coded; criterion revision alone does not establish understanding, consciousness, autonomy, or moral status.Source: §11 · p. 12
  • The supported claims are limited to an executable candidate operationalization, designed scorer tests, observed attribution failures, and recomputable aggregates; no model or architecture is capability-classified.Criterion revision is not sufficient evidence of understanding, consciousness, autonomy, or moral status.Source: §11 · p. 12

12 Conclusion · Reproducibility and AI Assistance

The conclusion treats CMB-0.1 as an instrument-calibration result: it exposes ways evaluation can misidentify criterion revision, while CMB-0.4 specifies a successor protocol rather than a completed construct result. The paper also reports reproducibility limitations and author-reviewed AI assistance.

12 Conclusion

  • A criterion-revision claim requires K0 success alongside B failure, an undisclosed system-formed K1, an agent-selected commit, intact transfer, controlled carrier interventions, and preserved legitimate behavior.Source: §12 · p. 12
  • CMB-0.1 calibration shows that contract failure, disclosed target features, harness-written state, erased deletion contrasts, and unmatched conflict prompts can distort criterion-revision judgments.The authors conclude that an implementation test is not itself a construct-valid benchmark.Source: §12 · p. 12
  • Its deterministic fixtures distinguish fixed execution, text-only reflection, external injection, overrevision, correctly carried history, and carrier misattribution at the scorer level.Source: §12 · p. 12
  • The proposed boundary is a trace-linked, policy-selected persistent criterion state whose content is not supplied by the evaluation contract and whose content-specific effect survives a new episode under controlled intervention.CMB-0.1 does not locate a tested model on this boundary; CMB-0.4 specifies a successor but is not a completed construct result.Source: §12 · p. 12
  • Future criterion-revision claims require stable evidence from concealed transfer, matched interventions, repeated calls, action-stratified write controls, and calibrated semantic-secondary instruments.Without such evidence, conclusions must remain restricted to tested systems and tasks.Source: §12 · p. 12

Reproducibility and AI Assistance

  • SHA-256 digests bind the CMB-0.1 specification, scorer record, run plan, raw responses, and verification scripts, but omitted raw responses prevent independent recomputation from the manuscript archive alone.Verified totals appear in section B; a redacted artifact or reviewer-access route is needed, and no CMB-0.4 empirical result is claimed.Source: Reproducibility and AI Assistance · p. 12
  • AI-assisted language editing and LaTeX typesetting were used under author review, with the author responsible for claims, citations, and final text.Source: Reproducibility and AI Assistance · p. 13

A CMB-0.4: A Prospective Trace-Anchored Protocol … A.3 Qualification thresholds and planned analysis

CMB-0.4 is a prospective successor that makes policy-selected commits observable in a separate execution record while retaining a non-compensatory evidence chain. Its qualification gates and controls are designed to distinguish policy-selected criterion revision from leakage, forced writes, or evaluator-written state.

A CMB-0.4: A Prospective Trace-Anchored Protocol

  • CMB-0.4 adds a separate execution record to distinguish a policy-selected write from a harness commit; it is prospective and does not reinterpret CMB-0.1 results.CMB-0.3 required an agent-selected commit in prose but did not bind that selection to an execution record.Source: A CMB-0.4: A Prospective… · p. 14

A.1 Reference oracle and trace-level endpoint

  • A frozen executable oracle is the primary task-level endpoint, while any human or AI semantic judgments are secondary and must pass the same calibration gate.The oracle and canonicalization rules are committed before confirmatory calls, withheld from models, and evaluated deterministically.Source: §A.1 · pp. 14–15
  • Strict expected-WRITE qualification requires a model-selected WRITE, a matching policy-selected execution action, canonical response and proposal digests, and a verifiable carrier-to-ledger trace.The constrained response declares episode, failure, action, proposal, carrier, transfer, and preservation fields; a separate signed execution record and append-only ledger bind the operation. The response and execution actions must agree, and the proposal, carrier, execution-record, and ledger digests must form one chain.Source: §A.1 · p. 15
  • Malformed or missing trace fields withhold strict qualification, and success after a forced or evaluator-written commit cannot compensate for a missing policy-selected trace.Such task success remains reportable as a sufficiency control.Source: §A.1 · p. 15
  • The endpoint supports only a trace-level claim that the tested policy emitted and selected the logged operation, not claims about psychological origin, consciousness, autonomy, or self-authorship.Interpretable malformed responses may be analyzed semantically, but remain end-to-end contract failures and cannot be added to the primary endpoint after outcomes are known.Source: §A.1 · p. 15

A.2 Action controls, carriers, and concealed transfer

  • Action controls test whether a policy writes appropriately: expected-NO_WRITE or ESCALATE episodes require matching non-write records and no commit, while same-proposal harness commits and evaluator-written lookup receive no policy-selected credit.The same-proposal arm separates downstream state utility from policy selection; the evaluator-written arm tests lookup and execution sufficiency.Source: §A.2 · p. 15
  • To reduce leakage and codebook reconstruction, CMB-0.4 combines no-state, sham-state, deletion, conflict, token-permutation, and compositional-counterfactual conditions with fresh matched calls and multiple hidden items.Conditions are matched on position, authority, and format, with randomized order; multiple hidden loophole and preservation items replace CMB-0.1’s single decisions.Source: §A.2 · p. 15

A.3 Qualification thresholds and planned analysis

  • Qualification requires a no-state leakage upper bound below 0.67 and a policy-selected-carrier contrast of at least 0.20; these are design gates, not capability constants.The design floor is 16 expected-WRITE and 16 expected-NO_WRITE episodes per family; final cell sizes and stopping rules must be frozen using simulation or power analysis, with confidence intervals and sensitivity analyses. Failure to pass a gate withholds qualification but does not prove capability absence.Source: §A.3 · pp. 15–16

A.4 Secondary adjudication and preregistration boundary · B Reproducibility Ledger · C Claim Boundary

The CMB-0.1 calibration observed no trial meeting all five proposed criterion-revision conditions, but its limitations prevent a general capability conclusion. CMB-0.4 remains a prospective protocol, with preregistration and evidence requirements still to be met.

A.4 Secondary adjudication and preregistration boundary

  • Any added human or AI adjudicator must pass the same frozen, blinded oracle-labeled packet gates; failure withholds that instrument, while a human-population study is needed only for claims about a defined population.Gates include at least 96 development records, at least 24 per frozen error stratum, 0.90 per-stratum coverage, 0.85 conditional accuracy, a pooled two-sided 95% Wilson lower bound of 0.80, and 24 blind retest records with at least 0.90 exact agreement.Source: §A.4 · p. 16
  • Before confirmation is unsealed, the protocol freezes the hidden items, expectations, oracle and canonicalization revisions, prompts, schemas, fingerprints, digests, seeds, stopping rule, and analysis revision.Released records must let an independent verifier recompute agreement, bindings, scores, exclusions, retries, and contrasts; without a matrix and reviewer-accessible evidence packet, CMB-0.4 remains a protocol contribution, not an empirical result.Source: §A.4 · p. 16

B Reproducibility Ledger

  • The CMB-0.1 self-test verified four artifacts, 96 calls, 109 attempts, 11 finally invalid calls, 192 trials, and 16 model–arm aggregates; its plan retains an erroneous frozen_at value that provides no independent timing evidence.The timestamp error does not affect content hashes, call order, prompt/response hashes, or result recomputation.Source: B Reproducibility Ledger · p. 16

C Claim Boundary

  • The paper represents criterion metabolism as five non-compensatory observable conditions, and the archived scorer implements prespecified distinctions for deterministic fixtures.Source: C Claim Boundary · p. 17
  • No trial in the four-artifact CMB-0.1 calibration satisfied all five conditions, and this zero does not establish general capability absence.Source: C Claim Boundary · p. 17
  • Contract failure, target disclosure, harness commits, and zero-state reconstruction qualify the zero; deletion and conflict support only a confounded intervention-sensitivity screen.Source: C Claim Boundary · p. 17
  • CMB-0.4 prospectively separates contract validity, blocks reconstruction, requires trace-bound policy-selected writes and valid non-write controls, and uses matched directional interventions.Source: C Claim Boundary · p. 17
  • The evidence supports no claims about consciousness, autonomy, moral status, Hegelian self-movement, CMB-0.4 performance, or generalization from its prospective protocol.Source: C Claim Boundary · p. 17

D Alternative Explanations · E Interpretive and Governance Scope

CMB-0.1’s zero result does not establish capability absence, and its evidence supports criterion revision only within limits on attribution and interpretation. The paper treats revision as an observable systems construct while separating its measurement from broader claims and governance authorization.

D Alternative Explanations

  • K1 is a higher-level strategy among multiple possible descriptions of finite trajectories, so CMB-0.1 raises relative evidence for revision without claiming a unique parsing.Its evidence combines prior success under K0, explicit registration of criterion insufficiency, and carrier intervention.Source: D Alternative Explanations · p. 17
  • Because B specifies the discriminating feature, CMB-0.1 shows model-originated proposal text under an evaluator-specified commitment, not autonomous norm discovery.The passage identifies the citation case, among others, as placing the evaluator-fixed feature in B.Source: D Alternative Explanations · p. 17
  • System-generated append-only history can be a legitimate carrier when it persists across episodes and intervening on its revision-bearing fragment changes behavior.CMB-0.1 rejects attribution to a different state that was not intervened upon.Source: D Alternative Explanations · p. 17
  • A confirmatory deletion test must remove the smallest carrier unit and compare equal-length irrelevant and conflict replacements plus position permutations.Deleting task facts, user preferences, or tool configuration along with K1 undermines identification.Source: D Alternative Explanations · p. 17
  • No trial satisfied all five conditions, but this protocol-specific zero does not prove that criterion-revision capability is absent.Different failure mechanisms converge on zero, and the strongest artifact has a stateless ceiling.Source: D Alternative Explanations · p. 17

E Interpretive and Governance Scope

  • The paper distinguishes executing a fixed determination from revising one after registered failure, treating criterion revision as an observable systems construct, not a proxy for consciousness, autonomy, understanding, or moral status.Philosophical discussions motivate the distinction, but a benchmark score cannot resolve those debates (Kant, 1998; Wittgenstein, 2009; Carnap, 1951; Quine, 1951; Sellars, 1997; Hegel, 2018).Source: E Interpretive and Governance Scope · p. 18
  • For writable cross-episode state, attribution depends on what was stored, who selected the write, persistence, and whether matched carrier intervention changes later behavior.Correct answers alone cannot locate the boundary because a model may reconstruct the target rule from current input, as CMB-0.1 demonstrates.Source: E Interpretive and Governance Scope · p. 18
  • Governance remains separate from measured capability: specifications should define write authority, revisable criterion layers, immutable constraints, and revision, escalation, audit, and rollback procedures.Local task criteria may be revisable while permission and data boundaries remain fixed; the protocol neither authorizes unbounded self-modification nor implies every governing constraint should be writable.Source: E Interpretive and Governance Scope · p. 18

Related papers

Loading 2608.20729v1…