Pith.

Artificial Intelligence · arXiv:2609.35149 v3 · submitted 28 Sep 2026

From Migration to Calibration: Aligning Agents with Deployment Requirements

Yaxiao Liu (PwC China AI Center), Pengbo Liu (PwC China AI Center), Yiwen Liu (PwC China AI Center), Yihua Guan (PwC China AI Center), Jiaxing Song (Tsinghua University), Xingjian Li (PwC China AI Center)

The paper says

“Deploying, migrating, or scaling an agent changes its model, support system, application, and users. Calibration should align the agent with destination requirements without requiring source imitation or universal superiority.”

more of the abstract

We propose three standard families: basic capability and operational reachability; technical and professional precision, including terminology, units, reporting periods, program fidelity, and report dependencies; and user context covering roles, use cases, authorized values, permissions, and budgets. A versioned mapping distinguishes requirements to preserve, transform, add, or retire. Four instruments—benchmarks, positive/negative examples, skills, and role-conditioned cost/utility—supply tunable reward channels within a fixed acceptance contract. Consultants annotate methods, branches, bounded loops, checkpoints, relative comparisons, and report refinement; a registry records applicable evidence. Adaptation revises configurations or trains a separate controller or skill while freezing the foundation executor. Independent hard gates govern qualification; RL and preference methods optimize candidates without supplying factual truth. Qualification and held-out improvement over the unchanged target are separate outcomes. Supplied intent labels permit a limited paired recount: small prompt/retry gains coexist with regressions and failed gates. FinQA supplements match prior summaries and retain provenance conflicts. These observations motivate destination-specific evaluation but do not establish calibration efficacy. We specify matched-budget tests, clean reload, complete-denominator reporting, and dependency revalidation. This is a methodological proposal with partially reconciled pilot evidence; confirmatory validation and the greater-than-80% zero-shot task-context acceptance target remain outstanding.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 29 Sep 2026 · of version 1 of the paper; version 3 is newer · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A carefully hedged proposal with zero empirical results; the source-as-non-ceiling design principle deserves attention, but the refutation protocol leans on unit-scoring machinery it never error-bounds. This line is wrong

2 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the references say.

The paper says, at §4.1, §5.2, §8.6, §9

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.68):

“Limitations and Responsible Use Summary extraction, validity labeling, and checkpoint alignment introduce measurement error.”

Pith says: Objection 1 of 2 · major

The per-unit preservation gate F(u,a_T) ≥ F(u,a_S) (§4.1) and the checkpoint non-degradation gate h_v(z_T) ≥ h_v(z_S) (§5.2) are pointwise inequalities over many units, and the Section 8.6 refutation measures violation counts to test whether source evidence adds value. Section 9 explicitly concedes that summary extraction, validity labeling, and checkpoint alignment introduce measurement error ([b531]) and that the evaluator is 'a possible failure point' whose risk held-out checks 'reduce but do not eliminate' ([b541]-[b542]).

more of objection 1

Because these are pointwise gates, small per-unit measurement error compounds: spurious violations block genuinely better targets, silently degrading enforcement to the source ceiling that Section 1 explicitly rejects as a design goal, while correlated labeler error across source and target biases the very violation counts the source-evidence arm of the §8.6 test measures. The paper states that uncertainty 'must be measured' (§4.1) but prescribes no concrete error model, no ground-truth audit of the extraction/validity machinery, and no gate-tolerance sensitivity analysis. I recommend adding to the protocol a prescription for a held-out ground-truth annotation audit of unit extraction and validity labeling, with declared agreement targets, plus a sensitivity analysis showing how much per-unit scoring noise the two gates tolerate before the §8.6 refutation becomes uninterpretable. I note in fairness that the §8.6 design's use of the same independently validated units across all four conditions mitigates shared, condition-common error; the residual problem is noise's power-destroying and bias effect on the source-evidence comparison, which the manuscript does not currently address.

The paper says, at §4.1, §8.6

The review did not quote this passage.

Pith says: Objection 2 of 2 · major

The preserved unit set U^S,v_i is the foundation of the entire preservation obligation, but the paper never specifies how these labels are produced: who validates units, how many annotators, what adjudication rules apply to disagreements between 'valid' and 'unresolved' or 'incorrect', and how 'applicable to the target task' is decided when the target contract differs from the source contract.

more of objection 2

Section 4.1 defines the label categories and mandates independent validation, and Section 7.2 prescribes audits for the relative-judge evaluation, but no analogous audit is prescribed for the unit-labeling step itself, even though the preservation gate is stricter than the relative-judgment evaluation and has direct release-blocking consequences (Figure 3, gate 3a). Because incorrect applicability judgments can create false preservation obligations that force target outputs toward source formatting that the paper elsewhere declares unnecessary (§4.2, 'verbatim retention is unnecessary'), the protocol should specify a procedure for validating and auditing the applicability mapping, including how excluded units and their reasons are recorded.

5 minor objections

The paper says, at Figures 3-4, §5.2

Pith says: minor objection 1 of 5

Several rendered strings have lost spacing or subscripts, e.g. 'provenanceprecede' in the Figure 3 caption and 'passv(zTv)' in §5.2; these should be typeset correctly, and a one-line notation table for ΔQ, F(·), h_v, τ_v, δ_v, and λ would help readers track the gates across Sections 3-5.

The paper says, at §8.2, §8.6

Pith says: minor objection 2 of 5

The two-by-two refutation study is the load-bearing experiment, yet the manuscript specifies pilot sizes ('30 independent tasks' per class) only as a costing pilot and defers power analysis to a future prespecified calculation. It would be strengthen the protocol to state explicitly that the interaction test in §8.6 requires a declared minimum detectable effect for the source-evidence main effect, so that a null result cannot be vacuous.

show 3 more

The paper says, at §5.2

Pith says: minor objection 3 of 5

The distinction between the 'strict improvement after deficit' policy and the 'non-degradation repair' policy is clear and important, but the text should state explicitly which of the two is the default for the §8.6 refutation when a source checkpoint saturates the metric, since the two policies can produce opposite conclusions about checkpoint-repair value and the paper warns against pooling their results.

The paper says, at §2.4

Pith says: minor objection 4 of 5

The discussion of refs 21-24 is considerably longer than the other related-work subsections and reads more like a scoping memo than a literature review; condensing it would improve readability, though the candor about what is and is not claimed as novel is commendable and should be retained.

The paper says, at §7.2

Pith says: minor objection 5 of 5

The statement that 'agreement among several model judges is not equivalent to independent truth, since their biases may be correlated' is exactly right; consider adding an explicit instruction to report pairwise inter-judge disagreement on the audit set in addition to agreement with human labels, since correlated model judges will inflate both measures.

What would settle it

On a held-out subsample of the verifiable numerical tasks specified in Sec. 8.6, take gold answers with known unit decompositions and inject known unit-level defects (drop a required qualification, swap a number, delete a mandatory disclosure), then run the full unit-extraction, validity-labeling, and F-scoring pipeline using two independent labelers/judges, and compute per-unit error rates and per-task false-violation and false-pass rates of the F(u,a_T) ≥ F(u,a_S) gate against the known defects. If the false-violation rate is comparable to or larger than the preservation-effect size the 8.6 experiment is designed to detect, the unit-level gate is unenforceable as specified and the falsification outcome is uninterpretable; the paper as written contains no mechanism to rule this out.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 29 Sep 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic agrees 47 s · 18,591 tokens read, 6,190 written
  3. Editor wrote the verdict above 26 s · 19,661 tokens read, 4,184 written
  4. Referee major revision: 2 major objections, 5 minor 59 s · 18,972 tokens read, 7,490 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Know two things before you set aside time for this one. It is a methodological proposal with zero empirical results, and it is unusually honest about that — the abstract says it, Section 8.6 says it, Section 9 concedes the evaluator is a possible failure point. Second: the genuinely interesting piece, the Section 8.6 falsification protocol, leans on measurement machinery the paper never error-bounds.

What's actually new: not the three-layer taxonomy, not harness search, not DPO/GRPO — the authors explicitly disclaim those. The contribution is the integration, and specifically one design principle: the source should inform the target without becoming a quality ceiling. The unit-level preservation condition F(u,aT) ≥ F(u,aS), hard predicates that can't be traded away for scalar quality, the paired-vs-independent access-regime split, the offline-only source comparison — all internally consistent and carefully hedged. The DPO objective in Section 6.2 is correctly labeled as an application of [12], not a new derivation.

The soft spot is the one the stress-test names, and I think it lands. Both decisive gates — per-unit preservation and checkpoint non-degradation — are pointwise inequalities over many units, and the extraction, validity-labeling, and alignment machinery producing those unit scores has no specified error bound or ground-truth calibration. Section 7.2 does require reporting judge agreement, order sensitivity, and agreement with human annotations, so it is not ignored. But nothing bounds unit-level false-violation or false-pass rates, which means the 8.6 experiment cannot cleanly refute the headline claim: a null result for source-evidence value could be measurement-noise burial, and a positive result could be a labeling artifact. One nuance in the paper's favor: the primary refutation task set is verifiable numerical questions with injected fields and independent references, so ground truth exists there and the objection is partly answered for the test that matters most. It bites harder on open-ended tasks and user judgment.

That gap is normal for a proposal — you'd expect the planned experiments to calibrate the machinery before trusting the gates — but the protocol should state it as a precondition, not an afterthought, because the gates carry the whole argument.

The citation pattern is clean: no self-citations, no fitted parameters, and refs 21–24 are cited as precedents that constrain the novelty claim rather than padded support. The authors know the recent harness-transfer literature and position against it accurately.

Who it's for: deployment teams doing model swaps, cross-border rollouts, and global scaling, plus evaluation researchers. It deserves a real referee, but the referee should make validator-calibration a stated precondition of the 8.6 experiment.

The strongest objection

The central empirical hypothesis (Sec. 8.6) is that source evidence and validated checkpoint repair add measurable value over equal-budget target-native optimization. Every measurement this hypothesis rests on — source-unit extraction, validity labeling, semantic checkpoint alignment, and unit-scoring F(u,·) — is produced by the same model-or-human judgment machinery the framework is meant to discipline. Section 9 concedes this ([b531]: 'Summary extraction, validity labeling, and checkpoint alignment introduce measurement error'; [b541]-[b542]: the evaluator is 'a possible failure point' whose risk held-out checks reduce but do not eliminate). The gap is that no error bound, error model, or ground-truth calibration for this machinery is specified, yet the framework's two decisive gates are pointwise inequalities over many units: the per-unit preservation gate F(u,a_T) ≥ F(u,a_S) (Sec. 4.1) and the checkpoint non-degradation gate h_v(z_T) ≥ h_v(z_S) (Sec. 5.2). Small per-unit measurement error compounds across units: spurious violations block genuinely better targets (so enforcement silently degrades to forcing the source ceiling, which Sec. 1 explicitly rejects as a design goal), while correlated labeler error across source and target systematically biases the very violation counts the source-evidence arm of the 8.6 test measures. The consequence is that the designated refutation experiment cannot cleanly refute or confirm: a null result for 'source evidence adds value' could be measurement-noise burial, and a positive result could be a labeling artifact. Section 7.2 requires reporting judge agreement, order sensitivity, and agreement with human annotations, but nothing bounds unit-level false-violation or false-pass rates, so the decisive effect estimate itself remains unbounded by anything in the protocol.

The reader's weakest_assumption identifies the trustworthiness of the validation machinery as load-bearing; my concern is the same point, sharpened to its most precise technical form. The paper is unusually honest: it labels itself a methodological proposal (abstract, [b11]), reports no performance ([b43]), supplies a concrete falsifiable protocol (Sec. 8.6), separates paired from independent access ([b109], [b370]), requires uncertainty intervals and near-tie audits (7.2), and explicitly concedes the evaluator is a possible failure point ([b541]-[b542]). It deserves credit for this. What remains unresolved is that the concession is existential rather than bounded: the unit-level preservation gate and the checkpoint non-degradation gate are pointwise per-unit inequalities, and per-unit measurement error compounds across the many units in a task, so the gate is either vacuous (with a tolerance that weakens certification) or over-strict (blocking legitimate target gains, making the source a de facto ceiling). Crucially, the same uncalibrated machinery feeds the headline refutation experiment, so the 8.6 outcome cannot be interpreted until the machinery's unit-level error is measured against known ground truth. This does not move the verdict: the paper makes no empirical claim, its protocol can in principle falsify the framework, and the limitation is acknowledged, so CONDITIONAL remains the right verdict; the condition should explicitly require ground-truth calibration of the unit-level and checkpoint measurement machinery before the 8.6 result is treated as evidence either way.

The referee's report

If implemented exactly as specified, the framework would provide an unusually disciplined and auditable methodology for a practically important problem: capability retention across model, jurisdiction, and scale changes. The paper's strengths are real and should be credited: it makes no empirical overclaims; it states a concrete, falsifiable hypothesis in Section 8.6 with explicit conditions under which the distinctive claim weakens; it separates paired and independent access regimes so that a hidden source oracle cannot masquerade as generalization; it mandates matched-budget target-native controls; it honestly concedes that the three-layer taxonomy, generic harness search, and DPO/GRPO use are not novel (Section 2.4 and Section 9); and its DPO/GRPO discussion (Sections 6.2, 7.3) correctly labels these as candidate optimizers, not correctness sources. The value of the contribution is conditional on implementation, which the authors state plainly.

Pith's reading of the paper (a machine's summary, not the authors')

The paper tries to establish that when a working agent's driving model is replaced, the agent is moved across borders, or it is scaled across heterogeneous markets, the right way to preserve its capabilities is an auditable three-layer calibration rather than an interface-compatibility check or a prompt rewrite. The formulation is a constrained behavioral adaptation: information preservation (retain independently validated source content applicable to the target), harness adaptation (align execution at semantic checkpoints and repair within explicit budgets), and user acceptance (enforce recipient-specific output contracts). The basic objective is non-degradation, aggregate $\Delta Q \geq 0$ together with per-unit preservation of every applicable valid source unit and target hard-contract satisfaction, with aggregate improvement as the stronger outcome. The reason this matters is that the paper's distinctive claim is testable: it proposes that explicit preservation obligations and validated checkpoint repair add measurable value beyond equal-budget target-native optimization, and it supplies the two-by-two experiment in Section 8.6 that would refute that claim. The paper is a methodological proposal and reports no empirical performance.

What carries the argument. The central machinery is the three-layer calibration contract whose operational interface is the semantic checkpoint. Information preservation works through traceable summaries: each answer is decomposed into claims, conclusions, necessary conditions, supporting evidence, and actionable recommendations, each unit retaining a span in the original answer, with source units labelled valid, incorrect, irrelevant, or unresolved; the preservation obligation applies only to independently validated applicable units, formalized as $F(u,a_T) \geq F(u,a_S)$ for each such unit $u$. Harness adaptation works through a directed acyclic checkpoint graph $G = (V,E)$ over subgoal dependencies; a checkpoint bundles a semantic subgoal, preconditions, an artifact schema, a validator, and a downstream dependency set, and an alignment relation maps checkpoints to target trace segments many-to-many, so unlike tool APIs are never equated. Validation is gated: hard predicates such as schema validity and supported provenance precede task-grounded scalar quality and cannot be traded away, and in paired calibration a non-degradation gate compares the target artifact with a reliable source artifact. User acceptance works through an output contract $C_t = (S,R,H,P_t,L_t)$ with hard predicates, contextual soft preferences, and provenance, applied section by section and revalidated at whole-document level. The repair loop is bounded: diagnose, adapt (parameter fix, tool substitution, local replanning), revalidate the same gates, invalidate dependent cached artifacts, and stop or escalate when budgets exhaust.

Core claim. On the paper's own terms, the central claim is that agent behavior emerges from a model together with its prompts, context construction, tools, execution policy, and interaction constraints, so replacing the model does not guarantee the surrounding configuration still works; the error is to treat interface compatibility or prompt-level scores as proof of capability retention. The proposed remedy is a joint contract with three interacting layers applied to every deployment scenario. Information calibration decomposes each answer into traceable information units, labels source units as valid, incorrect, irrelevant, or unresolved, and requires unit-level non-degradation $F(u,a_T) \geq F(u,a_S)$ for every applicable valid source unit, while allowing the target to exceed the source. Harness calibration defines a directed acyclic checkpoint graph over subgoal dependencies, aligns source and target traces at semantic checkpoints through an alignment relation that permits many-to-many tool correspondence, validates artifacts against hard predicates and scalar quality gates, and repairs deficits within explicit budgets. User calibration enforces an output contract $C_t = (S,R,H,P_t,L_t)$ — schema, reference templates, hard predicates, soft preferences, provenance — with hard requirements preceding soft preferences. The aggregate bar is $\Delta Q \geq 0$ with target-contract satisfaction; the stronger bar is $\Delta Q > 0$, and average gain cannot compensate for a per-unit preservation violation. The paper's own falsifiable form, in Section 8.6, is that the distinctive claim weakens if source evidence adds no measurable value to the same optimizer, if checkpoint repair performs no better than equal-budget final-only repair, or if gains disappear under independent judges and near-tie audits.

The premise it rests on. The load-bearing premise is that the validation machinery itself — summary extraction, validity labelling, checkpoint alignment, and model-neutral quality scoring — is trustworthy enough that the non-degradation gate is meaningful, and the paper concedes in Section 9 that the evaluator is a possible failure point and that these steps introduce measurement error; if that noise dominates the signal, the unit-level condition $F(u,a_T) \geq F(u,a_S)$ is unenforceable and the framework's accepted behavior reduces to whatever the imperfect validator approves.

If the paper is right (Pith's inferences, not the authors')

Model replacement, including foundation-to-post-trained transitions, is treated as a preservation question: retention across reasoning, evidence use, tool execution, and output adherence must be measured, and a smaller or specialised target that cannot meet the contract must report the deficit rather than promise to recover the capability.

Cross-border deployment is calibrated by holding underlying records and business intent fixed while changing the destination contract; shared core standards coexist with site- and market-specific profiles, and conflicting requirements are surfaced for resolution rather than silently overridden.

Global scale requires per-site and per-market group-level reporting so that passing a global average cannot compensate for failure of a mandatory market or site contract.

The three implementation routes — distill a smaller policy, optimise configuration with frozen weights, or train a separate controller with a frozen backbone — all run under the same evidence-constrained evaluation, and DPO and GRPO are treated as optimizers, not independent sources of truth, so admissibility checks must precede relative selection.

The minimal refutation experiment in Section 8.6 — source evidence present or absent crossed with checkpoint repair or final-output-only repair, budgets matched — is the test on which the framework's claimed added value stands or falls.

If the preservation gate's measurement noise — unit extraction, validity labelling, checkpoint alignment, model-neutral scoring — is not bounded, the non-degradation bar is unenforceable and collapses to whatever the imperfect validator approves; the paper concedes the risk in Section 9 but provides no measurement budget, so a natural extension is to report validator-noise bounds per checkpoint be

The semantic-checkpoint alignment idea generalises beyond agents: any migrated system whose internal executions differ in call counts or tool identities but whose observable subgoal artifacts should match could use the same alignment relation and gate hierarchy to audit capability retention.

The per-unit preservation condition $F(u,a_T) \geq F(u,a_S)$ is effectively a claim-level non-inferiority contract; turning it into a statistical test over units within a task — with preregistered margins and clustering — would make the framework's basic objective directly testable in the way the paper gestures at but does not fully specify.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Deploying, migrating, or scaling an agent changes its model, support system, application, and users. Calibration should align the agent with destination requirements without requiring source imitation or universal superiority. We propose three standard families: basic capability and operational reachability; technical and professional precision, including terminology, units, reporting periods, program fidelity, and report dependencies; and user context covering roles, use cases, authorized values, permissions, and budgets. A versioned mapping distinguishes requirements to preserve, transform, add, or retire. Four instruments—benchmarks, positive/negative examples, skills, and role-conditioned cost/utility—supply tunable reward channels within a fixed acceptance contract. Consultants annotate methods, branches, bounded loops, checkpoints, relative comparisons, and report refinement; a registry records applicable evidence. Adaptation revises configurations or trains a separate controller or skill while freezing the foundation executor. Independent hard gates govern qualification; RL and preference methods optimize candidates without supplying factual truth. Qualification and held-out improvement over the unchanged target are separate outcomes. Supplied intent labels permit a limited paired recount: small prompt/retry gains coexist with regressions and failed gates. FinQA supplements match prior summaries and retain provenance conflicts. These observations motivate destination-specific evaluation but do not establish calibration efficacy. We specify matched-budget tests, clean reload, complete-denominator reporting, and dependency revalidation. This is a methodological proposal with partially reconciled pilot evidence; confirmatory validation and the greater-than-80% zero-shot task-context acceptance target remain outstanding.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-09-29

The authors' main claim, as the classifier picked it out 0.30

“We formulate agent calibration as constrained behavioral adaptation across three interacting layers: information preservation, harness adaptation, and user acceptance; the layers apply to every scenario, not one-to-one to the three.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.98 · human participants 0.11 · controlled comparison 0.49 · randomized assignment 0.03 · states quantitative result 0.03 · compares with prior results 0.10 · mentions code or data availability 0.02

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Agents need calibration when deployment conditions change: replacing a driving model, including a foundation-to-post-trained transition; crossing jurisdictions; or scaling across heterogeneous markets and sources. background 1.00 0.06 0.06 asserted 1.00
Interface compatibility alone does not establish capability retention or target-contract satisfaction. background 0.51 0.12 0.10 asserted 1.00
We formulate agent calibration as constrained behavioral adaptation across three interacting layers: information preservation, harness adaptation, and user acceptance; the layers apply to every scenario, not one-to-one to the three. contribution 0.75 0.40 0.14 asserted 1.00
The basic objective is non-degradation on prespecified capability measures while satisfying target requirements; aggregate improvement is stronger. contribution 0.60 0.13 0.24 asserted 1.00
Information calibration preserves independently validated source content still applicable to the target task. method 0.85 0.11 0.04 asserted 0.98
Harness calibration aligns observable artifacts at semantic checkpoints and repairs them through iteration, tool substitution, or local replanning within explicit budgets. method 0.86 0.12 0.05 asserted 0.99
User calibration enforces recipient-specific output contracts: templates, schemas, and section-level preferences. method 0.79 0.10 0.04 asserted 1.00
A global e-commerce example shows how shared standards coexist with site- and market-specific adapters and validation. contribution 0.47 0.38 0.04 asserted 1.00
We distinguish trainable policies from frozen-backbone configuration or controller optimization, and evidence verification from relative judgment and DPO/GRPO optimization. method 0.53 0.14 0.04 asserted 0.95
Recent harness-transfer and judge-validity studies motivate target-native execution records, separate audits of task validity and near-tie ranking, and matched target-native optimization controls. background 0.91 0.06 0.05 asserted 0.79
We propose held-out evaluations for model changes, cross-border adaptation, and scale, including a factorial test of source evidence and checkpoint repair and group-level reporting to prevent aggregate gains from masking local failures. method 0.47 0.22 0.16 asserted 0.81
This is a methodological proposal; implementation and empirical validation remain future work. limitation 0.52 0.12 0.07 asserted 0.96

What kind of paper the classifier reports this is: position or commentary 0.90 · evidence from argument 0.96

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-29: 1 request, 8,771 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (7)
Figure 1 of the paper
Figure 1. Three calibration layers shared by model replacement, cross-border deployment, and
Figure 2 of the paper
Figure 2. Stepwise harness calibration. (a) Source and target traces align at semantic checkpoints
Figure 3 of the paper
Figure 3. Checkpoint validation gates. Hard predicates such as schema validity and supported
Figure 4 of the paper
Figure 4. Bounded checkpoint repair. Every repair returns an observable artifact to the gates
Figure 5 of the paper
Figure 5. Output-contract admissibility before preference learning. A recipient-specific contract
Figure 6 of the paper
Figure 6. User-contract calibration and three alternative implementation routes. Symmetric, val
Figure 7 of the paper
Figure 7. Joint calibration, frozen execution, and offline evaluation. Adaptation on calibration

Figures are the authors', reproduced from the paper.

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment