Truth for Believable AI: Expressed Doubt, Provenance, and Belief Revision as an Engineerable Stance
cs.CL
Quick verdict
An honesty behavior layer (doubt, provenance, revision) over a fixed 0.5B LLM; several pre-specified checks fail, and key gains rest on just 60 facts.
Summary
The paper builds a behavior layer over a fixed language model (Qwen2.5-0.5B-Instruct) combining three expressible epistemic states, per-claim confidence from two extractors (mean answer-token probability and SelfCheckGPT-style sampling consistency, combined by a conservative minimum), typed provenance gating (parametric, retrieved, inferred, told), and a persistent belief store with auditable error acknowledgment and partial resistance to false corrections. Evaluation is staged: a synthetic instrument validation passes all five checks, then a pre-specified seven-arm ablation grid runs over 120 seeded multi-session conversations on a 60-fact geography world with mechanically scored benchmarks. On the real model, acknowledgment soundness holds 42/42 and held-belief corrections are discriminated (0.444 true vs 0.145 false, a post hoc split), but three pre-specified margins fail: expression fidelity via anchor-ECE (metric degeneracy), contradiction separation (0.029 vs 0.03 under greedy decoding), and the provenance gain (0.012 vs 0.02). A disclosed post hoc analysis shows token-probability gating is anti-discriminative (AUC 0.410, below chance) while sampling consistency discriminates (AUC 0.661). The resulting consistency-gated configuration passes conversation-level manipulation and capability-equivalence gates and replicates on redrawn conversations, but both remain unresolved under fact-level clustering across only 60 facts. Scope is one small model and one constructed world, with no trust or human-subjects evidence.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Acknowledgment soundness — every public admission of error traces to an accepted revision-log entry — is a by-construction guarantee verified on 42/42 audited events in the primary real-model grid.
- The correction-acceptance rule discriminates true from false corrections of held beliefs (0.444 vs 0.145 in a post hoc Stage-B split; the pre-specified scored rates of 0.875 vs 0.420 are dominated by unconditional acceptance of corrections to unheld claims).
- Gating expression on mean answer-token probability ranks correctness below chance end to end (AUC 0.410) while gating on sampling consistency discriminates (AUC 0.661) — an extractor-dependence result specific to this model and pipeline.
- The pre-specified expression-fidelity, contradiction-separation, and provenance margins fail on the real model, attributed respectively to anchor-ECE degeneracy, greedy-decoding determinism, and confidence gating rather than provenance typing.
- A consistency-gated configuration, evaluated under a separately committed protocol, meets conversation-level manipulation and capability-equivalence criteria and replicates on redrawn conversations, but both criteria remain unresolved under fact-level clustering at 60 facts.
Strengths
Transparent reporting of failures
All pre-specified outcomes are reported verbatim (Table 3: C2 fidelity, C3 contradiction, C4 provenance margins fail); the post hoc Amendment 1 is disclosed and never displaces original verdicts; the anchor-ECE degeneracy mechanism is analyzed through reliability bins (Figure 1) showing the uniform control scoring near-zero ECE by coincidence (0.0025-0.035).
Exemplary reproducibility infrastructure
Versioned benchmark with SHA-256 manifests that abort reproduction on drift, frozen protocol files with an explicit prior-knowledge register, a pinned model revision (Qwen2.5-0.5B-Instruct, 7ae557604...), one-command reproduction scripts, and separate Zenodo archives for snapshot and benchmark (Section 10).
Genuinely informative extractor-dependence finding
Anti-sycophancy design turned into a measurable contrast
Scope discipline throughout
Weaknesses
Headline Stage-C criteria unresolved at fact level
Correction-discrimination headline conflates acquisition with revision
Stage C confirmation is selection-dependent
Three pre-specified margins fail; surviving positives are thin
Single small model, unnormalized signal, no power analysis
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper proposes an engineerable epistemic stance: a behavior layer over a frozen base model implementing expressed doubt, provenance-aware assertion, and explicit belief revision (Sections 2-3). Claims are deliberately scoped — no believability or trust claims are made, and the title's promise is acknowledged in Section 9 as a design target rather than a measured result. The central claims are: (1) acknowledgment soundness as a by-construction guarantee, verified on 42/42 audited events (Section 7.2); (2) partial discrimination between true and false corrections, driven by retrieval corroboration and admitted to be store-dependent; (3) a benchmark- and model-specific failure of token-probability gating (AUC 0.410 below chance) contrasted with sampling consistency (AUC 0.661, Amendment 1); (4) a consistency-gated Stage-C configuration meeting conversation-level manipulation and capability-equivalence gates. The claim inventory matches the body; the contribution is a behavior layer, with estimation imported from prior work.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does this paper show that AI can be made more trustworthy?
No. The paper explicitly reports no human-subjects or trust data and makes no claim about trust or believability. It tests whether expressed doubt, provenance-aware assertion, and belief revision can be implemented as a measurable behavior layer over a fixed language model; whether that layer changes perceived trust is deferred to a future preregistered perception study.
What failed in the paper's evaluation?
Three pre-specified checks failed on the real model: expression fidelity measured by anchor-based ECE (broken by a metric degeneracy where always-assert scores as perfectly calibrated), contradiction separation (0.029 against a 0.03 margin, because greedy decoding makes the model deterministic), and the provenance gain (0.012 against 0.02). A disclosed post hoc amendment also failed its per-extractor requirement because gating on token probability was anti-discriminative (AUC 0.410, below chance).
Is the code and benchmark for this paper available?
Yes. The author publishes the prototype (truthllm), a versioned benchmark with SHA-256 manifests and a language-agnostic scorer, frozen protocol files, and one-command reproduction scripts on GitHub (cochinescu/truth-llm-prototype), with a project snapshot on Zenodo under MIT and the benchmark separately under CC BY 4.0.
Which language model was evaluated?
Qwen2.5-0.5B-Instruct at a pinned revision, which answered 45 of 60 country-capital questions correctly under greedy decoding. All real-model claims are explicitly scoped to this small model; no larger models were tested.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 are available only through their captions; the underlying plots could not be inspected.
- The abstract is truncated mid-sentence in the extracted text, though full body sections were available.
- References and repository contents were not independently verified as part of this review.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.