Verifiable AI Assistance

Abstract

We investigate how generative AI can adapt to people and ambiguous information while the consequential outputs of a system remain open to independent checking. Tax preparation and tutoring expose complementary parts of this problem. Tax software must turn incomplete evidence into supported facts and repeatable calculations. A tutor must adapt explanations and practice without weakening correctness or confusing assisted performance with learning. We propose a research agenda that separates interpretation, acceptance of evidence, deterministic execution, and evaluation of outcomes. The agenda asks where errors enter, whether explicit uncertainty controls improve the tradeoff between error and useful automation, and whether personalization improves delayed, unassisted learning. We outline comparative experiments, failure measures, and reporting requirements. Hopper Labs is developing tax-preparation and tutoring software as environments in which to investigate these questions. This paper presents a conceptual framework and proposed studies; it reports no new empirical results and makes no claim that the proposed systems are generally safe or effective.

1. Problem and scope

Can a system benefit from the flexibility of a language model while keeping consequential decisions inspectable? A useful assistant must interpret unfamiliar documents, ask clarifying questions, and communicate in language suited to its user. Those capabilities do not by themselves establish that the facts it accepted, the rule it selected, or the explanation it produced are correct.

Tax preparation and personalized tutoring are useful settings for this question because both combine contextual judgment with operations that can often be checked. They also have different success criteria. A tax workflow aims to produce a supported, correct result within a declared scope. A tutoring workflow aims to change what a learner can subsequently do without assistance. A plausible answer or an engaging conversation is insufficient evidence in either setting.

Our proposed contribution is an evaluation framework connecting these two settings: identify the boundary at which a model's suggestion becomes an accepted fact or action, test that boundary, and measure outcomes beyond the conversation. This is a research agenda, not a claim that combining language models with deterministic software is novel. The initial scope is bounded tax tasks and quantitative tutoring tasks with independently reviewable answers. It excludes claims of complete tax coverage, general teaching effectiveness, or a solution to alignment.

2. What can be deterministic?

We use deterministic to mean that fixed accepted inputs, fixed rule and configuration versions, and fixed execution dependencies yield the same computational result. The conversation, document interpretation, and selection of a next question may remain stochastic. End-to-end repeatability requires freezing all consequential inputs, including external data and model-produced decisions; choosing a model temperature alone does not establish it.

Let f denote accepted facts, r a rule version, and c the complete execution configuration. A deterministic evaluator produces y = E(r, c, f). Replaying the same tuple should reproduce y. This property says nothing about whether f was true, r was appropriate, or E implemented r correctly. A repeatable error remains an error. We therefore distinguish replay consistency, correctness against independent reference judgments, and the evidential support for accepted inputs.

Verifiable assistance means that a specified claim can be checked against an explicit criterion and the evidence needed for that check is available to the evaluator. The criterion may be a numerical reference, a formal constraint, or an independently scored assessment. It is not a blanket guarantee about everything the assistant says. Some questions will remain ambiguous or unsupported and should be recorded that way.

3. A conceptual approach

The proposed approach separates four responsibilities: proposing an interpretation, deciding what evidence supports it, executing the accepted task, and evaluating the result. AI may suggest candidate facts, clarifying questions, explanations, or practice items. Acceptance criteria govern which suggestions may affect consequential computation or assessment. A supported calculation or check is then performed, with enough context retained to inspect the result.

For example, an ambiguous amount in a document should remain unresolved until there is adequate support for a value; repeatedly calculating from a guessed amount does not resolve the ambiguity. In tutoring, a story about a learner's interests may change the wording of a question, but the new wording still needs to preserve a valid problem. These are illustrative examples, not reports of experiments.

The same scrutiny applies to the surrounding language. A correct calculation can be accompanied by an unsupported explanation. A correct assessment can be undermined by a hint that reveals the answer. We would evaluate the explanation and assistance policy separately from the numerical or formal check. The research question is whether this separation produces a useful, measurable improvement and where it fails.

4. Tax preparation: automation with explicit uncertainty

RQ-T1 — Where do consequential errors enter? We would distinguish document extraction, interpretation of facts, selection of applicable rules, execution, and communication. The hypothesis is that deterministic execution removes some variability while leaving substantial risk at the earlier acceptance boundaries. A study should be able to disconfirm this hypothesis by finding that another stage dominates errors within the evaluated scope.

RQ-T2 — Do evidence requirements improve useful automation? We hypothesize that explicit support requirements and escalation for ambiguity can lower errors among accepted outputs at comparable coverage. The important comparison is the error–coverage tradeoff: refusing every difficult case can make an accepted subset look accurate without making the system useful. We would report completed work, abstentions, erroneous acceptance, and review effort together.

RQ-T3 — Can a changed fact be handled consistently? Proposed tests would introduce corrected, missing, contradictory, and duplicated information, then assess whether affected outputs and explanations change appropriately. Metamorphic tests would use transformations with a justified expected relationship, such as reordering evidence without changing its meaning. Such tests supplement independently reviewed cases; they do not establish that a tax rule is legally correct.

Each evaluation would declare its supported year, jurisdiction, forms, task boundaries, and reference sources. An unresolved interpretation would be reported separately from a known numerical answer. File-format validation or acceptance by a filing test environment would count as interface evidence, not as proof of legal correctness. This paper offers a software research proposal and supplies no filing positions or tax advice.

5. Tutoring: relevance, persona, and learning

RQ-L1 — Does relevance help learning? We would compare general practice with examples adapted to explicitly supplied interests and goals, while holding the underlying skill, exposure, and independently assessed difficulty as comparable as possible. The hypothesis is that relevant contexts can improve participation, but participation alone will not reliably predict delayed learning or transfer to unfamiliar contexts.

RQ-L2 — What should drive adaptation? A second comparison would test adaptation based on demonstrated work and errors against adaptation based primarily on declared preferences. Learner state would be treated as an uncertain estimate that can be revised, not a fixed identity. Interests would not stand in for ability, and demographic stereotypes or unsupported learning-style categories would not be used as explanations of performance.

RQ-L3 — Can a helpful persona preserve honest feedback? Here persona means the tutor's manner of communication: warmth, directness, pacing, and encouragement. It is distinct from a learner profile. We would test whether requests for reassurance, disagreement with correct feedback, or repeated requests for an answer induce false praise, incorrect grading, or premature answer disclosure. Being supportive should be compatible with correcting a mistake.

For quantitative exercises, proposed validation would check that generated questions are solvable, that answers and scoring criteria agree, and that contextual rewrites do not introduce ambiguity or change difficulty unintentionally. Open-ended responses may require independent rubric-based human assessment. A deterministic grader is useful only within the scope where its criteria are valid; it cannot by itself establish understanding.

6. Proposed evaluation design

Tax study. Begin with authorized public or synthetic cases and independent expert review of the reference answers, including explicit labels for ambiguity and unsupported scope. Compare three conditions using the same model version and comparable budgets: AI assistance with ordinary tools, including a calculator; AI assistance with a fixed deterministic evaluator; and that evaluator plus explicit evidence and uncertainty controls. Give conditions comparable access to relevant sources, record unavoidable differences, and use ablations to separate the contribution of execution from acceptance policy.

The unit of analysis would be an independent case, with related variants grouped during splitting and analysis. Development cases, pilot cases, and held-out evaluation cases must be separated to reduce leakage. Repeated model runs estimate variation; they are not additional independent cases. Adversarial or deliberately corrupted cases would be reported separately from ordinary cases, with their construction and prevalence stated.

Primary tax measures would include error among accepted outputs, acceptance coverage over all in-scope cases, and incorrect acceptance of known unsupported cases. We would also report correct completion without human changes, review time, total cost, latency, and replay consistency. Numerators, denominators, uncertainty intervals, and case exclusions would accompany each rate. Risk–coverage curves would expose the effect of different acceptance thresholds. A system that is more accurate only because it completes far less work would not automatically be judged better.

Tutoring study. First validate items and scoring offline. A subsequent consented study would randomly assign learners to general practice, interest-adapted practice, or practice adapted to demonstrated skill, with comparable exposure and available assistance. If several factors change together, a factorial or ablation design would be needed before attributing an effect to personalization. Baseline attainment would inform stratification; parallel assessment forms would be independently reviewed.

The primary tutoring outcomes would be delayed, unassisted performance and transfer to unfamiliar questions. Immediate assisted success, engagement, satisfaction, and hint use would be secondary measures. Persona robustness would also be evaluated in a separate controlled interaction study so an agreeable conversation is not mistaken for learning. Assessment would be scored independently where feasible, with assessors blinded to condition. Attrition, contamination between conditions, accessibility, and differing exposure would be reported.

Before confirmatory data collection, the study protocol would specify hypotheses, primary outcomes, analysis units, exclusion rules, stopping rules, and a sample-size rationale based on a meaningful target effect or error bound. Exploratory analyses would be labeled. Pilot work would establish feasibility and measurement quality; it would not be presented as confirmatory proof. No study described here has produced results reported in this manuscript.

7. Relationship to existing research

LeanTutor combines natural-language interaction with a theorem prover in a proof-of-concept mathematical tutor [1]. It is a relevant precedent for checking a bounded formal task. Our proposed agenda asks how acceptance of uncertain inputs and measurement of downstream outcomes should be evaluated across domains; it does not imply that tax interpretation or general learning can be reduced to proof checking.

Bastani and colleagues distinguish assisted practice performance from performance after assistance is removed in high-school mathematics [2]. Kestin and colleagues study an AI tutor against active learning in a particular college setting [3]. Together, these studies motivate measuring the instructional design, population, and outcome rather than treating all AI tutoring as one intervention. Neither establishes the effectiveness of the software discussed here.

Anthropic's persona-vector research studies model-internal directions associated with behavioral traits [4]. This is relevant to questions about honesty and sycophancy, but model persona vectors are different from the explicit learner interests and performance records proposed here. We do not claim to implement that technique.

AuditBench evaluates auditing methods against models with deliberately implanted hidden behaviors [5], while Petri supports automated exploration of model behavior through simulated interactions [6]. These motivate testing behavior across interactions and separating discovery of failures from estimates of deployment risk. Our proposed domain tests would investigate specific failure hypotheses; successful tests would not certify broad alignment or exclude undiscovered failures.

8. Products and code as research instruments

Hopper Labs is developing tax-preparation and tutoring software to make these questions concrete. The research role of that work is to provide bounded tasks, repeatable experiments, and observable failure cases. A product demonstration can show that a workflow exists. A comparative study is needed to determine whether a particular approach improves correctness, useful automation, or learning.

Future research updates can connect a question to an experiment, the relevant software version, a measured result, and remaining limitations. Public examples should use synthetic or appropriately authorized material. Reproducible evaluation packages could include a task definition, assumptions, reference judgments, evaluation procedure, and aggregate results. The availability of each artifact would be stated explicitly; this paper does not announce an open-source release of the underlying products.

This public agenda describes research questions and evaluation principles. It does not disclose proprietary implementation details, private source code, production configurations, or individual tax and learner records. Those disclosure limits also limit independent reproducibility. Any eventual empirical paper would need to say exactly what an outside evaluator can reproduce and what remains dependent on private infrastructure.

9. Limitations and threats to validity

A validator can be wrong, incomplete, or share the same blind spot as the system it checks. Reference answers can contain mistakes. Synthetic cases can omit the ambiguity of real documents or real classrooms. An acceptance policy can suppress measured error by declining difficult work, and a tutoring policy can raise short-term scores by giving away too much. These are reasons to use independent review, explicit denominators, and multiple outcome measures.

Results may depend on a model version, subject, learner population, tax scope, or unusually well-curated data. Commercial interests in the products create an incentive to emphasize favorable outcomes; preregistered confirmatory analyses, independent assessment, and reporting negative results would help address that threat. Any participant study would require appropriate consent, privacy protections, and oversight, with additional safeguards for minors.

No results, formal safety guarantees, or comprehensive benchmark are presented here. The manuscript is an initial working paper, has not been peer reviewed, and is not an arXiv submission. Its proposed experiments require further specification and review before execution.

10. Research direction

The central question is how much useful adaptation an AI system can provide while keeping its consequential outputs independently assessable. Tax preparation makes the boundary between uncertain interpretation and repeatable execution visible. Tutoring adds the requirement that assistance improve what a person can later do unaided. Building software in both domains offers a practical way to investigate those boundaries, provided implementation progress and empirical evidence remain clearly distinguished.

References

1. Patel, M., et al. (2026). arXiv:2506.08321v2. LeanTutor: Towards a Verified AI Mathematical Proof Tutor.
2. Bastani, H., et al. (2025). PNAS, 122(26), e2422633122. Online version includes the affiliation correction. Generative AI without guardrails can harm learning: Evidence from high school mathematics.
3. Kestin, G., et al. (2025). Scientific Reports, 15, 17458. AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting.
4. Anthropic (2025). Research report. Persona vectors: Monitoring and controlling character traits in language models.
5. Anthropic (2026). Alignment Science Blog. AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors.
6. Anthropic (2025). Research report. Petri: An open-source auditing tool to accelerate AI safety research.