Voice-First Written Assessment: Restoring Evidence of Reasoning Through Structural Redesign

Synthesis: Written assessment has always rested on an inferential claim: that a submitted text is dependable evidence of the reasoning that produced it. Webb-Davies, Ratner and Bassett argue that generative AI has broken that claim, and that the sector's main responses — detection, honor declarations, rules about permitted use — cannot restore it, because they govern conduct rather than evidence. Their proposal, Voice-First Written Assessment (VFWA), separates the capture of reasoning from its expression: an observed, ungraded first stage records how a student thinks in whatever language or modality suits them, and a second stage, completed at home with AI unrestricted, is graded only in relation to that baseline. It is a framework argument rather than a study: the paper reports no empirical data, no sample and no measured outcomes, and presents the model as a design direction awaiting pilots.

The inference the essay can no longer support

The argument is built on validity rather than misconduct. Drawing on Messick's distinction between construct-irrelevant variance and construct underrepresentation, the authors contend that a polished final essay does not show how a student generated claims, chose evidence or exercised judgment — or whether the student did those things at all. Their sharpest sentence is that a take-home essay is not made valid by attaching an unenforceable rule to it.

They add two further frames. Murphy and O'Neill's ecological validity asks whether assessment conditions resemble the settings where the assessed behavior actually occurs, which cuts against both supervised exams and unsupervised essays in different ways. Dawson's separation of authentication (whether the person assessed produced the work) from control of circumstances (whether the conditions support the inference) gives the design its target: a single evidentiary checkpoint where authorship and reasoning are both secured.

The equity argument runs alongside the validity one rather than after it. If academic prose is treated as the natural form of reasoning, language performance becomes part of the construct whether or not the assessor intended it, and the students it fails are disproportionately those whose thinking does not arrive pre-formed in prestige academic English. On that reading, widening access at the point where reasoning is captured is evidence protection, not a concession.

What the model actually requires

VFWA has two stages and one safeguard. Stage 1 is observed, uses a previously unseen prompt, is assessed only for sufficiency rather than graded, and permits the widest feasible range of expression: spoken response, structured outline, argument map, annotated diagram, informal English, another language. Translation tools are allowed to render responses into a shared language, so markers need not be multilingual.

"Voice-First" carries both meanings at once. Spoken voice is a legitimate evidential channel, not a fallback; written authorial voice means the student works in the repertoire that most directly enables thinking before it is disciplined into a single register. The authors reject the Ong-style hierarchy in which writing alone enables analytic thought, arguing instead that modalities make different aspects of reasoning observable.

Stage 2 is the graded submission, written at home with digital tools permitted and optional. What is assessed there is accountable governance of mediation: what the student retained, revised, added and justified. Change between stages is expected and treated as learning rather than as suspicion; the requirement is auditable connection, not consistency.

The model is aimed at divergent tasks, where there is no single correct answer and the point is to see complex reasoning, critical judgment and integration of knowledge. Its theoretical basis is the authors' earlier DETAP account, which treats academic writing as repertoire-based, digitally mediated meaning-making and locates competence in the accountable governance of linguistic mediation rather than in unaided prose production. VFWA is that account turned into an evidentiary architecture: the paper's four core features are an observed baseline, ungraded treatment of that baseline, later refinement under realistic conditions, and interpretation of the final submission against the baseline rather than as a standalone product.

The interpretive step is deliberately holistic. An assessor first checks for continuity of question, stance, framing or reasoning moves, and only if that fails traces whether a plausible developmental path explains the divergence. Confidence is graded confident, uncertain or not credible, and only the last two trigger a conditional viva — a narrow oral check about how a position developed or why evidence was selected, not a second assessment. Where the viva confirms ownership, no penalty follows.

Key Findings

1. The paper reframes AI in assessment as a validity problem rather than a cheating problem, arguing that rules which cannot be verified describe what should happen without showing what did.
2. VFWA separates conceptualisation capture from refinement, so that reasoning is secured at one authenticated point and the final product is read against it rather than in isolation.
3. Stage 1 requires four conditions: observed administration, an unseen anchor, ungraded sufficiency, and the widest feasible choice of language and modality.
4. Stage 2 is tool-agnostic: generative AI is permitted and optional, and the construct becomes the student's evaluative control over the mediation they choose.
5. Interpretation of the Stage 1 to Stage 2 link is a holistic judgment about whether both artifacts belong to the same intellectual project, not a checklist or a similarity score.
6. The conditional viva is bounded by design: it is triggered only when assessor confidence is insufficient on the artifacts alone, which is what makes the model scalable relative to universal oral assessment.
7. The authors position VFWA against product over-saturation, arguing that it aims for the minimum process evidence needed to restore interpretability rather than for portfolios that multiply artifacts without increasing evidential warrant.

Where it sits among other responses

The paper is careful about its neighbours. Oral assessment secures real-time reasoning but imports its own validity threats, since speech style, accent and confidence shape perceived credibility independently of understanding. Portfolio and process-heavy designs foreground development but add workload that pushes much of the work back into unobserved settings, so more artifacts do not automatically mean more warrant. Programmatic assessment supplies the underlying logic — multiple data points, low-stakes evidence generated for learning, holistic expert judgment — and VFWA is presented as that logic compressed into a single task.

Three illustrative configurations carry the model across disciplines: an argument map or recorded response to an unseen quotation in an analytical essay, a spoken reading of an unseen patient scenario before a clinical report, and a provisional issue analysis before a legal advice note. In each, the assessor looks for a recognizable intellectual thread rather than correspondence of content.

What this means for practice

• Start from one task, not the whole programme. Pick a single high-stakes written assignment, add an observed and ungraded capture of reasoning on an unseen prompt, and grade the final artifact against that baseline.
• Specify what Stage 1 must show, not what form it takes. Position on the question, warrant, competing interpretations and selection rationale should be visible whether the student speaks, outlines or draws.
• Keep the viva rare and narrow. Reserve it for submissions where continuity of question, stance or reasoning cannot be reconstructed, and ask how the argument developed rather than whether the student wrote it.

Limitations

• No empirical test. The paper reports no sample, no instrument and no measured outcome; whether Stage 1 artifacts improve assessor confidence or interpretive consistency is named as future research.
• Infrastructure is assumed. Multilingual recording, transcription, translation and secure capture at scale are prerequisites that not every institution can meet equally, and the authors call VFWA a design direction rather than a prescription.
• Workload is redistributed, not removed. Observed sessions, artifact management and occasional vivas move effort from post hoc investigation to upfront design, and that trade has not been costed.
• Sufficiency thresholds are unresolved. How a discipline decides that a Stage 1 artifact is enough without turning the stage into a second exam is left open.
• Construct boundaries are narrow. Where the learning outcome is mastery of written English, rhetorical performance or genre control, those should be assessed directly and securely rather than through VFWA.
• The viva depends on disciplinary judgment. The model concedes that interpreting complex evidence requires expert human judgment that cannot be automated, which places a capability requirement on markers that the paper does not measure.

Citation

Webb-Davies, K., Ratner, S., & Bassett, M. A. (2026). Voice-first written assessment (VFWA): Restoring evidence of reasoning through structural redesign [Preprint, version 1]. OSF Preprints. https://osf.io/preprints/edarxiv/tyh8z_v1