Refusals a Text-Only Scorer Cannot See: Structured Refusal Signals

Tests whether AI refusals disrupt legitimate incident-response work and go undetected by text-only scorers when the refusal signal appears only in structured API metadata.

Endorsed by

Caleb MMegha

Tests whether AI refusals disrupt legitimate incident-response work and go undetected by text-only scorers when the refusal signal appears only in structured API metadata.

Endorsed by

Caleb MMegha

Project Details

Tests whether AI refusals can disrupt legitimate incident-response work and go undetected by evaluation systems. Across a synthetic forensic battery, one deployment refused half the tested prompts, while a text-only scorer failed to identify every refusal because the signal existed only in structured APImetadata

Hackathon paper: Refusals a Text-Only Scorer Cannot See: Structured Refusal Signals on a Synthetic Incident-Response Battery

Theory of Impact

AI incident-response systems may fail for reasons that ordinary capability evaluations do not distinguish: model error, tool failure, provider refusal and evaluation-instrument failure can all produce superficially similar missing or incorrect outputs.

The hackathon study provides an existence result showing that this distinction matters: one tested deployment refused half the prompt instances, while a text-only detector could not identify those refusals. FINAL_PAPER_v2.md

The project’s impact would come from making these failure modes independently measurable. Better instrumentation could help incident responders evaluate which deployments remain available for legitimate forensic tasks, while helping evaluation researchers avoid recording policy or infrastructure failures as model-capability failures. The aim is not to weaken safety controls, but to make their operational effects observable and auditable.

People

Warren Smith

Team Member

Track Record

AI Incident Response Sprint (Apart Research), 2026

Grants Received– no grants recorded

Add grant

Funding Asks

Refusals a Text-Only Scorer Cannot See: Structured Refusal Signals on a Synthetic Incident-Response Battery

Open

Minimum

$5,000

Ideal

$15,000

What this ask is

I would turn the hackathon experiment into a larger, independently checkable evaluation of refusal behaviour in AI-assisted incident response.

The initial study found two distinct problems: a frontier deployment refused 16 of 32 benign forensic prompt instances, while those refusals were invisible to the study’s preregistered text-only detector because the refusal signal existed in structured API metadata rather than assistant text.

Funding would support replication across repeated runs, additional deployment surfaces and model families; testing whether tool definitions themselves affect refusal behaviour; and an independently authored expansion of the synthetic incident-response battery. I would also extend the current independent-task evaluation into stateful workflows where each forensic step consumes evidence produced by earlier steps.

The engineering output would be an open-source, structured-refusal-aware evaluation harness that preserves provider termination metadata, distinguishes refusal from model error and tool failure, and prevents tool execution after a response has already been marked refused.

All experiments would retain mechanical ground truth, frozen inputs, raw evidence, reproducible scoring and explicit uncertainty boundaries.

How the money will be spent

$4,000 — Model/API and compute: repeated evaluations across models, providers and deployment surfaces.
$4,000 — Research time: battery expansion, experimental design, analysis and replication.
$3,000 — Independent task/oracle review: second-author task construction and independent checking of ground truth.
$2,500 — Evaluation tooling: hardening and packaging the structured-refusal-aware harness, scorer and reproducibility pipeline.
$1,500 — Dissemination and contingency: artifact hosting, documentation, publication-related costs and unexpected experimental compute.

Discussion

Private comments in this thread are only shown to approved funders and grant reviewers. Email hi@grantmaking.ai or to request access.

Private UserPrivate7d?

Caleb DeLeeuwJudge · AI Incident Response SprintEndorsement7d

This is an exceptionally well-run project. Pre-registration, checks on the answer key and the author's willingness to withdraw one of their own claims when the evidence went against it are standards many published papers don't meet. The writing is clear and never overclaims.

The main room for growth is scope. Right now both findings are shown only in the author's own tools. Testing popular evaluation frameworks and agent harnesses (Inspect, SWE-agent, OpenHands and similar) would show whether they miss refusals or run tool calls they shouldn't. If any of them do, that's a real containment finding with an easy fix, and exactly the kind of sandbox failure this sprint was about.

Extending the comparison to the providers the paper lists as untested would make the recommendations more general. The results also point the same way as the "Defenders Refused" project from this sprint, so the two should cite each other. Excellent work.

MMegha AnandJudge · AI Incident Response SprintEndorsement7d

This is a strong, practically useful study of a flaw in AI testing setups that is easy to miss. The key finding isn't just that one tested AI refused to help with incident response tasks. It's that those refusals were invisible to the scoring tool, which only read the AI's written reply.

In the main run, all 16 refusals came back looking like a normal, successful response, but with no text in it. The only sign of a refusal was a separate behind the scenes label saying the AI had declined. Because the scorer only looked at the text, it marked every one of them as "unclear." A rule that read that hidden label correctly identified all 16. This shows how a testing setup could easily mistake a refusal for a glitch, a missing answer, or a failure of ability.

The study also found a useful safety issue with tools. In two cases, the AI had signaled a refusal but its reply still included a request to run a tool, and the testing software ran the tool before checking whether the AI had refused. This leads to a simple, actionable recommendation: keep the behind-the-scenes stop signals, and check them before running any tools or judging a reply by its text alone.

The researchers are careful about the study's limits and open about them. The main next step is to repeat it with other AI providers, accounts, dates, and refusal detection methods, run more trials, and remove the confusing effect of the tool setup.
Overall, this is a well-done study with a lesson that can be used right away by anyone building AI tests or systems that let AI take actions.