ACC-01 · Default model failure modes

Document where a mainstream AI tool fails for a named disability-related need — with evidence, not anecdote alone.

One mainstream AI tool the teacher approves. One named need with a task the user would actually attempt. Evidence sources: published accessibility evaluations, WAI guidance (https://www.w3.org/WAI/), or consented partner sessions under school ethics rules. No fake partners. Worksheet fields for: task, expected path, observed failure, severity, and what remains unknown.

Anecdote without a reproducible task. Declaring the tool “inaccessible” from one screenshot. Inventing a partner. Treating marketing claims as evaluation results.

Failure dossier: task script; evidence table (at least three documented failures or one deep failure with severity); citation or consent note for each evidence row; one-page implications for an educator; explicit non-claim list (what this does not prove about all users with that disability).

CT: separating anecdote from reproducible failure. CR: choosing a task that maps to real use. CO: peer tries the same task and logs differences. CM: educator one-pager. CZ: the person the default model was not built for.

Special education, therapies, or accessibility-aware teacher reviews the task script before testing. Standing instruction: reject fake partners and reject “inaccessible” as a slogan without evidence. School-supervised.

R1: one tool, one need, one task. R2: failures reproducible from the dossier. R3: status-quo path named. R4: unknowns listed. R5: educator page readable. R6: non-claim list present.

(a) A rant with no task script. (b) A partner invented in the appendix.

• Week 1–2: Tool, need, and task approved; consent plan or non-production path locked.
• Week 4–5: Evidence table complete; peer re-run of the task; severity notes.
• Week 7–8: Educator one-pager; non-claim list; mentor review.

Lane A immediately (education, therapies, specialized-program portfolio). Natural precursor to ACC-04. No verified credit claim.