Grounded draft
Every model-extracted step retains a bounded source reference and stays unconfirmed until reviewed.
Submitted Build Week prototype · Education
Eligibility is not access.
AccessCrash turns supplied process instructions into a source-linked path, asks a human to confirm it, then tests whether a capability-constrained person can actually reach the promised outcome.
StatusSubmitted on Devpost; public signed-out synthetic judge path and source available.
ProblemA process can declare a person eligible while still demanding capabilities they do not have.
MechanismSource → unconfirmed graph → human confirmation → capability twin → path test.
AuthorityGPT-5.6 compiles; the human confirms; deterministic code decides reachability.
Problem and mechanism
A second human profile can reveal a failure the normal path hides.
The same eligibility assumption is tested against two capability profiles. AccessCrash reports the first unavoidable barrier and the alternative routes that also fail, so a process owner can see the specific requirement breaking access.
- CompileGPT-5.6 turns bounded instructions into a cited, unconfirmed process draft.
- ConfirmA human reviews every step before the graph becomes assessment input.
- StressA capability twin traverses conditions and alternatives under the same external eligibility.
- CompareBefore-and-after process evidence shows whether the barrier was actually removed.
Verified public evidence
The complete public loop stays inspectable.
The public site uses a visibly labeled bundled synthetic graph. Live GPT-5.6 compilation remains disabled until authenticated identity and persistent quota controls are in place.
What the work demonstrates
Accessibility can be tested as process reachability.
The prototype translates a vague concern into a concrete graph, named capability assumptions and a repeatable before-and-after test.
Explicit confirmation
The reviewer owns the process model. The interface does not hide that responsibility behind AI confidence.
Deterministic reachability
The engine follows the confirmed graph and capability profile; the model has zero outcome votes.
Authority boundary
AccessCrash tests the supplied model of a process, not a person's eligibility.
It does not approve benefits, make legal or clinical decisions, certify accessibility, independently verify citations or prove that an institution follows its written process. Its result is only as complete as the confirmed graph and capability assumptions.
Model mayCompile bounded instructions into a cited, unconfirmed draft.
Model may notConfirm the process, decide eligibility or set the crash-test outcome.
Build Week 2026
See the complete bounded-authority collection.
ProofLatch governs release evidence, Gatehold governs agent clearance and ExitCanary tests SaaS exit fidelity.