← Selected work
OpenAI Build Week 2026 · four submitted projects

AI may reason. Authority stays bounded.

EV1 Labs built four working systems around one design discipline: use GPT-5.6 where language and ambiguity matter, then make the decision boundary explicit, testable and visible.

01 · EvidenceEach claim points to a public app, repository, video and submitted entry.
02 · BoundaryThe model role is narrower than the product's final operational decision.
03 · FailureFallbacks remain labeled; model failure cannot silently upgrade the result.
04 · ControlA human can inspect the input, authority and current prototype limit.
The collection

Four operational failures worth making visible.

The projects are separate submissions, not four skins on one product. Their shared architecture is a deliberate answer to overconfident automation.

Shared engineering pattern

Interpret, confirm, decide, re-run.

The exact sequence differs by product, but the authority never hides inside a paragraph generated by the model.

  1. 01 · BoundConstrain the input.

    Use a strict evidence packet, synthetic fixture or owned local state rather than an open-ended prompt.

  2. 02 · ReasonGive the model one job.

    Explain, map, compile or identify semantic overlap under a structured contract.

  3. 03 · DecideKeep authority explicit.

    Deterministic policy or an identified human checkpoint owns the consequential result.

  4. 04 · ProveGenerate new evidence.

    A changed commit, export, process or lane must be evaluated again rather than assumed fixed.

Honest collection boundary

Submitted prototypes are not universal guarantees.

The collection proves four implemented, public and reviewable product theses. It does not prove market adoption, complete category coverage, trusted evidence origin or real-world outcomes beyond each documented test boundary.

Public proofApp, source, narrated demo and Devpost submission for every project.
Still openProduction integrations, broader field evidence and sustained user adoption.
Apply the pattern

Have an AI workflow where the final authority is still unclear?

Send the decision, evidence and failure that matter. The first answer will separate what a model may interpret from what the system or a person must own.