EvidenceGate
OpenAI Build Week · Developer Tools

Evidence in two lanes. One inspectable gate.

EvidenceGate checks what a repository proves, what authoritative sources currently support, and where those two domains disagree before AI-generated code ships.

Inspect the invariant

Same task. Different evidence.

Open each self-contained report, follow its official citations, and download the canonical JSON bundle behind the deterministic decision.

01 FAIL

Incomplete patch

The happy-path test passes, but required web-search, provenance, citation rendering, and negative-test evidence is missing.

Gate
fail: 1/14 required criteria verified
Source fixture
2 records · 2 bound citations
Bundle hash
3f7f66bef1f853761c629bfce3f86f60fbb5fcb8ee220fcaf25e12eef5c868a5
02 PASS

Corrected patch

Repository checks verify the implementation while validated official sources support the current external requirements.

Gate
pass: 14/14 required criteria verified
Source fixture
2 records · 2 bound citations
Bundle hash
6b11bd8461f8a5da2166d1dc2f863ca71ab3d6915257c549801d8b6c7fff1cf9
No-build judge path

Review it in three moves.

No installation, account, API key, or source rebuild is required for this path.

01 / FAIL

Open the incomplete report

See why a passing happy-path test cannot prove web search, source provenance, or safe clickable citations.

02 / PASS

Open the corrected report

Inspect verified repository evidence beside supported external-source evidence for all fourteen required criteria.

03 / VERIFY

Download the bundles

Compare the stored decision, source registry, assessments, model-run metadata, and canonical SHA-256 hash.

Internal evidence

What does this repository demonstrate?

Diffs, source analysis, commands, tests, and runtime probes establish implementation facts—but cannot establish whether an external API remains current.

External evidence

What do authoritative sources state?

Allowed official documentation establishes current requirements—but never proves that this repository implemented them.