389 candidates went into a funnel. Zero came out admitted.
This is the full corpus readout: cluster-by-cluster review, candidate packets, license/provenance rules, the Horizon pilot preflight, and the exact reason “cool repo” was not treated as “safe building block.”
Each reviewed disposition has a source path, canonical repository URL, reason, and next gate.
That is a useful evidence discipline. It is not the same as a build receipt. No candidate checkout, build, browser run, screenshot baseline, license scan, owner, or rollback manifest was completed for the pilot.
Why 389 is still light
The corpus was reused from an existing merged sweep, not restarted. The external 850k-repository data plane and local catalogue database were not mounted in this run. The next sweep should be per industry and capability, not one undifferentiated “AI builder” query.
Interpretation: 389 is a reliable current funnel count, not a claim about all useful GitHub code.
02 / Review history
What the agents did inside the corpus
Six review clusters completed; cluster 007 planned.
Cluster 001 · 15
React admin-dashboard lane. Every record held; the core issue was missing admission proof rather than no visual value.
Schema, migration, database-to-API slices. JustVugg/dbcli emerged as a read-model sidecar candidate; write surfaces and package/import mismatch kept it held.
Horizon admission execution plan and read-only preflight. Eight-point-two GiB and Playwright were available; scanner, pinned checkout, fixture host, owner, and rollback were missing.
Run G1–G6 on a disposable pinned checkout with synthetic fixtures.
JustVugg/dbcli
Potential read-model discovery sidecar for a constrained data plane.
Write surface and package/import mismatch; safe wrapper not yet implemented or tested.
Read-only adapter contract, fixture database, denial test, and schema receipt.
Ducksss/payload-components
Strong registry-contract reference with manifests, provenance, recovery fields, and documented build/smoke gates.
Payload-specific; useful as a pattern, not a drop-in Actionist block.
Extract the manifest/registry contract without importing unreviewed runtime assumptions.
Rejected, not merely held: Schema-Weaver/pg-migration-engine is BSL-1.1; production use requires a commercial license. The packet keeps it reference-only.
04 / Admission plan
The seven gates the pilot never reached
Cluster-007 is a plan and preflight, not execution.
G1
Source + license provenance
Pin the commit, record license and obligations, scan source/dependencies/assets, and preserve the receipt.
G2
Extraction + adaptation
Record what is copied, wrapped, transformed, or discarded; map ownership and token/registry slots.
G3
Deterministic build
Install from the pinned source, run typecheck/build/tests, and capture the output.
G4
Browser smoke + write denial
Prove the candidate renders and cannot create unintended external side effects in the fixture environment.
G5
Screenshot baseline
Compare the adapted block against the approved visual contract, not just “it loaded.”
G6
Owner + rollback
Name the accountable owner, rollback path, data boundary, and recovery behavior before admission.
G7 / final verdict
No gate receipt, no admission.
The plan explicitly stops before the final verdict when prerequisites are missing. That is why the research produced zero admitted blocks without being “empty.”
05 / Evidence shelf
Open the underlying corpus work
Local source links are locked on the public host.
currentAM-CORPUS current state
All proven counts, pilot disposition, evidence gaps, and next action.