YASHKHOU/VERIFY GitHub ↗
VERIFICATION INFRASTRUCTURE / 2026

AI made creation cheap.
Proof is now the bottleneck.

Three open developer tools for checking generated artifacts, verifying agent-written code, and making autonomous actions recoverable.

01ARTIFACTgenerated.xlsxPASS
02CODEagent/pr-184VERIFIED
03ACTIONcrm.updateREVERSIBLE
01 / ARTIFACT VERIFY

CI for files, not just code.

Fail generated XLSX, PPTX, DOCX and PDF deliverables before they reach a client, board deck or finance model.

$ artifact-verify check model.xlsx
PASS · 7/7 checks
✓ office.structure
✓ xlsx.formulaErrors
✓ content.placeholders
Inspect Artifact Verify →
02 / REPO VERIFY

Agents don't grade their own work.

Turn a task into executable acceptance evidence: scope, tests, build, probes and a machine-readable verification digest.

$ repo-verify verify
VERIFIED · 5/5 checks
✓ scope.allowedPaths
✓ command:test
✓ command:build
Inspect Repo Verify →
03 / ACTION GUARD

Give agent actions an undo strategy.

Classify actions as reversible, compensable or irreversible. Add idempotency, approval gates, rollback and tamper-evident receipts.

await tx.run({
  classification: "reversible",
  snapshot,
  execute,
  rollback
})
Inspect Action Guard →
THE SHIFT

Generation is abundant.
Verification is scarce.

93%

of Claude/Cowork conversations in Anthropic's 2026 Economic Index produced an artifact.

~25%

average silent document corruption reported in Microsoft Research's long delegated-workflow study.

5/5

planted failure classes contained by the transparent starter benchmark. Small by design; inspect and rerun it yourself.

Run the starter benchmark ↗ · Sources: Anthropic Economic Index, June 2026 ↗ · Microsoft Research, DELEGATE-52 ↗. Statistics are source-attributed research context, not product claims.

Use the free tools.
Keep the evidence.

git clone https://github.com/yashkhou/verify.git
cd verify && npm install && npm test

MIT licensed. No account required. Start with the verifier closest to the failure mode you already have.