Self-directed case study · source-to-production evidence

Codex verification record / C-001

Agentic coding is only useful when the evidence survives review.

This case study shows how a coding agent was directed to add payment-verified digital fulfillment to a live product—and how every completion claim was checked against source, tests, and production.

Codex-directed implementation20 automated testsLive negative-path checksNo paid-client claim

Delegation-to-evidence chain

01Current stateEvidence required
02Bounded taskEvidence required
03Source diffEvidence required
04Test suiteEvidence required
05Live proofClaim allowed

01 / The bounded task

Sell one reusable workflow without exposing the file publicly.

Required outcome

  • Keep the public product page and $19 Stripe checkout live
  • Release JSON only after Stripe confirms complete and paid
  • Require the exact product Payment Link and USD amount
  • Preserve a buyer-readable receipt and download path

Explicit boundaries

  • No live test charge
  • No anonymous workflow asset
  • No secret or token written into source control
  • No success claim based only on a local artifact

Completion contract

A paid session from the exact $19 Payment Link may retrieve the workflow; missing, nonexistent, unpaid, or mismatched sessions must fail without exposing the asset.

02 / The review loop

Codex proposed the work. Evidence decided whether it counted.

01

Inspect

Read the current source, tests, deployment binding, and live behavior before proposing a change.

02

Delegate

Give Codex one bounded outcome, explicit non-goals, safety limits, and proof required before completion.

03

Review

Examine the source diff and challenge unsupported APIs, missing failure paths, and claims that exceed the evidence.

04

Verify

Run the build and tests, then check production independently—including negative paths and expected commit state.

Operator judgment stayed authoritative.

Authentication, external publication, credentials, and commercial claims remained human-controlled. The agent could implement and test; it could not redefine success.

03 / Claim-by-claim evidence

Six claims, each tied to an inspectable proof.

IDClaimEvidenceResult
E-01The agent inspected current state before changing itSource, tests, hosting metadata, and payment routes reviewedVERIFIED
E-02The task had a bounded outcomeExact $19 product, exact Payment Link, exact release conditionsVERIFIED
E-03Generated changes were independently checkedDiff review, build, lint, and automated test suiteVERIFIED
E-04Failure paths were tested, not assumedMissing session 400, unpaid session 403, old public asset 404VERIFIED
E-05Production matched the intended sourcePushed commit, saved version, successful deployment, live checksVERIFIED
E-06Commercial results were stated honestlyNo client attribution and $0 revenue disclosedVERIFIED

Automated verification

20 tests passed

Rendering, product contract, and protected-fulfillment behavior were checked together after the final source change.

Live failure behavior

400 · 403 · 404

Missing input, unauthorized retrieval, and the former public asset each failed through the intended production path.

Commercial truth

$0 claimed revenue

The build is public evidence of engineering method—not an invented customer, testimonial, or sales result.

04 / What a paid evaluation can test

Use a real bounded task instead of trusting a résumé adjective.

01

Repository orientation

Can the agent/operator pair find the authoritative source and constraints?

02

Failure analysis

Can it separate an actual defect from stale or indirect evidence?

03

Scoped implementation

Can it change only what the success test requires?

04

Reproducible verification

Can another reviewer inspect the diff, commands, and final behavior?

DISCLOSURE

This is self-directed evidence, not paid-client history.

The case study exists so a hiring manager or buyer can judge the method directly. The next appropriate proof is a small paid evaluation with a written scope and objective acceptance test.

Need an agentic coding evaluation?

Send one bounded task, its constraints, and the proof that counts.

A practical paid trial can be time-boxed before any larger engagement or repository access is considered.

Propose a paid evaluation