Public evidence index / Agent Job Output Review
Agent Job Output Review & Contract Kit
A fixed-scope review of up to five recurring AI-agent or automation jobs. Each job is examined against one question that process exit codes cannot answer: did the run produce the useful artifact it was supposed to produce, or did it merely finish?
Proposed price: $349, one-time. There is no checkout on this page and no payment is taken here. The first step is a text-only scope worksheet, not an order.
Who this is for
Engineering leads and founders running scheduled agents, collectors, enrichment jobs, reconcilers, or report builders that report success on a cadence. The common failure is not a crash. It is a job that exits zero every night for weeks while emitting a template, a stale copy of yesterday's file, or a change that nothing downstream ever accepted. Dashboards stay green because the dashboard is watching the process, not the artifact.
This engagement produces the artifact-level definitions and tests you are missing, in a form your own systems can execute.
What you receive
- A dated PDF findings report covering each reviewed job, with a status, the evidence chain behind that status, and the specific gaps that prevented a stronger conclusion.
- Up to three draft machine-readable useful-work contracts in JSON: what the job claims, what artifact counts as proof, what honest idle looks like, and what states are refusals rather than passes.
- A Python standard-library test harness, shared as a template and tailored to your contracts, with synthetic fixtures written for your job shapes.
- A source bundle of everything above, delivered within 5 business days after complete intake.
Being direct about what is and is not paid work: the harness template is shared and is not the value by itself. What you are paying for is the customer-specific claim-to-evidence mapping, the refusal reasoning, the contract definitions, and the fixtures. You validate, deploy, and operate everything. Nothing delivered can start, stop, pause, disable, or schedule a job. There is no actuator in this kit.
Two questions, kept separate
Most job monitoring collapses two distinct questions into one boolean. The review keeps them apart:
- Did the job run? Answered by process state, run identity, timestamps, and exit status.
- Did the run prove the promised useful artifact? Answered only by an observable downstream artifact tied to that specific run.
Status taxonomy
| Status | Meaning |
|---|---|
| USEFUL | The run produced the named artifact, and evidence ties that artifact to this run. |
| HONEST IDLE | There was legitimately nothing to do, and the job emitted a distinct signal saying so. |
| DEAD WORK | The job did something and finished, but no accepted downstream artifact resulted. |
| ERROR | The job failed and said so. |
| UNKNOWN | The evidence source could not be read, so no claim can be made. |
UNKNOWN is counted and reported as UNKNOWN. It is never converted into idle, a pass, a failure, or an automatic strike. A changed output is not automatically useful. A completed process is not automatically useful.
Intake preflight: the contractability test
Before any paid work is accepted, each candidate job is checked against a published bar. A job is contractable only if all six hold:
- A named intended useful artifact — what this job is supposed to produce, stated concretely.
- An observable evidence source where that artifact can be seen.
- A distinct honest-idle signal, separable from success.
- An unavailable-evidence state that is distinguishable from idle.
- Stable run identity and timestamps, so an artifact can be tied to a specific run.
- Sanitized representative outputs plus proof that the downstream artifact exists.
Jobs that fail this bar receive a NOT CONTRACTABLE row in the matrix with the reason. They do not receive a guessed contract. If no submitted job passes, no paid engagement is accepted and you keep the preflight result.
What the initial request form accepts
Only: job count, languages, schedule, intended artifact types, evidence types, and sensitive-data constraints. The form accepts no uploads. Do not send logs, code, configuration, credentials, output samples, or customer data at this stage; that material is rejected rather than stored.
Handling of material you send later
After written scope acceptance under the normal customer agreement, you supply sanitized excerpts for accepted jobs only.
- We never ask for credentials, secrets, personal data, customer content, production access, or full logs.
- We cannot claim personal data is incapable of arriving. If an accepted evidence package contains any of the above, review pauses and a sanitized replacement is required before work continues.
- Accepted working files are deleted 30 calendar days after delivery. Final deliverables remain with you. Ordinary scope and billing records are retained.
Sample matrix
Three fictional jobs, shown at the level of detail the report uses.
| Job | Observed | Evidence check | Status |
|---|---|---|---|
| Lead research agent, nightly | Exit zero; output file changes every run | No downstream accepted lead receipt exists for any run in the sample window | DEAD WORK |
| Filing collector, hourly | Exit zero; empty result set | Official source request timed out; the job cannot distinguish "nothing published" from "could not ask" | UNKNOWN |
| Ledger reconciler, daily | Exit zero; wrote a reconciliation file | Downstream receipt carries the run id and a timestamp after run start; artifact is fresh, not a copy | USEFUL |
Refusal example
A summarization job emits free-form log lines and nothing else. There is no artifact store, no receipt, and no consumer that records acceptance. Result: NOT CONTRACTABLE — no observable evidence source; success and idle are indistinguishable in prose logs. We do not write a contract for this job. The report states what would have to exist first.
Harness fixtures
Each contract ships with synthetic fixtures covering, at minimum:
- Useful success with a tied downstream artifact
- Honest idle
- Nonzero error
- Stale or byte-identical repetition of a prior output
- Placeholder or template output
- Unavailable, permission-denied, timeout, and malformed evidence
- Changed output with no proven downstream artifact
These fixtures validate that the contract classifies each case correctly. They say nothing about your running fleet. Only your own execution against real evidence can do that.
Owned method, bounded claim
This method comes from controls we run on our own recurring jobs, which distinguish useful work, honest idle, dead work, errors, and pauses. Negative controls in that system have caught exit-zero outputs that looked healthy while repeating a prior artifact or concealing an unavailable probe.
That is the whole claim. It is not evidence of external demand for this offer, it is not a certification of any kind, and it is not proof about your fleet or your jobs.
Contract starter
A free browser-local tool that generates starter JSON and a test matrix from a few fields. It runs entirely in your browser, makes no network calls, and uploads nothing. It cannot inspect or validate any real job, and its output is a starting point rather than a finding.
Runs only in your browser. It writes a draft contract and made-up test cases. It cannot look at a job, and its output is not a finding.
Ready. Fill in the fields, then build.
Draft contract JSON
Synthetic test matrix
Framing sources checked 2026-08-09
- Google SRE: Monitoring Distributed Systems
- OpenTelemetry generative AI semantic conventions
- Python unittest documentation
These inform vocabulary and structure. They are framing sources, not endorsements of this offer, and none of these projects is affiliated with it. Source pages can change and are checked again when a customer scope is prepared.
Scope and next step
Fixed scope: one engineering team, at most five recurring jobs reviewed, at most three draft contracts. Delivery within 5 business days after complete intake. A green process result is not evidence of a useful artifact, and unavailable evidence is never a pass.