Reliability Retainer

A green dashboard proves a job returned. Never that it did the work.

A scheduled job can report success five runs in a row — exit code 0, dashboard green — while the process it was supposed to run has been dead for days. The lie isn't in the bot. It's in the wrapper. This is the evidence discipline that catches it.

Report an incident → Book the Week-1 audit → See the packages

15 minutes, or skip the call: I'll send the Week-1 scope doc. Either way I can show a wrapper lying, live, in under two minutes.

nightly_job.log — 5 consecutive runs

$ tail -f nightly_job.log

[run 1] wrapper exit 0 — status: SUCCESS

[run 2] wrapper exit 0 — status: SUCCESS

[run 3] wrapper exit 0 — status: SUCCESS

[run 4] wrapper exit 0 — status: SUCCESS

[run 5] wrapper exit 0 — status: SUCCESS

[freshness-check] payload artifact age: 4d 02h — expected <1h. STATE: RED.

$ cat RELIABILITY-AUDIT.md

Real findings. Names stripped, mechanisms kept.

CRIT — dishonest exit code

A log handle inherited by a background server bricked every future run of a wrapper in under a second. The success ledger got written anyway. No exit code caught it, no heartbeat caught it — output just quietly froze for four days.

HIGH — fail-open restart loop

A trading bot's restarter faithfully relaunched a process an exhausted API quota kept killing — 187 times in a single day, each one logged as a normal restart, not an incident.

CRIT — silent degrade-to-default

A pipeline ran "healthy" for five weeks on default inputs because its real data sources had died. Its own graceful-degradation logic was, by design, hiding exactly that.

HIGH — unverified attribution

A paid "speed edge" data feed, once actually audited for attribution, had fed 0 of 20 trades. The free fallback fed all 20. The dashboard never distinguished the two.

For teams running scheduled jobs, wrappers, or agent automation in production, where a false "success" costs money before anyone notices. Nothing runs unattended? Not for you yet.

$ ./retainer --phases

Three phases, each sellable on its own

WEEK 1

The Audit

from ~$5k, standalone

$5k is the first-engagement rate, priced to earn the first testimonial. After the first delivered engagement this bundle is $8–15k.

Every scheduled task / wrapper / agent job swept against a 6-principle hazard catalog
Every finding: file, line, severity, hazard class
Live-evidence appendix for every CRIT
Ranked fix-lane plan (doubles as the Month-1 quote)

Acceptance: every CRIT reproduced with evidence you verify yourself.

MONTH 1

The Harness

$5–50k fixed, scoped by the audit

Wrapper repairs: honest exit codes, sentinel-gated writes
Freshness asserts at every job's output seam
Identity-verified instance guards, fail-closed
A status surface: alive vs. producing vs. inputs-fresh

The gate: we kill your payload, your wrapper still exits 0 — the surface goes RED anyway.

ONGOING

The Watch

$2–8k/mo

Scheduled freshness sweeps + composite liveness checks
Regression run on every material change
Monthly report: incidents caught, time-to-detection, pass/fail — not vibes
Paging on state transitions only, tiered. Never level-spam

Detection is read-only. Remediation is opt-in, separately scoped.

$ whoami

Who's running this

Jaime Bell. I run a fleet of ~60 trading/agent repos: every incident above is mine, found on my own fleet before any client's.

github.com/jaimenbell · jaimenbell123@gmail.com

$ grep -v "in scope" SOW.md

What this explicitly is not

No trading/business logic. We never touch what the agent decides — only whether its runs are honest about themselves.
No model quality. We maintain the eval/regression machinery; labels and prompts stay yours.
No infra ops. No on-call, no cloud cost work, no CI/CD ownership beyond the reliability checks themselves.
No auto-remediation. The watch detects and pages. It never restarts, kills, or repairs anything itself.
No prevention guarantee. The promise is honest, fast detection — not zero incidents.
Scoped to the harness repo. Adding a job to the watch is in scope. Building the job is a new quote.

Recognize one of these?

If an agent or scheduled job of yours has done something in the same family — reported success it didn't earn, asserted something that wasn't true, or quietly went dark behind a green light. Tell us about it. No pitch on the first message.

Report the incident → Book the Week-1 audit →

15 minutes, or skip the call: I'll send the Week-1 scope doc. Either way I can show a wrapper lying, live, in under two minutes.