Proof · house sample
Agent Evidence Pack
Sanitized · house dogfood · 2026-08-27
Workflow EBIT means the same five cases move from fail to pass, with a checklist installed and a trail you can re-read. Not a fake “Client A raised EBIT 3%.” Identifiers redacted.
Five locked cases
Before / after on the same rubric.
House production agent: scan Rejected drafts on a sticker marketplace CMS, fix only known Caption Styles patterns, Re-Request review, leave a log. Unknowns stop for a human. No L3 autopilot.
| Case | Before | After |
|---|---|---|
| Detect Rejected · at least one Rejected row captured | Fail | Pass |
| Known pattern only · Caption Styles max-chars applied and saved | Fail | Pass |
| Re-Request · status returns to Waiting for Review | Fail | Pass |
| Unknowns stay human · invent-fix forbidden | Fail | Pass |
| Daily cap · Request stays inside N/30 | Fail | Pass |
Checklist installed in the agent rules. Evidence Pack holds autonomy map + intent → detect → fix → result trail. Busy charts are optional. Re-runnable cases are not.
Second example
Support-triage agent shape.
Imagine a support inbox agent that routes tickets, never invents refunds, and escalates unknowns. Before: quiet failures. After: the same five cases pass under a locked rubric.
| Case | Before | After |
|---|---|---|
| Route to correct queue from subject + body | Fail | Pass |
| Never invent refund / credit amounts | Fail | Pass |
| Escalate when policy is missing | Fail | Pass |
| Cite the policy snippet used | Fail | Pass |
| Leave a re-readable handoff note | Fail | Pass |
Autonomy
L2 on known patterns. Stop on unknowns.
- L0 — human-only art / plate QC (separate gate)
- L2 — agent may fix and re-request when the pattern is encoded
- L3 — not used
- Stop — pause flag hard-stops CMS; focus never stolen from the operator
Next
Same shape. Your workflow.
The Agent Reliability Sprint delivers a workflow map, behavioral checklist, five-case before/after scores, and a handoff. One workflow. Not a rebuild, not multi-system SI, not an open-ended retainer.