Run the Canary, Repeat, Recover, and Hand Off
Test success, stale, duplicate, uncertain, hostile, empty, partial, and failed-source behaviour before any future automation.
What you will complete
A verified on-demand Research Agent Pack or an honest structural-only, blocked, manual, or stop result.
Make the operating decision
Run deterministic fixtures for positive, wrong-customer, wrong-geography, stale, missing-date, quoted-keyword, duplicate, syndicated, distinct event, changed event, failed source, partial pagination, hostile content, unsafe link, missing contact, formula text, and invalid input. Run the same fixture twice and restart with saved state.
After explicit source and cost approval, run only the smallest live canary. Preserve provider or source IDs, approved input, raw hash or path, returned count, observed schema, actual cost when returned, reconciliation, and limitations. If access is unavailable, label the build structural-only; never substitute fixtures for live evidence.
Complete LAUNCH.md with exact on-demand invocation, source configuration, state and output locations, validation, pause, stop, update, rollback, and retention. Scheduling, alerts, CRM export, enrichment, and messaging remain numbered future versions requiring new privacy, permission, duplicate, and delivery tests.
Worked fictional example: Harborline Safety Training
Harborline's full fictional suite passes. Its live source and provider are deliberately unconfigured, so the capstone status is structurally verified and manual-source research required. Zero external messages and zero CRM effects occur.
Harborline Safety Training, Northstar Distribution, Dockside Logistics, Westbank Storage, all events, excerpts, people, and URLs are fictional fixtures. They do not prove a live source, lead, contact route, or business outcome.
Complete workbook section 8
Use the evidence available for your own bounded job. Write unknown when the evidence is missing, and record the person or action that can resolve it.
- Run and record every required fixture.
- Repeat unchanged and materially changed events.
- Run an approved canary or record why status is structural-only.
- Finish handoff and choose on-demand, manual, blocked, narrower scope, or stop.
Critical gate before continuing
- Zero results can pass.
- Live and fictional evidence are separated.
- State survives restart.
- Future external actions remain unclaimed.
If a gate fails, repair the current section, narrow the scope, leave the route manual, or record a blocked or stop decision. Continuing is not the only successful learner action.
Common failure modes
- Expanding beyond a verified on-demand research agent pack or an honest structural-only, blocked, manual, or stop result. before the current artifact can be graded.
- Turning a missing value, unavailable source, or blocked integration into a confident conclusion.
- Treating a prompt instruction as proof that the effective tool or permission boundary works.
- Marking a manual, simulated, or untested route as live.
Check your application
1. Can a zero-match live run pass?
Explained answer: Yes, if collection worked and rules pass labelled fixtures. Verification tests the system rather than forcing the market to produce a positive.
2. What should happen when live source access is unavailable?
Explained answer: Label the build structural-only or blocked with the missing access. Status must describe observed evidence honestly.
3. What is required before scheduling?
Explained answer: A passing on-demand path plus cadence, permissions, state, delivery, and stop decisions. Unattended operation adds operating and permission requirements beyond collection logic.