Guides Try the AnchorCore AI startup sample

Try the AI startup

The second demo corpus lives in the repo's sample-ai/ folder — clone the repo (or grab the folder from GitHub) and point a source at it, exactly like sample/. Release zips ship the app only.

A fictional AI product company ("Relay"): model routing, evals, pricing decisions and an incident postmortem — built to exercise what the payments corpus doesn't: policy-gated routing, threshold launch gates, and a metric that had to be corrected.

Connect it#

In the app → Library tab → Add source:

  • Connector: Local folder
  • Path: the absolute path to the sample-ai/ directory — click Browse… to navigate, or paste it (e.g. C:\Users\you\anchorcore\sample-ai)
  • Name: anything, e.g. "Relay"

Then click Sync now.

What's inside#

FileWhat's in it
sprint14.mdSprint goals, what was committed and carried over
people.mdPeople and who owns what
decisions/d1-model-router-v2.mdA tiered-routing decision with constraints
decisions/d2-eval-gate-thresholds.mdLaunch gates as absolute thresholds
decisions/d3-pricing-usage-tiers.mdA pricing decision with margin reasoning
decisions/d4-eval-set-contamination.mdA corrected metric that blocks a launch
meetings/m1-sprint14-planning.mdPlanning notes: actions, an open question, a stale number
meetings/m2-incident-postmortem.mdA postmortem: what happened, why, what changed
specs/model-router-spec.mdThe router spec with a numbered request flow
specs/eval-pipeline-runbook.mdThe eval runbook with thresholds and escalation

Two things in here are deliberate test content, not mistakes: the planning notes still quote the old eval number (0.94) that d4 corrects to 0.89, and the postmortem's open question about publishing an incident note was never closed. They exist so you can ask about them and watch the answer cite the correcting document, and try disputing a record in the Records tab.

Things to try (in the Ask tab)#

1. The 60-second wow

  • "What did we decide about model routing, and why?"
  • Expect: the tiered-routing decision with its reasoning, cited — 62% less spend on the lookup slice, policy constraints included.

2. Policy-gated answers

  • "Which tier can a tenant with a local-only policy reach?"
  • Answered from the router spec and the postmortem, not from guesswork.

3. A corrected number

  • "What is our citation accuracy on the reasoning slice?"
  • → 0.89 on the rebuilt holdout set, with d4 explaining why 0.94 was wrong and which launch gate it blocks.

4. Owners and sign-off

  • "Who approves a change to routing policy?" → Noor and Sam.
  • "Who signs off pricing changes?" → Ravi, per people.md and d3.

5. Postmortem follow-through

  • "What did we change after prompts reached the cloud provider?" → the three changes, each with an owner and a due date.

Source: docs/sample-ai.md in the public repo.