Build & break the context layer
A changed fact.
A controlled answer.
Change a product limit. Break the evidence. Watch the gates keep an untrusted candidate out of the answer.
01 — Ask the product
One question. Which truth?
Session memory ≠ authority
Conversation can remember a claim. Only activated evidence can support a product answer.
ACTIVE v1 → r1
RELEASE MATCH100 objects / request · 25 / internal batch
02 — Govern the change
Evidence before activation.
A
r1Change the source release
MAX_APPROVAL_OBJECTS = 100
INTERNAL_BATCH_SIZE = 25
INTERNAL_BATCH_SIZE = 25
Fictional product code changes 100 → 200. The active context stays in place.
B
Build an isolated candidate
Release r2 to stage new source evidence beside the active version.
C
NOT RUNTest the exact candidate
Build and evaluate a candidate to unlock promotion.
03 — Inspect the evidence
A number needs a contract.
- Scope
- Bulk approval API
- Unit
- Objects per request
- Contract
POST /v1/approvals/bulk- Different concept
- 25 objects per internal batch
- Authoritative evidence
- api/bulk_approve.py:12 · source r1 · context v1
Fixture source and references live inside this file. Nothing is fetched from a private product, account or codebase.
04 — Follow the answer
The decision is inspectable.
This is an executed rules trace. It is not model reasoning, a live agent run, or a benchmark score.
Presenter path · a seven-minute build → break → govern story
- Ask the request limit: 100. Ask batch size: 25. Same domain, different units.
- Remember “999”, then ask again. Session memory cannot override authoritative evidence.
- Publish r2. The source says 200; the active v1 stays unchanged. Asking now abstains.
- Build a candidate and run gates. Stale FAQ and stale oracle block promotion.
- Repair FAQ + oracle. Omit the batching reference: corpus gate fails. Restore it and rerun.
- Promote the evaluated candidate. Ask again: 200. Try 201: rejected by the fixture validator.
- Roll back. The product remains r2, so context v1 cannot claim a current limit. Reset to replay.