# Evidence before the answer — free technical clinic brief

AI-native training by LockedIn Labs · Edition 1, 28 October 2026

A 120-minute facilitator plan for developers and engineers. Use the [retrieval-limit workshop](https://training.lockedinlabs.ai/training/guides/retrieval-limits) and its original Python runner. Prepare Python 3.11+ with SQLite FTS5 before the timebox. The exercise uses synthetic, in-memory rows and makes no model, account or network call. Setup, reading and breaks are additional.

## Before the session

Each participant predicts the difference between missing evidence, inaccessible evidence and a limited candidate population. Record the actual Python and SQLite versions. Read the source and use a disposable folder. The exercise's trusted fixture context is not real authentication or physical tenant isolation.

## Run the clinic

| Planned minutes | Activity | Evidence to retain |
| --- | --- | --- |
| 0–15 | Agree the question, permitted source population and context admission rule. | Own baseline prediction and a simple source map. |
| 15–40 | Run post-filter, overfetch and pre-limit. Locate actual ranked candidates and admitted context. | Exact command, query population and returned IDs. |
| 40–60 | Change only the result limit. Predict before running. | Separate original and changed reports; no rewritten reference. |
| 60–85 | Inspect revoked and corrected sources. Explain what the final read withholds. | Actual grant/revision decision and retained context. |
| 85–105 | Swap builder and receiver. Receiver reproduces a withheld-source case. | Each person's actual action, evidence path and assistance. |
| 105–120 | Identify the next bounded real-store or provider test. | Owner, unresolved concurrency/refresh question and acceptance evidence. |

An empty final result does not prove the corpus has no eligible source. Overfetch does not grant permission or guarantee full coverage. A current-source gate does not establish that a generated answer is supported. This sequential SQLite experiment does not establish approximate-vector-search behavior, concurrent revocation safety or production authorization.

## Individual record

```text
Synthetic alias / role:
Runtime / exact file / command:
My prediction before the original run:
Candidate population / limit / actual candidate IDs:
Admitted context / source gate / evidence path:
One changed input / separate prediction:
Observed change / unchanged permission boundary:
My contribution / assistance:
Receiver action / evidence found / missing evidence:
Next real-store/provider test / accountable role:
```

Keep observations, predictions and suggested repairs separate. A successful local run is practice evidence, not a production assurance or field qualification.