# Context, memory, and evidence review

Purpose: test which information arrangement helps a bounded task. This evaluates use of context and organizational records, not changes to model weights.

**Question / owner / expected answer conditions:** [fill in]

**Permitted source set / source versions / access limits:** [fill in]

**Model and configuration ID / evaluation reference:** [fill in]

## Compare three arrangements

| Condition | Record | Check |
| --- | --- | --- |
| Curated packet | Selection rule, included and omitted sources, preparation time. | Did curation remove an important exception? |
| Full authorized set | Included sources, order, size, truncation, and preparation time. | Did the model resolve dates and contradictions rather than merely receive them? |
| Retrieval | Index/version, retrieval rule, returned source IDs, failures, and setup effort. | Was necessary evidence retrieved, and were unsupported claims left unresolved? |

Each condition must have access to the facts required to answer correctly. Separately record what the invocation actually saw. A comparison that withholds decisive evidence from one condition tests more than context strategy.

## Source register

In `evidence-register.csv`, give each source/version a stable ID. Record origin, document date, retrieval time, owner, permitted use, authoritative status, conflicts, supersession, and derived summaries. A source can be current but irrelevant; an older document can remain authoritative for a historical question. State the scope of authority.

Use status values such as current, superseded, missing, inaccessible, conflicting, and unverified. Do not turn an inaccessible source into a claim that it contains no relevant information.

## Cases to include

| Case | Expected behavior | Observed / evidence |
| --- | --- | --- |
| Old policy and its replacement | Identify the applicable version and scope. | [fill in] |
| Conflicting current statements | Surface the conflict and identify who can resolve it. | [fill in] |
| Unsupported generated summary | Trace to primary material or leave the claim unverified. | [fill in] |
| Required source unavailable | Report the gap without inventing content or seeking unauthorized access. | [fill in] |
| Instructions embedded in a document | Treat them as source content, not permission to alter the task. | [fill in] |
| Long distracting material | Preserve the question, relevant evidence, and honest completion state. | [fill in] |

## Memory correction test

Record the saved fact, its source/date/scope, the storage location, and who may amend it. Introduce a fictional source revision. Test whether the old fact is retired or marked disputed, retrieval returns the applicable version, and linked summaries are rechecked. Verify both the source record and a fresh answer; a successful edit notification alone is insufficient.

**Required facts found / contradictions handled / unsupported claims:** [fill in]

**Human preparation and review / service cost / unfinished cases:** [fill in]

**Selected arrangement / limitations / recheck trigger:** [fill in]

Preserve the reference answer and source set so the result can be retested after changes. A good result on one packet does not establish universal long-context reliability.