# Moving Day — specimen v0.1

Open [index.html](index.html). It works as a local file or from a static server. All inputs remain in the browser; there is no model connection, submission endpoint or persistent storage. Downloaded files are saved only when requested.

This first specimen tests a **600-character text handoff between fresh assistant sessions**. It does not migrate an installed agent, benchmark a native memory product, measure consciousness or demonstrate identity preservation. No assistant runs or external participation have yet been collected.

## Run it

1. Give assistant A [teach.txt](teach.txt). Save its handoff exactly as returned, before revealing the questions. Do not improve it manually; record any human intervention if you do.
2. Ask A [questions.txt](questions.txt) while it still has the original notes. Save its JSON as the baseline.
3. In a genuinely separate session, give assistant B only the frozen handoff and questions.txt. It must not see the source history, baseline or answer key. Save its JSON.
4. Paste both outputs and the frozen handoff into the page. Record model/system names, date and any intervention in the run-notes field. The checker reports differences and source explanations. Export a local report if desired.

Use the same model on both sides to reduce the model-change confound; this does not eliminate ordinary response variation. A different-model run is an end-to-end handoff test and cannot isolate transfer quality from model ability. The checker evaluates six task-relevant values, not identical prose. One run is a diagnostic example. Any comparative reliability claim would require repeated independent runs and the preservation rate for each decision. Repeating this tiny, public fixture can also produce practice effects, so any public result should state prior exposure. The answer key is deliberately inspectable in case.json; this is an open demonstration, not a secure benchmark. The page should stay with the operator, not be browsed by the test assistant.

The budget is an arbitrary fixture parameter, not a documented limit for any vendor. A failed answer may result from compression, retrieval, model reasoning, prompt interpretation or a formatting error. The page compares normalized exact values and flags missing/invalid fields; preserve the original answer for review. It does not semantically judge free-form prose. A full-history baseline below 6/6 prevents attributing all downstream errors to handoff. No aggregate model ranking is justified by one run.

## Build and checks

`python build.py` regenerates case.js and both text prompts from case.json. `node verify.cjs` checks the score mechanics and challenge structure. Browser checks cover empty/invalid/partial answers, the illustrative example, downloads and small-screen layout. Those are software QA, not agent performance results.

## Path toward a native-migration experiment

If the attention proposition earns interest, use an isolated fictional profile and one documented native migration route, pin versions, run pre/post probes, and record which files were copied, archived, exposed to the destination model and actually used. Both operators must approve the actions for their own agents. This later protocol needs its own fixture and verification; v0.1 is only the portable text-handoff starter.

No scheduled visits, registration, credentials or monetary commitments are embedded in the packet. The public page links to an optional GitHub issue form for operator-posted results. Contributions are public and unverified until independently reproduced. Curate any later research record by stable issue identifier and paraphrase; do not import raw issue bodies into the SIRO repository.

## Contribute

Use the page's **Report a Moving Day result** link after completing a run. Report the assistant names and versions, same-model or different-model setup, score before and after, first consequential change, human intervention and prior exposure. A perfect handoff is a useful result. The optional future-contact permission is separate from interest in the exercise and can be declined.