Repository navigation
Validation and Evidence
OpenSocrates separates source implementation, package verification, live host delivery, complete reading, actual application, and useful outcomes. A green offline suite cannot establish the later layers.
| Evidence | Supported conclusion |
|---|---|
| Versioned source and deterministic checks | The named contract is implemented and checked at that commit |
| Native package/build/installer checks | The exact platform package and covered lifecycle behavior work in the tested environment |
| Host registration or account upload | The package is registered, or its upload format was accepted |
| Live entry/loading observation | The named host surface delivered the observed entry/content |
| Complete canonical read | The exact procedure was read at the stated evidence level; agent assertions remain labeled |
| Application and artifact review | The method was reported or observed as used, and the particular result was assessed |
| Quality/reliability evaluation | Only the tested protocol, cases, host versions, and measurements support the claim |
No row substitutes for another. Keep unknown/unavailable states explicit. CLI and Desktop Code, web Chat and Desktop Chat/Cowork, and Antigravity workspace/global conversation-app cells remain separate. An IDE observation does not prove conversation-app loading, and another account is not the same cell.
The default native decision path reports read acknowledgments as
agent_reported and keeps applied: unverified. Native Claude's stateless entry
has no transcript reader, Stop repair, or native application receipt. Account
content also has no native receipt. These instrumentation limits do not create
an additional completion gate: finish the requested artifact when its actual
requirements, full reads, and applicable method conditions are met.
The exact footer Powered by OpenSocrates is conditional on an eligible method
being read completely and actually applied. It is model-reported attribution,
not independent application evidence, a quality certificate, or a reason to
display it for reader guidance alone.
Record the exact commit, commands, environment, observed results, failures, checks not run, and remaining live limitations. Native Mac and Windows CI are platform-specific. Clean-machine acceptance and purge/reinstall on a previously used machine are different results. Preserve failed practical observations when a guide changes; source checks do not regrade those earlier outputs.
- Mac observations
- Windows observations
- Clean-machine acceptance
- Purge/reinstall acceptance
- Codex SessionStart timing
- Decision-point evidence contract
Earlier Claude selector, payload-read, hook, Chat, and Cowork receipts remain
versioned historical records under docs/. They describe their named older
versions and do not qualify the new v1.5 profiles. AI-assisted adjudication is
not human gold, held-out truth, or causal proof of answer-quality improvement.
Receipts and reports must not retain raw private prompts, transcripts, credentials, screenshots, workspace content, or hidden reasoning. Commit release-critical procedures and sanitized evidence to the repository; a Wiki revision alone is not their durable record.