Three recorded evidence examples
The reader preserves three actual public-service observations from 21 September 2026. Opening an example does not submit a new question or call an AI. Start with the seven-step learning guide.
| Example | Recorded selection | What it demonstrates |
|---|---|---|
| Current care-home question | 55 records, 115 relationships, 523,326 bytes | Inspect the question, source passages, partner qualifications, provenance and gaps. |
| Unknown-term control | No records or relationships; 4,385 bytes | Missing evidence stays visible. There is no answer to manufacture. |
| Original historical care-home package | 35 records, 50 relationships, 257,105 bytes | An older package keeps its exact identity even after the current assembler changes. Its original engine was not recorded; the replay engine is separately identified. |
All three packages report insufficient evidence and contain no AI answer. The public SDK receipt records exact complete-package reconstruction. The service report preserves the earlier failed attempt and explains the corrected SDK comparison.
Why the files are small
A catalogue lists the selected records. Choose a record to retrieve its complete text and metadata through bounded parts. The reader verifies their hashes before presenting the complete value. Relationships and diagnostics are separate streams. The complete original package is also available for a client that can accept its size.
The 225-file archive contains 2,766,291 bytes. The browser does not download every file when the page opens. The publication declaration binds the approved cases, exact original inputs and exporter files. A hash match establishes delivery integrity, not current law or correct interpretation.
Publication boundary
These fixed public questions are explicitly approved demonstration examples. The service does not retain anonymous visitors' questions. This archive is not an AI-answer audit log, a new evidence assembly or a finding that the evidence is complete. The retained public SDK check and the separate website publication check have different scopes.
The public browser procedure checks the actual reader after publication. Its nine offline controls do not by themselves establish a live browser pass.