Chart
Payer criteria
Draft appeal
◌ Nothing written yet. The gate stamps each sentence as it arrives.
0of 0 entries read
0of 0 sentences verified
$0.000this appeal cost
0.0 smodel time elapsed
◌ Nothing written yet. The gate stamps each sentence as it arrives.
Fifty synthetic denials per run, effort low, through the same deterministic gate. Agreed with the key means the model's warranted decision matched the grading label and the gate passed or declined accordingly.
Agentic rows: the model reads the chart by tool call from an index instead of receiving it whole; entries read is the mean per case against the mean chart size, and key entries read is the share of grading-key evidence entries the model opened. Gate passed, key says no counts drafts the gate passed on a case the key labels absent: every sentence was in the record, but the record did not satisfy the criterion, which is the referential limit the physician's signature covers. A live CHART row adds a third tool, a query against the author's pre-existing coverage index for when the authority's criteria last changed; its answer appears under the criteria panel and never in the draft. Muse Spark Open on Baseten reports no thinking-token count on the Messages path, so its cost per appeal omits reasoning tokens. Fifty cases is a small set: one case is two percentage points.
Pure layout over the stored result files of the three kept live-CHART runs (README Table 1, last three rows): the entries each model opened, in order; its sentences with the gate's stamps; its verdict; the call's cost and latency. Nothing is re-run here. Pick another case with the case picker above, or press C to return to the replay.
The figures of the paper are frames of this replay: each link lands on its frame, and the arrow keys walk the trace from there. The tables are in the README on GitHub; Table 1 is also the Scoreboard tab. Read from README.md when the page is built, so this list is the paper's.