Evidence ledger
14 records. Every published claim cites one by id, and a claim that cites an id which is not in this ledger fails npm run validate-content rather than being published quietly.
A record documents what was observed and what that observation cannot support. An entry existing here does not make the claim above it true; it makes the basis for the claim checkable.
Evidence ledger
Measured5
Produced by running code in this repository and preserved as an artifact.
- EV-001Tagged transcript baseline, held-out split, corpus 2026-09-07.1
Twelve synthetic held-out transcripts, run once with the deterministic Tagged transcript baseline.
- EV-002Tagged transcript baseline, development split, corpus 2026-09-07.1
Twelve synthetic development transcripts, run once with the deterministic Tagged transcript baseline.
- EV-003Corpus composition, version 2026-09-07.1
The full corpus: 24 transcripts of 427 to 484 words, six case families, twelve development and twelve held-out cases, two of each family in each split.
- EV-004Injection and reversal behaviour of the deterministic baseline
The four embedded-instruction cases and the four decision-reversal cases across both splits.
- EV-005Prototype inference cost of the deterministic baseline
One run of twelve held-out cases.
Public source3
A publicly reachable document. Its existence does not make its claims true.
- EV-012How to Meet WCAG 2.2 (Quick Reference)
Accessibility requirements applied to the comparison and article surfaces.
- EV-013The Pudding
Editorial reference for placing interaction at the point a question arises.
- EV-014Next.js documentation
Implementation reference for the application shell.
Interview0
An approved, anonymized note from a conversation with a real person, with an actual sample size.
No records of this type exist. None have been collected.
Assumption3
A stated input chosen by the author. Not measured and not sourced.
- EV-006No model provider is configured in this deployment
The provider seam used by the lab and by run-benchmark.
- EV-007Assumed review rate used for review-inclusive cost
The review-inclusive cost formula.
- EV-008No commercial comparator has been obtained
The comparison surface and the benchmark.
Hypothesis3
A proposition the author believes is worth testing. No evidence is claimed for it yet.
- EV-009Hypothesis: capture reliability, not summarisation, is the hard part
The meeting assistant category.
- EV-010Hypothesis: workflow placement determines whether notes are used
The meeting assistant category.
- EV-011Hypothesis: trust failures are asymmetric
The meeting assistant category.