EVALUATED 27 OCTOBER 2026
Evidence,
not a score.
A copied quotation can be checked. A convincing interpretation still needs your judgment.
What the system checks
The hosted investigator returns source quotations, interpretation and unknowns as structured data. Each quotation must appear in its attributed, selected source after typographic quote and whitespace normalization. Unchecked references or invalid quotations withhold the answer. No retry or hidden fallback invents a replacement.
This does not prove a source is true or a conclusion follows from it. A model can cite the right document and still reason badly. That limit is part of the interface.
Sixteen deterministic tests
Tests cover fabricated and misattributed quotations, excluded sources, spliced passages, blank quotations, unsupported references, uncited interpretations, explicit no-evidence responses, input limits and labeled export. All sixteen passed. These tests validate the guardrail, not model accuracy.
Seven actual model calls
Six initial examples and one follow-up used hosted Gemma 4 through the Worker preview. These are captured responses, not simulated demos or a representative benchmark.
| Scenario | Displayed outcome | Model time |
|---|---|---|
| Clock discrepancy | quotes matched | 3.5s |
| Does the boat prove death? | quotes matched | 4s |
| Birthday not in the record | withheld | 2.7s |
| Unproven murder premise | quotes matched | 3.9s |
| Evidence deliberately excluded | withheld | 2s |
| Explicit ghost fiction | fiction | 3.1s |
| False premise after refinement | withheld | 5.1s |
“Quotes matched” means the quoted words occur in selected sources. “Withheld” means the receiver declined to display an interpretation that did not pass its checks. “Fiction” deliberately permits invention.
What failed, and what changed
The initial birthday and excluded-evidence questions were withheld. That prevented an unchecked answer, but did not give a useful substantive response. The initial murder-premise answer rejected a proven murder, yet its unknowns still speculated about an assailant. The prompt was tightened to avoid carrying a false premise into unknowns. The follow-up was withheld, so no improvement in useful-answer rate is claimed.
A separate guided browser run returned three matched quotations and working source links. Notebook and export flows were checked separately. Timing depends on provider load; these observations are not a speed promise.
A bounded public demo
Generation is limited to 240 admitted attempts per UTC day across this app, plus an approximate six-per-minute limit per IP and Cloudflare location. Failed admitted calls count too. Reading evidence and using the notebook remain available when the AI allowance is exhausted.
The budget database stores a date and count, not conversations. Questions and selected fictional documents are sent to Cloudflare for inference. Notebook notes stay in the browser. Provider infrastructure policies are separate from application storage.