The output is prose you have to trust.

Five questions a reader asks. Five dead ends.

All five stop at the same wall: the run was not kept, so there is no record to answer from — not the searches, not the pages, not the browser steps, not the model calls.web-research/DEEP_RESEARCH.md · docs/RESEARCH-POC.md §12

The five, as failure modes

• not recorded“Which page said this, in those words?” Prose and a list of URLs survive; the page actually read, and the bytes the sentence rests on, go with the run.
• a different run every time“Run it again, exactly.” Nothing was recorded to re-execute from, so a second run is a second opinion, not a check.
• half of real“What did it cost?” The number printed is an estimate at batch rates, about half the real spend.
• nowhere to put it“This claim is wrong.” Nowhere in the run can a reader’s disagreement be recorded against the claim it judges.
• answered anyway“Did anything fail?” Failures are answered over, confidently.

Our own earlier tool was no different.

It answered confidently when every browser agent had failed, and printed costs at half the real spend. The tool this project replaces — a browser-use clone driving Gemini through OpenRouter — ran the same eight questions the kernel is now measured on, and its eight runs are the baseline in the benchmark set.

16–122 calls · 157–739 s · $0.11–$0.94 estimated per question · 2 of 8 runs with no report. The dollar column is the tool's own estimate at batch rates — the scout run it priced at $0.107 cost about $0.21. The two error rows ended without a report and are still on the benchmark, so the kernel's Q7 and Q8 are judged against nothing.bench/sets/gold-8/questions.json (baseline calls, seconds, estimated_usd, status) · bench/sets/gold-8/caps.toml (the batch-rate note) · docs/E1-2026-09-16.md Q7 (OpenRouter 402)

next: how it works — the whole system as composed, append-only, content-addressed ledgers.