Twenty-eight packages in five days, measured on 16 September: a bench line of 47 gated rows with no miss, a gate of eight checks with one miss drawn, and eight questions judged against the old engine — five of eight, one short of the bar. The miss became the first fix; the judge's finding became the MVP.
The eight questions were first answered by the old browser-based engine (web-research, 21 August and 9–10 September). The kernel was designed on 11 September as composed, append-only ledgers, built as 28 packages, and measured on the 16th against the same eight questions.docs/DESIGN-kernel.md · docs/PACKAGES.json stage poc · git history 2026-09-11 → 2026-09-16 · web-research/runs/<id>
Every claim in a report resolves to a byte range of a page the kernel fetched and stored. Kill the process at random and it resumes without repeating a call. Replay a run from the ledger at $0 and the report is identical. Refuse a unit that would cross the budget before it is invoked. Seven of the eight gate checks passed on the first measurement.
Check eight failed: one live run at concurrency 6 cited a page a sibling task had fetched, and the same output was rejected on replay at concurrency 1. FIX-590 made a claim resolve against its own task's retrievals, and merged the same day (#673). Checks 1, 4, 5 and 8 have run in CI on every merge since.docs/GATE-POC.md — the checks · docs/PACKAGES.json FIX-590 · PR #673
Every value is a bench.result event; the verdict is recomputed from value and target when the report renders, never read from a stored flag. Measured 2026-09-16 on a MacBook Pro (macOS aarch64, 10 cpus, 32 GB), release profile, commit 79c3370.docs/BENCH-POC.md · docs/GATE-POC.md · docs/E1-2026-09-16.md
The old engine's report beside the kernel's for each of the eight questions, judged with the method written down first. Not-worse on five, worse on three. The kernel is far cheaper — it drives no browser — and that is where it lost: its share of primary sources roughly halved wherever the old engine had found them, because the interim fetcher was locked out of the sites practitioners write on.
Of 179 fetch rows in the gate ledger, stackoverflow.com failed 7 of 7 (403 challenge) and reddit.com 3 of 3 (robots). The pricing page the Q1 answer needed was fetched fine — the miss was extraction, not fetching. Ten defects were filed; the judging method became the first evaluation doctrine.docs/E1-2026-09-16.md — method, verdicts, the finding, ten defects · doctrine/eval/v1
The eight verdicts are recorded as Decision.feedback events in the gate ledger; the document is the reasoning they cite.docs/E1-2026-09-16.md — verdicts, tally · Gold-8 Verdicts page
next: the MVP — 42 of 43 packages merged, 53 questions run for $8.27, and what the reading found.