Glossary.

Terms, in the order a run meets them.

• research-stackA self-hosted deep-research engine. It answers a question from web sources and records every search, fetched page, claim, check and dollar, so the report can be audited claim by claim. shown on the home page
• deep researchAn AI system that plans searches, reads many web pages and writes a cited report on a question, over minutes rather than seconds. shown on the problem page
• auditable deep researchDeep research whose every claim can be traced to the exact words of its source, with the run behind the report kept as a record. shown on auditable deep research
• ledgerAn append-only record of one kind of fact. research-stack keeps thirteen: the question, the run, the pages fetched, the claims, the budget and more, all as rows of one database table. Nothing is edited in place. shown on the reference
• eventOne row of a ledger: a fact recorded at one moment, identified by a hash of its own content. shown on how it works
• runOne attempt to answer one question, from the first search to a report or to a stated gap. shown on the worked example
• claimA statement the run is prepared to rest the report on, stored with the exact quote from a fetched page that supports it. shown on the worked example
• anchor · quote checkAn anchor ties a claim’s quote to the exact bytes of the stored page it came from. The quote check, rule A1, refuses any claim whose quote is not found in those bytes. shown on the worked example
• admission ruleA check every new row must pass before it is written. There are fifty-three. A row that fails is refused, and the refusal is itself recorded. shown on the reference
• refusalA recorded no: what was asked for, which rule refused it, and the evidence beside it. shown on the reference
• gapHow a run ends when it could not establish an answer: it reports what it could not find instead of writing around it. shown on the worked example
• A7 · the answer checkA figure or name in the final answer must appear in a quote of a claim the answer cites. An answer that breaks it is refused. shown on the MVP build log
• charterThe run’s locked list of lines of inquiry and the cells each must fill. Once locked it is never edited; an addendum adds to it. shown on how it works
• cover loopThe step that finds the charter’s cells no claim answers yet and sends the run back to dig for them, up to three extra rounds. shown on how it works
• unitOne of the model-backed workers of a run, each with one job and its own spending cap. shown on the reference
• gateTwo uses on this site. A milestone’s gate is its acceptance run: the POC gate was eight checks, the MVP gate 53 questions. The kernel’s gate is the single entry check every write passes. shown on the MVP build log
• POC · MVPProof of concept and minimum viable product: the two build stages logged on this site. shown on the roadmap
• packageA unit of planned work with a written ‘done when’ clause, merged as one pull request, such as EXTRACT-1. shown on the roadmap
• replayRe-running a finished run from its record with no network and no model calls; the same report bytes come out. shown on how it works
• DeepResearch Bench IIA public benchmark of deep-research tasks whose rubrics were written by experts; DRB-II for short. shown on the MVP build log
• rubric item · recallOne checkable point an expert wrote for a benchmark task. An information-recall item checks that a specific fact appears in the report. shown on the MVP build log
• the clean tenThe ten DRB-II tasks research-stack is scored on, with no run allowed to read the source its own rubric was built from. shown on the MVP build log
• JevTypeSafe’s typed-decision model. It cannot write text, only return a probability, a choice or a score, and research-stack uses it as a cheap measuring instrument. shown on the Jev page

the whole idea in one page: auditable deep research