Auditable deep research.

Why most deep-research reports cannot be audited.

A deep-research tool searches the web, reads pages and writes a report with citations. What reaches you is the prose and a list of links. The run that produced it is not kept: which searches were made, which pages were actually read, which passage a sentence rests on, and what the tool looked for and did not find. A reader who wants to check one sentence has to redo the research.

When independent researchers did redo it, the gap was large. DeepTRACE, an audit of deep-research systems, reports citation accuracy ranging from 40 to 80 percent across the systems it tested. Cited but Not Verified found that the strongest models’ cited reports reach only 39 to 77 percent factual accuracy once each citation is checked against its source.arXiv:2509.04499 · arXiv:2605.06635

the five questions a reader asks and where each one dead-ends: the problem

What an audit trail has to hold.

• the questionAs it was asked, and any narrower question the run decided to answer instead, stated before the digging starts.
• every search and pageEach query and each fetched page, stored as bytes, so a quote can be found again later and nothing depends on the live web staying the same.
• every claim, with its quoteThe statement the report rests on and the exact words of the stored page behind it. A quote that is not found in those bytes is refused.
• every check and refusalWhat was asked for, which rule said no, and the evidence, so a reader sees what was thrown away as well as what was kept.
• what was not foundThe gaps, stated in the report in the run’s own words instead of written around.
• what it costEvery paid model call, reserved against a cap before it is made and settled after, so the price of each answer is a sum of rows.

research-stack keeps all six as rows of thirteen append-only ledgers in one database. Nothing is edited in place, so the record of a run cannot drift from what the run did.docs/SPEC-kernel.md §1–§3

the same record drawn as a machine: how it works · one sentence of a real report traced to its bytes: the worked example

How auditability can be measured.

A February 2026 paper, From Fluent to Verifiable, proposes four measures of claim-level auditability for deep-research agents. In our words:

• provenance coverageHow many of a report’s claims point to a source passage at all.
• provenance soundnessWhether the passage a claim points to actually supports it.
• contradiction transparencyWhether evidence that disagrees is shown rather than dropped.
• audit effortHow much work a reader needs to check one claim.

The paper defines the measures; it does not publish results for any system.arXiv:2602.13855

Where research-stack stands.

Every claim in a research-stack report carries a quote that must be found, byte for byte, in the page the run stored. At the proof-of-concept gate every resolved quote was re-read from the store: 490 of 490, with 0 mismatches.docs/GATE-POC.md — check 4

The minimum viable product ran 53 questions over 20 campaigns for $8.27, and every report was read.docs/GATE-MVP.md

Auditable is not the same as right. On ten DeepResearch Bench II tasks it scores 16.42 over 672 expert-written rubric items, well behind the commercial systems.docs/GOLD-EXT-CLEAN.md

The record also shows its own mistakes. In four of our passes, 12 to 33 percent of the claims that state a figure stated one that none of their own quotes contained. The answer check refuses such a figure before it reaches a report, which is why one run ended with no answer rather than a wrong one.issue #1419

how to check any report’s citations · the glossary · what is broken