Citations and verification

The central claim of this product is that its citations are checkable. This page describes exactly what is checked, by what, and what happens when a check fails.

What a citation is here

A citation is not a link. It is a chain of data, and each link in it is stored:

claim sentence  →  [n]  →  sources[n]  →  { url, title, snippets[] }
                                              ↑
                    the verbatim text the executor extracted from that page

The executor never returns a bare fact. Every piece of evidence carries the source URL, the source title, the claim it supports, and a verbatim snippet — the actual text from the page, capped at 500 characters. That snippet is the provenance unit; everything downstream is a view of it.

A single page routinely supports several distinct facts, so a source carries every snippet extracted from it, not one. Keeping only the first meant a chip could show a quote that had nothing to do with the sentence it was attached to.

Citations of your own files

A citation of a file you uploaded points at a character range in the text the app extracted from it: corpus://<document>#chars=<start>-<end>. Opening it shows that text with the cited passage highlighted, and names where in the file it is — the page of a PDF, the sheet of a spreadsheet, the slide of a presentation, the cell of a notebook — from the positions recorded at ingest. A PDF can also be opened at the cited page. The passage is shown as text, never rendered, so nothing in a file can run in the app.

Some formats are read row by row or cell by cell rather than as running text (spreadsheets, notebooks, presentations). Their quotes are checked against that linearised text, and the passage says so beside the quote.

A run limited to selected files can only cite those files, at the versions frozen when it started; a later upload or delete does not change what it cited.

Reading the report

Every factual claim carries [n] markers. Hover one and you get the source title, the domain, and the verbatim snippet supporting that claim. The sources panel lists the same table in full.

Markers that cannot be resolved render a visible ⚠ unverified chip. This is the design, not a bug report: a citation apparatus that renders broken markers as clean text is worse than no apparatus, because it looks like evidence.

Rendering is deliberately conservative. Reports are model-generated Markdown rendered in your browser, so raw HTML is never passed through — no rehype-raw, no dangerouslySetInnerHTML, enforced by a check in CI. The citation chips are produced by a dependency-free Markdown plugin rather than by injecting HTML. Remote images in report Markdown are not rendered either; they become links, so injected content cannot exfiltrate a reader's IP through a tracking pixel.

Today only single [n] markers become interactive chips. Grouped markers like [3, 11, 18] still resolve in the sources table but are not hoverable.

Citation resolution rate

Every approved report records what fraction of its in-text markers resolve to a real source. It is shown on the run and on every row of history, and history can filter on it.

Not measured is a distinct state from 0%. A report that made no citable claims, and a report where every marker is broken, are opposite findings; storing them as one number would make an unmeasured value indistinguishable from a total failure. The column is nullable and nothing renders null as zero.

The same function computes the number for the product and for the evaluation harness, so a published figure and a displayed figure cannot disagree.

What is checked, and where

Two of those are worth spelling out.

The snippet must be text that was really fetched. The executor's output is checked against what the tools actually returned during the run. A model that writes a plausible quote from memory does not get to attach it to a real URL — that is a fabricated citation, the precise failure this product exists to prevent.

Contradictions are surfaced, not resolved. When two sources cannot both be true, both sides are quoted from snippets the detector was shown, and any pair whose source URL was not in the evidence is dropped. The pairs appear in the report and the count appears at the review gate. Nothing picks a winner.

How often a citation actually supports its claim

Resolution says a marker points at real evidence. Whether that evidence supports the sentence is a separate question, and the product does not answer it at runtime (below). The evaluation harness answers it offline, with an independent model as judge — one that is none of the models under test, named in the result along with the model that actually answered each ruling.

The 3.0.0 release run measured 96.4% citation support over ten fixed questions, with Gemini 2.5 Flash running every agent and Claude Sonnet 4.5 judging, against a 95% threshold. That is one run on one model routing: quality tracks the models you route to, and support means "the snippet backs the sentence", not "the sentence is true". The method, the other runs and every caveat are in what has actually been measured.

What is not checked

Stated plainly, because a verification story with a hidden gap is worse than none:

• The critic grades per-task evidence, not the finished report. Nothing at runtime re-checks that the synthesizer's use of [n] is faithful to the snippet it points at. That is measured offline by the evaluation harness and by the citation-fidelity benchmark, not by a runtime gate.
• Resolution is not truth. A marker resolving means the source and its supporting snippet exist and match. It does not mean the source is correct, or that the claim is true in the world.
• Evidence de-duplication is URL-exact. The same article at two URLs becomes two sources.

Offline verification

The .bundle.json export is the strongest form of this. It carries the report, every claim with its cited indices, every evidence snippet with a SHA-256 content hash, the sources, the contradictions, the model routing, the costs, and the full approval chain — and a bundle_hash covering all of it.

python -m research_engine.verify_bundle path/to/research.bundle.json

Seven checks, no AI, no network, no database. Exit code 0 or 1. Full specification: Research bundle format.

Chat citations

Follow-up chat over a finished report answers from that report and its sources, and says so. The scope selector states what each option will and will not read, and a scoped answer reports which grounding produced it.

Project chat answers from approved reports in that project and cites them as [R1], [R2], resolving to the report title, date, and link — and through the report to its original sources. A claim with no resolvable citation gets the same ⚠ chip as anywhere else. Citations are computed from what was actually retrieved and then narrowed to the markers the answer used, because listing an unused excerpt as a citation would be exactly the sources theatre the ⚠ chip exists to prevent.