We show our work.
Every claim, back to its source.
A sonde is an instrument you send into an environment too large to survey by hand — it goes where you cannot, measures what it finds, and returns with data you can check. Lumisonde is that instrument, pointed at the flood of technical information: deep research, automated end to end. Coordinated AI agents pose their own questions, read widely across the live web, weigh and integrate what they find, and deliver the result as reports — insight included, no human researcher in the loop.
Automation alone earns no trust, so we audit it. Every claim links back to the source it came from; numbers are traced into archived source text before anything ships; and each report publishes the verification checks it passed alongside the ones it didn't. A claim that cannot be traced loses its wording, or doesn't ship at all.
What a receipt looks like ↓What a receipt looks like
Every [N] in a report opens the record behind it: the query that surfaced the source, what we extracted from it, and the section it fed. Here is one, unedited, from Hard, Verifiable Agent Data: Download for Code, Build for the Rest.
In the original SWE-bench, 32.67% of successful patches involved solution leakage (the fix visible in the issue text or comments) and 31.08% of passing patches were suspicious due to weak tests; filtering these problems dropped SWE-Agent+GPT-4's resolution rate from 12.47% to 3.97%[6].
In SWE-bench, 32.67% of successful patches involve 'cheating', meaning the solution is directly provided in the issue report or comments.
In SWE-bench, 31.08% of passing patches are suspicious due to weak test cases; among them, 12.75% of passing instances are incorrect fixes.
How a report is made
Lumisonde is not a single model writing an essay. An orchestrator delegates each phase to a specialist agent — retrieval, outlining, writing, verification, review — and holds the boundary between them. The critical rule is that evidence is gathered and frozen before any writing begins: nothing downstream can cite outside that pool, which is what makes every sentence traceable.
Targeted searches run against each question. Candidate sources are read, screened for relevance and authority, and reduced to the facts, claims, and quotes worth keeping. What survives becomes the evidence pool — frozen, with provenance attached.
The report is framed around decision questions — "when is X worth it?" — rather than a fixed template. Major clusters found in research must be covered, or explicitly scoped out with a reason.
One controlling verdict, argued section by section, citing only the frozen pool. Evidence is graded and the language must match its grade: a controlled comparison may "establish"; a result that doesn't isolate the variable, or a vendor-authored figure, may only "suggest" — and vendor studies are named as vendor studies.
Deterministic checks that can block publication: every [N] resolves to a real source, every cited URL exists in the evidence store, every figure traces to a source extract, named systems appear in the evidence, and quotation marks are reserved for verbatim source language.
An independent pass re-examines the draft and applies minimal, audited edits. It sees the report, not the reasoning that produced it — then the gate runs once more.
The report ships with its receipts: per-claim source statements, the source list with authority labels, the retrieval funnel, and the audit ledger.
What the checks do — and do not — guarantee
Traceability — every figure links to the source it came from
Internal consistency — markers, sources, and figures reconcile
That a third-party figure is correct, or that an experiment reproduces
That a volatile value — a price, a model version, a benchmark — hasn't changed since we retrieved it
A report is anchored in time: it reflects its sources as of its publication date. We say this plainly rather than implying permanence.
Limitations
The writers here are language models, and language models can be wrong in ways the checks don't catch. Verification establishes that every claim traces to its cited source — it cannot establish that the source itself is right, and sound evidence can still be framed into an unsound argument. Lumisonde is in beta: the reports, the checks, and this site are all improving in the open.
If you find something wrong — a number, a citation, a conclusion — we want to know: contact@lumisonde.com.