How Outsample works, and where it fails.
This page explains what runs when you ask a question, using one real run and its measured numbers. The limits we know about are at the end.
Every question runs three stages.
| stage | what happens | real run |
|---|---|---|
| Retrieve | Hybrid search over the corpus: dense embeddings plus keyword recall. This stage casts wide. | 295 papers |
| Gate | A relevance pass keeps what bears on the question. If too little survives, the answer is flagged as thin evidence. | 50 kept |
| Synthesize | The model writes the answer from the gated set and cites the papers it used. | 2 cited |
The real-run column is a disagreement query on cross-sectional momentum, 2026-06-11. 23.98 s, $0.0407. The full run is shown further down this page.
An answer reasons over the gated set, typically tens of papers. No retrieval system reads fourteen thousand papers per answer, and a product that implies it does is lying to you.
Answers are constrained to what retrieval returned.
Synthesis writes from the gated papers. It does not assert past what it found: thin retrieval is flagged as thin on the answer, and a question with no support in the corpus ends in a bail. Citations name the papers an answer used, and each links to its source.
Every answer carries its counts and its cost. From the run above:
retrieved 295. used 50. cited 2. $0.0407.
When the evidence is not there, it says so.
Asked where the literature genuinely disagrees on cross-sectional momentum, retrieval returned 295 candidates and the gate kept 50. Two papers engaged each other directly. The answer mapped their dispute, then stopped:
“No other paper pairs in the corpus engage directly enough with cross-sectional momentum to constitute a genuine contradiction.”
It did not pad the result with pairs that do not exist. A question the corpus cannot support ends in a statement of the limit and a list of what was actually found.
Two kinds of paper, marked in the citations.
Most of the corpus is full text. A smaller set, mostly classics that sit behind journal paywalls, enters as reference entries.
| Full text | Parsed body in the corpus. Synthesis reads the paper itself. | 13,454 (89%) |
|---|---|---|
| Reference entry | Metadata, abstract, and a link out to the publisher. | 1,664 (11%) |
Citations mark which kind they cite. A reference entry can ground an answer only as far as its abstract goes; the link out covers the rest.
Measured 2026-07-31 against the shared library. Full composition, sources, and field coverage: /corpus.
Current limits, as of July 2026.
Retrieval is top-k.
An answer reads the gated set. A relevant paper can be missed at retrieval, and the counts on each answer show what this one read.
Output tracks corpus depth.
Ask about a niche the library barely covers and you get thin-evidence flags or a bail. Coverage by field is on /corpus.
It runs on a frontier model.
A misread of a paper's claim is possible. Citations are checkable, which bounds the damage without removing it.
Search coverage is partial.
Search runs on summaries and embeddings: 13,935 of 15,118 papers today. The rest are reachable by browse and metadata filters.
Heavy tools are slow.
A strategy run with its adversarial review measures 3 to 4 minutes; a corpus question about 30 seconds. The times published here are measured medians.
It does not backtest.
Outsample tells you what to test and what is wrong with the idea on paper. Whether it survives data is your job, on your infrastructure.
The Free tier includes three full synthesis runs. Run one on a topic you know cold and read the citations it returns. That test is worth more than this page.