BRIGHT benchmark · reasoning-intensive retrieval
A reasoning-first retrieval system built on patent-pending architecture. Scored on BRIGHT, the hardest public benchmark for retrieval that requires reasoning rather than keyword matching.
Every question in all twelve BRIGHT domains, scored with the benchmark's own measure. For reference, the benchmark paper's keyword-search baseline sits near 14 on the same scale.
| Domain | Score | Average | Questions | 90%+ |
|---|
Question numbers are BRIGHT's own. A perfect score means every relevant document the benchmark lists for that question was returned, in the right order.
Eight stages, each one measured on its own before it was kept. The architecture's own mechanisms are named in the section below; their internals are not published.
The question is broken into labelled facets, the distinct things a reader would need to know. Each facet becomes its own search, alongside the original question.
Every facet is searched two ways at once, by meaning and by wording, and all paths are merged into a single ranked pool. Measured on its own, pooling the facet searches this way raised relevant-document capture from 71% to 100% of a pool of 200.
The pool is narrowed to what is worth reading, balancing relevance against redundancy across the whole set rather than one document at a time.
Selected documents are read in full. Claims are extracted with the quote that supports them; nothing advances on a summary alone.
Each extracted claim is checked back against its own source text. Claims that fail are dropped before they can influence the ranking.
Confirmed terms and entities accumulate across rounds, so each later round searches with everything already established rather than starting over.
The stable search order and the verified-evidence order are merged into one final order.
The switch to the fused answer happens all at once under a governing check, or not at all; any failed check keeps the stable answer. On a 20-question set the governed cutover moved 64.31 → 66.00, two questions better and none worse.
U.S. provisional application filed 3 April 2026: Hybrid Quantum-Classical Inference Architecture with Entropy-Controlled Latent-State Cutover (GACP), GIRC, GTRI-L System. The names below are the architecture's own; the mechanisms behind them are not disclosed here.
Designed as hybrid quantum-classical and domain-independent; it runs on classical hardware today and the BRIGHT results above were produced that way.
The first stage turns one question into a small set of labelled facets. Below are the labels from a facet set for a coding question; the search text each one carries is not published.
# question → facet set (stage 1 output, coding domain) F1 variable declaration F2 iteration over a sequence F3 conditional branching F4 string formatting F5 error handling
Each facet runs as its own search and the pool is the union of all of them plus the original question. The labels are the only part of the first stage shown publicly.
GRAIL is open to investment and is currently in conversations with investors and strategic partners about the next stage.
What the numbers above represent: a patent-pending architecture, benchmarked in public against the hardest reasoning-retrieval standard in the field, and domain-independent by design. The same system runs over legal, biomedical and technical corpora without being rebuilt for each one.
Available on request: the technical dossier, the measured result for every component, the patent filing, and a demonstration run on a corpus of your choosing. Write to:
Full per-question score tables, run records and verification logs are available on request under NDA.