Six stages, each able to fail loudly rather than quietly. Research resolves 136 fields with evidence labels; entity-scale resolution fixes what kind of business this is; an agent simulation produces the behavioural evidence; composition assembles the document; a red team attacks it; and three validation gates decide whether it ships at all.
Six stages run in sequence, and each one can fail loudly rather than quietly. Research resolves 136 fields with evidence labels; entity-scale resolution fixes what kind of business this is; an agent simulation produces the behavioural evidence; composition assembles the document; a red team attacks it; and three validation gates decide whether it ships.
Eleven agents resolve 136 named fields from SEC EDGAR XBRL filings, company filings, a self-hosted SearXNG metasearch cluster and five independent search backends. Each field is stamped VERIFIED, CREDIBLE_ESTIMATE, DOCUMENTED_GAP or NOT_APPLICABLE_BY_BUSINESS_MODEL, with the retrieved URL attached.
Before any advice is written, the pipeline resolves what kind of business this is — solo operator, local SMB, small-cap public, global enterprise. This is what stops a $3.9B contractor being told to improve its Google rating, and it is graded as its own rubric category.
Personas are constructed from the research, then run through a discrete-choice model across designed scenarios. The largest run to date produced 2,046,200 individual choices from 40,200 personas. Choices are computed by multinomial logit, not one model call per agent — which is why populations this size are affordable.
A structured deliverable is built from the resolved profile and the simulation output: sections, claims, citations, charts, recommendations. Nothing reaches the page that is not traceable to a resolved field or a simulation result.
An adversarial pass extracts every claim from the composed deliverable and tests it against the evidence ledger. A red team that returns zero claims is treated as a failure of the red team, not a pass for the report.
Three gates run before delivery: self-contradiction detection, a mechanical checker with zero-tolerance rules, and the 9-category rubric. A designed page only ships if it grades higher than the deterministic template it would replace.
A deep run holds an 8-hour wall-clock budget across all six stages. Everything the run did — every request, every model reply, every decision — is written to a JSON trace that ships with the report.
Every resolved field carries one of 4 evidence states, and one of them means we looked and could not find it. That state is printed in the report body, beside the claim it would have supported. A missing number is never defaulted, rounded, or quietly dropped — the gap is the output.
VERIFIEDCREDIBLE_ESTIMATEDOCUMENTED_GAPNOT_APPLICABLE_BY_BUSINESS_MODELNine categories, scored on the rendered HTML. The scoring is split into two bands, and the split is the whole design: 584 floor points are awarded for having no detectable defect, and 416 earned points start at zero and are awarded only for analytical properties the document actually demonstrates. A flawless report that proves nothing scores 584, not 1000. A perfect 1000 is deliberately unreachable.
| Category | What it scores | Points |
|---|---|---|
A | Zero fabrication | 191 |
B | Entity scale | 95 |
C | Financial rigour | 143 |
D | Simulation insight | 143 |
E | Recommendation substance | 143 |
F | Research completeness | 95 |
G | Narrative coherence | 76 |
H | Honesty | 38 |
I | Visual and technical | 76 |
| Total | 1000 |
Zero fabrication carries the most weight of any category, at 191 points, because it is the failure that makes every other score meaningless. Six zero-tolerance rules can veto a report outright regardless of its total.