A decision-grade record freezes the brief, the answer key, the budgets, and the graders before anyone sees output. The unit is the business case: six generations of one brief are one case, not six. Artifact properties — how many citations, how firm the figures, how consistent a report is with itself — are diagnostics. They are not pass rates, and a report can be perfectly self-consistent and consistently wrong.

This report reads a June 2026 third-party assessment as dated third-party evidence. Its results are external measurements of artifact properties on one satellite-sector brief, six reports per system, pooled. Where this report derives a quantity from those results — a checkable-share ceiling, a contradiction count per report — the derivation is labeled modeled, the assumptions are stated, and the underlying figures remain the evaluator's. Nothing here converts that table into our pass rate.

The bake-off ran four systems on one commercial-diligence brief: Binocs, Claude Opus 4.8 and GPT 5.5 single-shot, and GPT Deep Research. Six dimensions were scored: citation infrastructure, claim commitment, analytical coverage, analytical character, internal consistency, and unit specificity. On the published cells Binocs leads breadth outright — roughly ten times the citation volume of any comparator at low source concentration — and ties the twin on self-consistency, 66.7% against 66.0%.

Every dimension is a property of the artifact. The design contains no answer key, no correctness estimand, and, by its own qualifications, no verification of underlying source claims. Firmness is scored as strength: 82% point estimates reads as strong. In a diligence record the direction reverses for the unsourced share: a firm figure without a source is the most dangerous object in a memo, because it reads as checked and cannot be checked. Hedging an unsourced estimate is calibration, not weakness.

Coverage rewards volume. The 1,062 entities and 282 unit types are pooled across six regenerations of one brief, so per-report breadth is unreported and regeneration variance counts as coverage. The topic table shows 23% of Binocs's quantified values are its own scoring and rating outputs — self-produced numbers, not observed market facts — against 3%, 1%, and 10% for the comparators. Market sizing, the load-bearing quantity in commercial diligence, is 9% of Binocs values and 33% of Claude's.

A diligence reader can check a figure only if it carries a citation and the citation resolves. Multiplying the bake-off's own cells (modeled): Binocs 0.56 × 0.84 ≈ 47% of figures are even checkable, against 59% for GPT Deep Research, 40% for GPT 5.5, and 34% for Claude Opus 4.8. The next link — does the source support the figure — was measured for no system. The chain fails closed: the end-to-end verified share is unknown for all four, bounded above by those ceilings. On the one chain a buyer walks, the twin leads and nobody finishes.

Firm-but-uncheckable exposure, treating commitment and citation as independent (modeled): about 36% of Binocs figures present as firm point estimates with no source attached, against about 18% for the twin. Contradiction burden per report, from the same consistency table: Binocs about 47 contradictions per report (284 across six), Claude about 10, GPT 5.5 9, the twin about 6. The 66.7% rate is the like-for-like read and the count is the reviewer's workload; both are true, and a memo that repeats more figures hands the reviewer more adjudications.

Evidence staleness sits in the bake-off's appendix uninterpreted: median source age 579 days. For a fast-moving sector brief, the median Binocs citation is about nineteen months old, against 176 days for Claude. Forward-looking content is 8% of values against 28% for the twin. The published profile is a wide, firm, historical base — the shape of a data appendix, not of the projection a committee acts on.

LiveDRBench scores deep research as claim precision, recall, and citation validity against a frozen key: OpenAI Deep Research reaches 0.629 precision, 0.519 recall, and 0.550 F1 weighted, with category spread from 0.314 to 0.721 F1. BrowseComp puts persistent-web retrieval at 51.5% for Deep Research with a 0.91 calibration error — browsing systems are overconfident about answers they cannot find. Those are bounds on the twin's architecture and on the workflow category. They are not Binocs results, and no equivalent correctness measurement of Binocs exists anywhere.

The failure mode that matters for Binocs is reading artifact breadth — citation count, firmness, entity volume — as factual correctness. The bake-off did not carry a key.

Mapped onto this report's comparison set, the bake-off ran a raw-model pair (Claude Opus 4.8, GPT 5.5 single-shot), a declared agentic configuration (GPT Deep Research), and the vendor surface (Binocs) — three of five required comparisons, with no current-process cell and no isolated comparisons, scored on artifact properties without a key. That is a useful harness contrast, and it cannot attribute anything: Binocs changes model, retrieval, orchestration, and output format at once relative to every comparator. The missing pair that would price the product is Binocs versus the twin at a frozen retrieval budget on a sealed brief with an adjudicated key.

Vendor materials describe a report-count fair-use limit, per-report overages, and an optional expert-human layer. Base subscription terms, service boundaries, and staff hours are not observed. Subscription price therefore cannot be read as software-only or full cost.

Full cost per accepted memo is the estimand: base fees, overages, all attempts, retrieval and tool charges, vendor-side human work, and measured buyer-review minutes to a partner-checkable standard. No complete accounting or correctness gate exists in any record.

The bake-off's most decision-relevant pattern is one it names in passing: the two step-by-step systems behave alike on trust dimensions, and the two single-shot models fall behind. Read structurally, trust behavior tracks architecture, not vendor. GPT Deep Research is a first-party product of the platform that supplies frontier models. It matches Binocs on consistency, leads on citation coverage and forward content, and trails on breadth and unit vocabulary. That is a twin one release from parity on the remaining columns, with distribution Binocs does not have.

There is no exclusive corpus: the input is the open web and the outputs belong to clients. Orchestration, retrieval policy, and the slide renderer are reconstructable in kind. Switching cost is workflow habit. A deal team that already holds a frontier subscription can insource the twin for marginal dollars, so customer insourcing is a live exposure, not a hypothetical. What would be defensible — a verified correctness lead on sealed briefs at a fixed budget, or scarce-source access rights — is exactly what no record measures.

The terms and site footer name different entities, and an undated privacy page appears to retain language from the prior product. Without a current DPA, this remains an open diligence request, not a finding. Target material should not be uploaded until the corporate map, retention schedule, and subprocessors are verified.