From a messy deal room
to a working deliverable.

Native Lev and three API conditions

Financial accuracy and document accuracy are different tests.

The current common comparison contains 8 case/stage pairs from 4 synthetic deal packets, with one trial per stage. Every chart uses the same completed inputs across Lev and all three API agents. These correlated stages do not establish an overall winner.

How large are the errors?

A strict field score can count one input error several times as it propagates. These are the largest absolute deviations in the same matched cohort, measured against the specified underwriting assumptions. They are separate from a practitioner's materiality judgment.

Scoring v1.1 corrects a revised-occupancy risk key, accepts equivalent citation objects, and excludes three ambiguously named ratio/capacity fields for every provider. Financial denominator: 12 per stage. Dependent output errors can share one incorrect starting input; field counts are not independent incident counts. Key-field mismatches use the declared scoring tolerance, not a practitioner materiality threshold. Original v1 grades remain public. Read the complete correction.

Consumer product comparison

Claude Chat on the same initial deal and revision

This smaller comparison is shown separately so consumer-product coverage does not narrow the main Lev-versus-API cohort.

Coverage is part of the result

36 completed workflow stages. A larger test matrix remains.

The target is 20 packets × 2 stages × 3 trials per condition. The packets come from four shared authoring families; there are zero authentic customer deals in this public release. Missing runs remain missing. Consumer product access and data parity are disclosed separately.

API conditions use the same open file, calculation and artifact harness. Claude Chat uses its native tools and displayed Opus 5 High setting. Lev Agent is a separate native product condition; these runs do not test the dedicated Index or Memo-builder surfaces. ChatGPT agent access remains pending. Provider compute and data access are not proven identical.

Six workstreams

What this expansion tests

Lender evidence probes

Can it distinguish program fit from approval?

Fixed public-source summaries test closed programs, loan-size ranges, occupancy exceptions, leverage and missing DSCR. The automated checks below do not grade every sentence of the explanation. Current category mismatches concern the missing-DSCR case; these answers still identify the missing financial information in their explanations.

Prompts, source links and reference answers · All original lender runs

Actual retrieval

Live research has a separate verification step.

The brief asks each condition for at most eight searches and eight page opens. The API harness enforces these limits; the native product is not mechanically capped by this harness. Native tools and data access differ from the public API search bridge. Completion means an answer was returned; it does not establish that every candidate is correct. Search-service cost is unmeasured. No web-wide recall or lead-conversion claim is made.

Source-audit examples: Lev retained an office loan whose 2027 maturity had already been extended to 2029, despite describing the extension correctly. Other findings include a Lev contact-name/email mismatch, unsupported closing dates in API answers, and a GPT-5 lease comp missing a free-rent concession. These are field-level findings, not complete research scores.

Frozen briefs and published audit records · Source-linked findings

Inspect the evidence

Every recorded workflow stage

What is needed for a qualified release

Independent CRE reviewers must qualify the references and blindly assess professional readiness. Human correction time, acceptance and reviewer agreement are currently unmeasured. Real customer test packets require documented release rights before public distribution. A full cohort, repeat trials, native lender matching and accessible consumer-agent conditions remain outstanding.

Lev owns and funds this project. The prompts, synthetic data, versioned graders, run records and original deliverables are open. Lev's backend remains proprietary. Results that favor another system are published under the same rules.

python3 tools/collect-expansion.py
python3 tools/plot-expansion.py
npm run build
npm run check:site

Updated 2026-09-07T14:56:15.798161+00:00. Regrading published outputs does not require paid API calls. Fresh executions incur inference charges.