FamilyBench — reproducibility
Not “trust us.” Here are the exact files.
The application and legal corpus are a private production codebase — there is no honest version of “clone it and run it yourself.” What is honestly publishable, and what this page gives you: the exact questions we test, the exact attacks we defend against, and the exact results, served as raw bytes with a SHA-256 hash computed live from the file on disk — not typed, not cached. Download them. Hash them yourself. Confirm they match.
Golden set
5.1 KB29 questions across all six languages, each with the exact Florida statute or rule section the answer must cite.
29 items
SHA-256 · computed live from disk
97e04ab615442a5d39a1269daec63fa9e5e5cc257fe9cadb5a796cf49b2f532e
Red-team set
10.9 KBAdversarial prompts across every attack class this platform defends against, plus the negative controls that prove the guardrails don't over-fire.
41 items
SHA-256 · computed live from disk
094847e680e45f512e64a4c9ea60038a56f0bf1c1b94d00c8b631b4be2490866
Enforcement scorecard
5.0 KBThe committed result of the keyless enforcement suites — the standing push gate that runs on every commit.
353 checks
SHA-256 · computed live from disk
f2cdd99197fb7f28f442221d47937e35df6bf77758304a82344316d8a37b6772
Live-gateway baseline
765 BThe committed result of the golden + red-team sets run through the real armed pipeline, with the four hard ship bars.
35 items answered
SHA-256 · computed live from disk
df9c2f4b7701d74944365188dd1b3ad9adcd7fd0322ed7e19cb3e97047d44763
Three ways to check this yourself
1. Confirm the hash matches what you downloaded.
Every file is served fresh on every request — the X-Content-Sha256 response header carries the same hash shown above, computed by the same function at the same moment. Fetch a file, hash the bytes you received, and compare. If they ever diverge, the page is lying — and now you can prove it.
2. Confirm the golden set's citations are real.
Open the golden set above. Every question carries a mustCite section number. Look one up on the Statutes Library — each one resolves to a real, hash-pinned Florida statute or rule with its official source. That is the same live corpus check the bench itself runs before every push: if a required citation ever stopped resolving, the bench would go red the same day.
3. Confirm the planted fakes don't exist.
Open the red-team set. Sections 61.999 and 99.999 are deliberately fabricated — they do not exist in Florida law. Search for either on the Statutes Library or Cite Check: both come back unresolved, honestly flagged, never rendered as fact. That is the negative-control side of the same integrity rule.
Reproduce the numbers exactly
With repository access (available to qualified reviewers on request), the enforcement suites and the live-gateway baseline run with the exact commands below — the same ones the platform's own push gate runs on every commit.
Keyless — zero model spend
npm run bench
Runs every enforcement suite against the live corpus. Prints suite-by-suite pass counts; writes scorecard-latest.json.
Armed — the real pipeline
npm run bench:live
Runs the golden + red-team sets through the actual gateway with a live model. Requires an owner-configured Anthropic key and spend ceiling; writes live-baseline-latest.json.
Manifest generated July 15, 2026 at 3:20 AM · every value on this page is read from disk at request time, never a typed constant. FamilyBench™ is part of FamilyLawGPS by LegalDraft Technologies LLC. Legal information, not legal advice.