1 ·REEVE
OSHA compliance · Live · production
Use casePhoto-to-hazard analysis (construction-site photos → hazard categorization with verbatim 29 CFR 1926 citations) and OSHA-rule retrieval grounding.
Underlying modelAnthropic Claude Sonnet (4.x family). Vendor-managed inference.
Fine-tuningNone. Off-the-shelf model with retrieval-grounded prompting.
Training opt-inWe do not opt into Anthropic's training-data program for this workload. Customer submissions to Reeve are never sent to any training pipeline (ours or vendor).
Future training corpusLimited to de-identified public regulatory data (OSHA citations, 29 CFR text, court filings). Because those records are public, we cannot guarantee a specific customer's prior incidents are absent from the public source data.
What it seesConstruction-site photos submitted by the customer plus the relevant 29 CFR 1926 paragraphs surfaced by the citation index. Nothing else.
What it does NOT seeCross-tenant data. Photos from other GCs. Worker PII outside the submitted photo. Anything from the audit chain unrelated to the active photo.
Output styleVerbatim citations only — never paraphrased. Every hazard claim is anchored to a paragraph and line in 29 CFR 1926.
Human-in-the-loopEvery penalty-bearing output is reviewed by the customer (the GC) before submission to OSHA. The system drafts; the GC signs. The system does not auto-file on the customer's behalf.
EvaluationMethodology under documentation. Contact security@vendriclabs.net for current status.
VersioningModel upgrades to a new Sonnet version are reviewed before customer-visible rollout. Pinning + regression testing posture under documentation.
2 ·VENDRIC
Substrate layer · Live · provenance under Reeve in production
Use caseSubstrate layer providing memory, cognition, prediction, and provenance primitives that downstream surfaces (Compli, Reeve) inherit.
Underlying modelNot a customer-facing inference surface. Substrate-layer building blocks only — no user-visible model runs at this layer.
What it seesInternal data structures originated by Compli and Reeve operations. No direct user input.
What it does NOT seeEnd-user prompts or conversations. Vendric does not have a user surface.
Training opt-inNot applicable — no inference endpoint.
Persistence formatEncrypted persistence format · spec frozen. Tamper-evident audit chain primitives live in production via Compli.
Human-in-the-loopNot applicable at this layer. Human-review checkpoints live in the downstream product surfaces (Reeve today; future Compli verticals tomorrow).
StatusLive · substrate. Provenance layer beneath Reeve in production.
3 ·METHODOLOGY
Every Model Card entry above is intended to be falsifiable. If a model changes, the entry changes. If a behavior shipped on Tuesday and was removed on Friday, the card reflects Friday. The card is part of the product — not a marketing surface decoupled from the engineering reality.
3.1 — What we will not claim
Per FTC guidance and the AI Trust + Security brief informing this rebuild, every AI claim on a Vendric Labs surface is backed by a specific output, customer type, or measurable behavior. We do not use the phrase “AI-powered.” We do not claim “revolutionary” capabilities. We do not benchmark on internal data and report the result as third-party validation.
3.2 — Human review checkpoints
Reeve: every form submission, every penalty estimate, every photo-to-hazard categorization is shown to the GC before any external action. The system does not auto-file with OSHA.
Vendric: substrate layer. No direct user actions originate here. All user-facing decisions happen at the downstream product surface (Reeve today; future Compli verticals tomorrow).
3.3 — Reporting concerns
If an output appears wrong, biased, or unsafe, email security@vendriclabs.net. We treat model behavior issues with the same urgency as code bugs.