Khnum Forge
Khnum Forge
Named for the Norse god of justice and lawful order. Tyr is the digital team built for a Social Security / disability practice — where the case is the medical record, and every fact has to be traceable back to the page it came from.
Most "legal AI" reads documents and writes summaries. A disability practice needs more than that: it needs the case worked the way an experienced advocate works it — the forms understood, the medical evidence lined up against the actual legal standard, the vocational argument built, and every assertion tied back to its source. That domain knowledge is built into Tyr, not something your staff has to prompt out of a blank model.
Tyr is a vertical — a purpose-built practice team, distinct from the general-purpose personas. What follows describes the Social Security / disability build specifically.
Under the hood, Tyr is grounded in the real machinery of a disability determination — the same sources of truth an advocate reasons from:
The actual SSA / disability forms, mapped field by field — so a scanned or completed form is read into structured data, not guessed at.
The Blue Book Listings, the five-step sequential evaluation, residual functional capacity (RFC), and the Medical-Vocational Guidelines — the "grids" — encoded as the framework the case is measured against.
Diagnoses connected to the Listings they bear on, so medical evidence is read for what it means to the claim — not just transcribed.
The step-five vocational argument, built on public-domain U.S. Department of Labor data (occupational codes and real job-number sources) — how a residual capacity erodes the occupational base.
Document AI optical character recognition with a dedicated lane for handwriting and low-confidence regions — the reality of a paper medical file, not a tidy PDF.
Evidence indexed and retrieved as structured facts with their sources attached, so the case is assembled from the record rather than paraphrased from memory.
The reason to trust the output isn't a promise that "the AI is careful." It's the architecture. In Tyr's records pipeline, the deterministic rules run before any model does, and the parts that must not hallucinate are handled by code, not by a language model at all.
Every extracted fact runs through deterministic rules and lookups — dates, codes, form fields, framework checks — before a model is ever asked to reason about it. The model works on validated inputs, not raw guesses.
Source evidence is carried verbatim, with page identifiers preserved end to end. A model is never allowed to stand between you and what a record says — the words that matter stay the record's own.
What is allowed to leave the pipeline is decided by deterministic code — fail-closed and allowlist-only. It doesn't ask a model for permission; nothing ships unless the gate's explicit rules pass.
A date read as "sometime in March 2019" stays that — it isn't rounded to a false, specific day. Confidence travels with each fact; the work never manufactures certainty it doesn't have.
Where a person confirms a reading, resolves an ambiguity, or overrides a finding, that's a recorded, structured choice — part of the case's provenance, not an off-the-record edit.
Every assertion in the work product resolves back to the exact page — and region — of the original document it came from. You don't take a finding on faith; you click it and land on the source.
Accountable decisions stay human — and Tyr makes that concrete. Sign-off is per-artifact and content-hash-bound: the approval is tied to a cryptographic hash of the exact bytes that were approved. Change a single character afterward and the sign-off no longer matches what's in front of you — so an approval can't silently drift onto altered work. And it's toggleable per matter, so a firm sets the gate where its own practice requires.