Skip to content

Latest commit

 

History

History
35 lines (26 loc) · 1.34 KB

File metadata and controls

35 lines (26 loc) · 1.34 KB

ADR 0001: Require a human-approved manifest before evaluation

Status: Accepted Date: 2026-07-22 Decider: Chelsea Kelly-Reif

Context

The product eventually might help map prose requirements to evidence. Allowing runtime extraction or model interpretation into the acceptance path would make the most consequential and least measurable step implicit.

Decision

M0 accepts only a strict, source-bound manifest authored or approved by an accountable human. The evaluator cannot read a PDF, infer clause meaning, draft criteria, or activate proposed mappings.

Options considered

Option Complexity Main risk
Human-approved manifest Low Manual setup and possible human error
Rules-based clause extraction Medium Brittle false certainty
LLM extraction and direct evaluation High Unreviewed interpretation controls acceptance
LLM draft with human activation Medium Appropriate future research seam

Consequences

  • The core evaluation can be deterministic and independently tested.
  • “Unverifiable” remains visible instead of being optimized away.
  • M0 does not demonstrate document ingestion or mapping productivity.
  • User research must determine whether manual mapping is tolerable.
  • Any future model-assisted mapper is a drafting tool with its own eval and no activation authority.