{
  "schema": "steno-design-research@1",
  "title": "How Steno was researched and built",
  "prepared": "2026-09-20",
  "source_basis": "Private Steno design documentation and recorded engine proofs; summarized here without private source, paths or operational detail. Page claims must stay within this record.",
  "research_method": {
    "literature_staging": "Zotero library holdings",
    "evidence_sweeps": "Structured subagent query sweeps over the staged literature; one recorded sweep ran 46 queries with mandatory counterevidence families",
    "reception_checks": "Scite search_literature DOI calls tallying supporting/contrasting/mentioning citations and editorial notices; consensus-style synthesis across sweeps",
    "honesty_limits": "Citation quota exhaustion mid-matrix was recorded; unchecked rows were stamped reception-unverified rather than assumed"
  },
  "deterministic_core": {
    "design_thesis": "Parse agreements into code-like structure (sections, cross-references, norms) and run deterministic checks against an operator position library, with objective gates before release; not LLM-as-judge.",
    "llm_allowed_roles": ["document type classification", "party/date/governing-law extraction", "section summarization for UI", "clause-category suggestion", "paraphrase-tolerant matching via local embeddings", "plain-English rendering of deterministic findings"],
    "llm_forbidden_roles": ["deciding position application", "selecting or inserting redlines", "choosing document mutation targets", "approving privileged material for write-through", "gating finalization or export"]
  },
  "formal_engines": [
    { "engine": "eFLINT", "role": "per-document deontic evaluation (rights, duties, violations)", "attests": "deontic structure of curated fragments" },
    { "engine": "Catala 1.1.0", "role": "typecheck of curated models", "attests": "structural soundness only; not per-clause computation from arbitrary prose", "correction": "An earlier write-up overstated Catala as verifying the clause; a design review corrected the scope and re-based the cross-check onto the per-document engines." },
    { "engine": "clingo 5.8.0", "role": "answer-set programming defeasibility checks", "attests": "whether exceptions or priorities override an obligation for curated fragments" }
  ],
  "hallmark_wiring": "Engine results, cross-check corroboration between per-document engines, and the exception gauntlet feed the Hallmark evidence marks shown on this case page.",
  "recorded_proofs": "Go-live proofs (2026-05-29) verified the engines on the live cluster for the limitation-of-liability clause family; coverage beyond curated fragments is future work.",
  "epistemic_skills_used": [
    { "skill": "Gauntlet", "contribution": "adversarial design review; caught and corrected the Catala overstatement; re-based the cross-check" },
    { "skill": "Manifest", "contribution": "mission custody across sessions with recorded checkpoints and operator acceptance gating" },
    { "skill": "Evidence-locked UAT", "contribution": "preregistered expected and disconfirming observations for operator-facing acceptance" },
    { "skill": "Did It Land", "contribution": "verification-at-the-consumer discipline behind the recorded-check pattern and this case study's evidence records" }
  ],
  "claim_limits": [
    "Design documentation and recorded engine proofs, not a claim that every engine runs on every clause today.",
    "Per-clause computation from arbitrary prose is future work.",
    "The public evidence remains the two recorded drafting checks and the Hallmark design specimens.",
    "No legal advice and no coverage claim beyond the recorded clause family."
  ],
  "amended": "2026-09-23: one phrase removed from epistemic_skills_used and one from claim_limits. Nothing was added or rerun."
}