Tier-4 Agent Host conformance

Tier-4 Agent Hosts run the Bancroft layer DAG (or a module-contributed equivalent) and apply the resulting write_uw_section / write_multiple_sections tool calls back into the file. Hosts MUST:

• Honor the layer dependency graph in BANCROFT_LAYERS (UW_PROTOCOL_v1.md Part IX).
• Validate every agent tool call against the schemas in WRITE_UW_SECTION_TOOL and WRITE_MULTIPLE_SECTIONS_TOOL.
• Strip any _meta / _notes Claude includes inside section_data and apply the canonical _meta constructed from buildMeta().
• Append a pipeline_log entry for every successful agent run.

Why shape assertions

LLM nondeterminism makes byte-equality testing impractical at this tier. Each fixture's expected output is a JSON Schema describing the shape an acceptable result must have — required fields, value enums, plausible ranges — not a literal expected file.

Fixtures

fixtures/<scenario-id>/
├── before.uw.md                Input deal file
└── expected-after-shape.json   JSON Schema for the post-run file shape

Provided scenarios

Scenario	Layer	Validates
fixtures/l6-risk-rating	L6	Layer produces a risk_assessment section with overall_rating, risk_score, and key_risks; pipeline log gets a new entry; existing sections are unchanged.
fixtures/l0a-scope-deterministic	L0a (Scope)	Scope agent walks the asset-class fallback cascade for every required input and stamps results provisional. Shape contract: must_write_sections (noi_model, debt_structure, valuation, market_analysis, gaps), must_stamp_meta (source: asset_class_default, actor: agent/L0a, provisional: true), must_advance_stage_to: scope.

Profile contract

A separate profile/ subdirectory asserts that every entry in BANCROFT_LAYERS declares the right consumed_profile. This is a lightweight contract test — not LLM-driven — that catches accidental profile-policy drift introduced by future layer additions.

profile/<scenario-id>/
└── expected-layer-profiles.json   { layer_id: consumed_profile, … }

Scenario	Validates
profile/consumer-profile-contract	Sorted (layer_id → consumed_profile) map matches the recorded baseline. Update only when intentionally changing a layer's profile.

Running

Tier 4 is operator-driven, not part of CI — LLM calls cost money and are nondeterministic. The reference runner ships a lint-only pass that parses each before.uw.md, validates each expected-after-shape.json as a well-formed JSON Schema, and runs the profile-contract assertion:

node scripts/run-conformance.mjs --tier=4

Live agent runs against the shape are out of scope for the reference runner; hosts validate their own post-run files against the shape using ajv or equivalent.