Feature Specification: Policy Generator — AI-drafted decision tables and DRDs

BLOCKED ON THE PIVOT (2026-08). This spec depends on 036, which was archived when HoloDeck pivoted to Temporal-first orchestration (see docs/product-specs/040-holodeck-temporal/spec.md). The workflow.schema.json, holodeck workflow run, and DRD surfaces referenced below were removed with the 036 overlay engine and no longer exist. The decision-table format and its loader survive; if this feature is revived it re-targets the D3 table-step design.

Feature Branch: 039-policy-generator Created: 2026-07-25 Status: Archived — depended on 036, which was replaced by Temporal-first orchestration (040); the workflow-engine surfaces it targeted no longer exist Author: justinbarias (with Claude) Depends on: docs/product-specs/036-deterministic-spine — built only after 036's MVP ships Corpus: corpus-manifest.md (pinned, hashed)

Why this spec exists now, unbuilt

036's schema is soft today and hardens with every task after T4. Writing this spec now — before building any of it — lets the generator's requirements land in 036 as one-line additions rather than migrations later. That is not hypothetical: 036's source: annotation was added to T2 early for exactly this reason, before any generator existed.

Two requirements have already been pushed back into 036 as a result:

036 task	What it adds	Why the generator needs it
T3a	provenance: block on DecisionTable	Without it the engine cannot tell an LLM-written table from a hand-written one
T6	Review gate — run refuses generated_by without reviewed_by	Closes the authoring-time hole in "the LLM is never the spine"

If nothing else from this spec is ever built, those two still earn their keep — they make 036 honest about a capability the ecosystem will add with or without us.

Objective

Draft a candidate workflow.yaml DRD and its tables/*.dmn.yaml from a policy document, for a human to review, correct, and commit. The generator never produces a runnable determination on its own: 036's review gate refuses to execute what it emits until a person has signed it.

In scope: decision tables, the DRD that composes them, and TODO markers where edge nodes must be hand-authored.

Out of scope: generating edge agents, gate schemas, or agent.yaml files. The generator marks where they go; a human writes them. Inventing the type system and the prompts is a different, much larger problem.

The one rule, extended

036's invariant is the LLM is never the spine — the model feeds typed inputs, never verdicts. A generator moves the model upstream of the rules themselves. The invariant survives only if generated policy cannot execute unreviewed:

AI may draft policy. AI may never enact policy.

Enforced in 036 by FR-030/SC-009, not here — the engine is the right place for it, and it must hold whether or not this generator is the thing that wrote the table.

User Scenarios

User Story 1 — Draft tables from a policy section (Priority: P1)

An FDE points the generator at a policy document section. It emits one or more tables/*.dmn.yaml with provenance.source citing the clause, generated_by set, and reviewed_by absent. The FDE reviews, corrects, adds their name as reviewer, and commits.

Independent Test: run against a pinned corpus document; assert every emitted table parses, validates against workflow.schema.json, passes T3's static FEEL check, and is refused by holodeck workflow run until a reviewer is recorded.

User Story 2 — Draft the DRD (Priority: P2)

The generator proposes the decomposition: which decisions exist, how they compose, and where facts must enter. Edge nodes appear as TODO markers naming the fact required and the clause requiring it.

User Story 3 — Refuse to map discretion (Priority: P1)

Where a clause is not mechanizable — "the Secretary is not satisfied that the matter directly prevented…" — the generator MUST emit an explicit unmapped marker citing the clause, not an invented rule.

This is the acceptance-critical behaviour. A generator that quietly invents a threshold for a discretionary clause produces a table that looks authoritative and is wrong in the one way that matters. Silence here is worse than failure.

User Story 4 — Fidelity evaluation (Priority: P1)

Generation accuracy is scored against a pinned golden corpus of policy documents with hand-authored expected tables.

Requirements

• FR-001: MUST emit tables/*.dmn.yaml conforming to 036's published schema, passing T3's static FEEL validation.
• FR-002: MUST populate provenance: generated_by, source, source_doc, source_sha256. MUST NOT populate reviewed_by — only a human does that.
• FR-003: MUST emit an explicit unmapped marker, citing the clause, for any provision it cannot express — never an invented rule (US3).
• FR-004: MUST emit a candidate workflow.yaml DRD with TODO markers where edge nodes and input_data are required (US2).
• FR-005: MUST NOT emit agent.yaml files or gate schemas.
• FR-006: MUST score against pinned corpus snapshots, never a live URL.
• FR-007: MUST report table fidelity as a number and decomposition quality qualitatively (see Open Question 1).

Success Criteria

• SC-001: ≥80% rule-level match against the golden corpus. (Threshold provisional — see Open Question 2.)
• SC-002: 100% of known-discretionary clauses produce an unmapped marker rather than a rule. No partial credit; this is a safety property, not a quality metric.
• SC-003: 100% of emitted tables are refused by holodeck workflow run until reviewed (inherits 036 SC-009).
• SC-004: Every emitted table parses and passes static FEEL validation — a generated artifact is never merely plausible, it is always well-formed.

Corpus

Pinned in corpus/, hashed in corpus-manifest.md. Domain: Australian employment services compliance.

The anchor is the Reasonable Excuse Determination 2018 (7pp), which contains the full difficulty gradient in one document:

• s5(2)(a)–(j) — a closed enumerated list of ten factors → a gate schema, not a table
• s5(3) — nested sub-tests → mechanizable
• s6(3) — "the Secretary is not satisfied… directly prevented" → must refuse to map (SC-002)
• s6(4) — four conjunctive conditions defeated by any of four exceptions → the hardest positive case

A generator that drops s6(4)'s unless limb produces a table that wrongly denies excuses to people who did engage with treatment. That is the concrete failure SC-001 exists to catch, and it is why the corpus is real policy rather than documents we wrote ourselves.

Known risks

The corpus costs what the generator saves. Golden expectations mean hand-authoring the tables the generator is meant to produce. Accepted: keep N small (4–6 sections) and treat the corpus as a durable asset, not throwaway.

Real policy may resist tabulation. If the mechanizable fraction is low, the fidelity number is measured over a narrow base. That is itself a finding worth publishing — what fraction of real policy is determinable is a question the target buyer will ask. The Determination 2018 is encouraging: three of four provisions are mechanizable, and the fourth is a clean negative case.

Policy drifts under you. The Guidelines were already at v1.24 (1 July 2026) against search metadata's v1 July 2025 — at least annual revision. Mitigated by hashing; a mismatch means re-derive expectations, not "the test is broken".

Optics. This domain determines welfare penalties. The framing obligation on 036's sample README applies here too: the point is constraining automated decision-making, and that must be stated, not inferred.

Open Questions

1. Does the fidelity SC score decomposition, or only table content? Recommendation: tables numerically, decomposition qualitatively — "is this the right set of decisions" has no clean oracle.
2. Is 80% the right SC-001 threshold? Provisional until the corpus is built; set it from a measured baseline, not a guess.
3. How is an unmapped marker represented — a provenance.unmapped: [] list, a sibling report file, or inline comments? Must survive review and reach the reviewer's eye.
4. Cross-source consistency: does the generator produce the same table from the statute (Determination 2018) and from its provider-facing restatement (Guidelines pp. 271–273)? A useful robustness check, possibly SC-005.
5. Delivery shape — a HoloDeck agent with response_format set to the table schema (dogfooding), or a dedicated CLI verb (holodeck workflow draft)?

Not doing

• Generating edge agents, gate schemas, or prompts
• DMN XML import (separate, deferred — see 036/dmn-yaml-mapping.md route 2)
• Auto-committing, auto-reviewing, or any path that lets generated policy execute unreviewed
• Fine-tuning or training on the corpus — it is an evaluation set, not training data