blueprint

name	blueprint
description	Converts a Synth candidate into a formal promotion proposal. Names before/after behavior, defines benchmark gates, lists affected targets. Requires Attestation (human sign-off) before Augment or Factory may proceed.
layer	L8

Blueprint

You define what "done" looks like before anyone writes code.

What this skill does

Blueprint reads a candidate.json from Synth, reads the current SKILL.md of the affected module, and produces a blueprint.json that specifies: the exact before/after behavior change, the benchmark gate that must pass, the build targets affected, and the verification mode required at Forge. Blueprint is the last human checkpoint before experimental implementation begins.

Blueprint does not write implementations. It does not modify modules. It produces the spec that Augment or Factory works from.

When to use

Blueprint activates when:

1. A candidate.json exists in .wabblespec/state/experiments/candidates/ with status: pending
2. Human explicitly invokes Blueprint with the candidate ID
3. The affected module's SKILL.md is readable

Inputs

Input	Path	Required
Candidate	.wabblespec/state/experiments/candidates/{id}.candidate.json	Yes
Module SKILL.md	modules/{layer}/{module}/SKILL.md	Yes — read before/after states from here
Module skill-rules.json	modules/{layer}/{module}/skill-rules.json	Yes — verification_mode and authority
Framework manifest	framework.yaml	Yes — confirms module ID and layer

Decision logic

1. Read candidate.json fully.
2. Read the affected module's SKILL.md. Identify the specific section that describes the behavior to be changed.
3. State before_behavior: quote or paraphrase the current SKILL.md description of the behavior.
4. State after_behavior: what the SKILL.md should say after the change.
5. Determine if this is AUGMENT (modifying existing module) or NEW (net-new module). This routes to Augment vs Factory.
6. Define benchmark_gates: minimum one gate with a metric, threshold, and developer outcome link. Use benchmark.schema.json fields.
7. Identify affected_targets: which L3 platform packages are affected by this change. List them explicitly.
8. Determine verification_mode for Forge: always Attestation for self-modification. Review or Demonstration for other changes.

Surface open questions to the human before writing if any of the above cannot be determined from the available inputs.

Reference Routing

Situation	Reference
Blueprint receipt write	engine/shared/references/script-delegation-contract.md → receipt-writer.py --type blueprint

Output contract

One file: .wabblespec/state/experiments/blueprints/{candidate-id}.blueprint.json

{
  "blueprint_id": "matches candidate_id",
  "candidate_ref": "path to source candidate.json",
  "change_type": "AUGMENT | NEW",
  "affected_module": "module-id",
  "before_behavior": "Quoted or paraphrased current SKILL.md description of the behavior.",
  "after_behavior": "What the SKILL.md will say after the change.",
  "benchmark_gates": [
    {
      "metric": "rate | count | latency_ms | boolean",
      "threshold": 0.8,
      "direction": "lower_is_better | higher_is_better",
      "developer_outcome": "false_completion | missed_test | stale_evidence | rework",
      "rationale": "Why this threshold was chosen."
    }
  ],
  "affected_targets": ["cli", "web"],
  "verification_mode": "Attestation | Review | Demonstration",
  "attestation_required": true,
  "created_at": "ISO-8601",
  "status": "awaiting-attestation"
}

After writing, request human Attestation. Blueprint status changes from awaiting-attestation to approved only after human signs off in writing. Augment and Factory must not proceed before status is approved.

Failure modes

Before/after ambiguity — before_behavior or after_behavior is vague ("module will work better"). Fix: quote specific SKILL.md lines for before; write exact replacement text for after.

Benchmark theater — benchmark_gates lists a metric with no developer_outcome link. Fix: every gate must name one of the four developer outcomes from benchmark-discipline.md.

Scope creep at Blueprint — blueprint changes more than one behavior. Fix: one blueprint per candidate. Multiple behavior changes require multiple Synth candidates.

Not tested

Blueprint cannot verify that the after_behavior is implementable or that the benchmark threshold is achievable. Augment implements; Benchmark validates. Blueprint's job is precision of specification.