Fitness Methodology¶

The Agent-Led Fitness Methodology is an optional V&V loop that checks whether your Dazzle app is fit for the purpose described by your spec.md.

It differs from dazzle ux verify --contracts in that it asks semantic questions — does this persona actually make progress through their lifecycles? does the DSL cover everything the spec implies? — rather than mechanical ones.

When to use¶

Run on every CI cycle if your project's [dazzle.maturity].level is mvp
Run on every PR if your project's maturity is beta
Run weekly (soft mode only) if your project's maturity is stable

Running¶

# Full cycle
dazzle fitness run

# Just findings
dazzle fitness findings

# Story paraphrase-confirm loop
dazzle fitness confirm-stories

MCP users:

mcp__dazzle__fitness.run()
mcp__dazzle__fitness.findings(axis=conformance)

Configuration¶

Add to pyproject.toml:

[dazzle.maturity]
level = "mvp"          # or "beta" / "stable"

[dazzle.fitness]
max_tokens_per_cycle = 100000
independence_threshold_jaccard = 0.85

[dazzle.fitness.independence_mechanism]
primary = "prompt_plus_model_family"

Required DSL additions¶

Every entity participating in fitness must declare fitness.repr_fields:

entity Ticket "Support Ticket":
  id: uuid pk
  title: str(200) required
  status: enum[new, in_progress, resolved] required
  assignee_id: ref User

  fitness:
    repr_fields: [title, status, assignee_id]

v1 emits a non-fatal lint warning if this is missing. v1.1 makes it fatal.

Entities with lifecycles must also declare a lifecycle: block (see ADR-0020).

Findings¶

Findings live in dev_docs/fitness-backlog.md. Each row has:

axis: coverage vs conformance
locus: implementation | story_drift | spec_stale | lifecycle
severity, persona, capability_ref
evidence_embedded: self-contained evidence envelope, durable after the underlying ledger has expired

Three corners¶

The methodology triangulates across three independent sensors:

spec.md — your natural-language oracle
DSL stories — /bootstrap's interpretation of your intent
Running app — what the code actually does

Each cycle measures independence_jaccard between corners 1 and 2 to verify the sensors haven't collapsed into a single (correlated) signal. When they do, all findings from that cycle are marked low_confidence=true and cannot auto-correct.