Project Evolution

How Dazzle grew — and where to read the primary record of that growth.

Most of this site is greenfield documentation: it tells you how to build with Dazzle as it is today. This page is the opposite. It is the curated front door to Dazzle's historical record — the analyses, retrospectives, and snapshots produced while the framework was being built — so the reasoning behind today's design stays inspectable without cluttering the learning path.

These linked documents are point-in-time

Every document in the archive below is a dated snapshot and carries a banner saying so. Many describe gaps that have since been closed or decisions that were later superseded. They are kept for provenance, not as current guidance. For durable decisions, read the ADRs; for the released timeline, read the CHANGELOG.

The thesis that stayed constant

From early on, Dazzle has been built as a prior-correction substrate: a DSL whose restricted grammar, strong conventions, and conformance gates exist to counteract the statistical biases an LLM brings from its training data. The DSL is consumed primarily by AI agents, so it favours precision and formal correctness over human ergonomics. That framing — restrict the grammar, bias the inference, filter the output — is the through-line connecting almost every decision recorded below. The Counter-Prior Catalogue is the operational expression of it.

How it grew

A few overlapping arcs account for most of the archive:

• A typed UI substrate. The UI runtime moved off Jinja2 to a frozen-dataclass Fragment tree rendered to HTML (ADR-0023), with contract verification asserting DOM shape against the AppSpec. Much of the gap archive comes from hardening this layer.
• An autonomous-improvement loop. What is now /improve began as /ux-cycle: a loop that walks the framework, finds gaps, files issues, and fixes them. Its framework_gap_analysis strategy periodically synthesises what it has learned into a standalone document — those syntheses are the framework-gap analyses archived below. They are the loop thinking out loud about where the framework was thin.
• Authorization as algebra. Access control converged on a formal predicate algebra: permit: for who, scope: for which rows, with as: clauses binding filters to personas (ADR-0010), statically validated against the FK graph and backed by a provable-RBAC package.
• Tenancy and identity. Row tenancy pivoted toward framework-owned discriminators and generated Postgres RLS; identity grew into a global-identity + organization + fenced-membership model with two-phase auth.
• A failure-mode lens. Most recently, that whole history was turned into a threat model: the Model-Driven Failure Modes catalogue generalises 4GL/CASE/MDE history into named risks Dazzle scores itself against. The dated gaps below are the empirical instances; the catalogue is the generalisation.

The archive

The dated framework-gap analyses (April 2026 /ux-cycle → /improve output) and the snapshots, briefings and retrospectives that used to be linked here were moved out of the working tree under ADR-0057. They live under archive/docs-history/ on the archive branch, alongside the superpowers plans and specs and the research notes.