Three theses
Of the eighteen, these three are load-bearing. The rest of the argument rests on them.
-
The keyboard has been demoted.
Writing source code is no longer where human leverage lives. Humans who spend their day at the keyboard are competing with the model at the one thing it is unambiguously better at. The job has moved.
-
Specification is the new programming.
A PRD-as-directory, a diagram canon, and a Gherkin acceptance suite are now the artifacts of senior engineering. The compiler used to live downstream of source code. The compiler is now the model, and it lives downstream of the contract.
-
If your contract is weak, autonomy is hallucination at scale.
A loose PRD plus an unconstrained agent produces a confident, fluent, deeply wrong system. Velocity without specification is a faster path to a worse outcome.
Failure modes — how the factory breaks
Eight named failure modes. A factory without a named catalog has them anyway; it just discovers them at production scale.
- F1Context poisoning
- F2Semantic drift in parallel swarms
- F3Fan-out / fan-in synthesis errors
- F4Automation theater
- F5Orchestrator as attack target
- F6Supply-chain attacks on the harness surface
- F7Verifier poisoning
- F8Monorepo blast radius
Each failure mode is solvable. None is solved by hoping it doesn't happen. Every one needs a named control point — a contract gate, a routine, an invariant, a tool restriction.
The invariants
The non-negotiable claims that hold across every harness the architect emits. Seven shown here; the full table sits in §08 of the document.
- Strict TDD No production code without a failing test first. The Three Laws of TDD, literally enforced.
- F.I.R.S.T. tests Fast, Independent, Repeatable, Self-validating, Timely. Slow tests fail CI.
- Mutation gate Mutation kill rate above threshold on critical modules. Catches tests that execute without asserting.
- Adversarial validation A red-team subagent on a different model family attempts to break the code on every PR touching critical-tier modules.
- Acceptance discipline Gherkin scenarios protected from agent modification. Six-layer defense via CODEOWNERS, branch protection, and CI.
- Output verification An independent, different-family model verifies that the agent did what it said. Same-family verifiers share blind spots.
- Worktree-per-task Every code-writing task runs in its own git worktree. No shared workspace, no cross-task corruption.
The standard
One rule sits above the theses and resolves them when they disagree: the standard the factory computes toward. Every gate, every sub-agent, every control point exists to satisfy it. Read §01 of the manifesto for the full statement; the rest of the document derives from there.
Where to go next
- The full v0.6 manifesto — eighteen theses, the full failure catalog, the invariant table, the integration-surface taxonomy.
- The dsf repository — the corpus, the architect, the surfaces, the atlas.
- Request beta access — the operator side.