Load the spec-writing standard first. A cell is in exactly one mode: maintenance (common) or rebuild (rare, deliberate). There is no per-edit regeneration of specs, and no gap analysis for routine work.
Maintenance mode
Ordinary contract-guarded agent work. The instruction shape is: "fix this; do not touch the contract; if the fix forces a contract re-evaluation, stop and report." The nlspec is a guardrail here, not a build input. The two-pump rule applies: maintenance never amends the nlspec directly. Maintenance does not waive the security carve-out: cells handling credentials, authentication or authorization decisions, irreversible or destructive operations, or data migrations are human-reviewed regardless of cell locality (see slice-design's review scope).
Condemn triggers
The methodology runs on replace economics: budget-bounded cells are cheap to rebuild from their contracts, so drift needs triggers, not baselines. Condemn a cell when any of these fires:
- budget breach — the primary tripwire; degraded internals show up as volume first (duplication, defensive scaffolding, flag soup);
- contract-check failure — actual surface no longer matches the stated contract, or something reaches past a contract into internals;
- undemonstrable claim — a claim routed through the cell can no longer be demonstrated;
- seam spray — a supposedly cell-local change touches other cells;
- friction — a change that ought to be cell-local repeatedly fails to be, or a fix takes multiple attempts inside one cell. Drift announcing itself before a mechanical wire trips.
Check budgets mechanically on every slice; keep them slightly tight so the wire trips while the rebuild is still cheap.
Rebuild protocol
The order is fixed:
- Harvest. Run the catalytic converter over the dying cell's tests. A condemned cell's tests are the only record of everything its ephemeral slice specs never wrote down.
- Condemn. Mark the cell for rebuild; route pending maintenance around it.
- Rebuild. Design the replacement with slice-design, from the nlspec, conditioned on the surviving elements — a rebuild driven by requirements change keeps what is still expected to succeed; it is never a fresh stochastic sample.
- Demonstrate. Prove the routed claims per /proof-of-work before the rebuild lands.
Archived slice specs from the cell's history are available as forensics; they are not load-bearing inputs.
Catalytic converter
Elevates behavioral evidence from code into claims, up to inclusion in the nlspec — checkable memory of hard-won decisions, without freezing how they were implemented.
- Intake: test cases only. Never opinions, rationale prose, or language-specific "lessons" — the moment the converter accepts those it becomes a third authority.
- Elevation test: does this test encode intent that would otherwise survive only by accident — a boundary decision, an invariant, a regression a future rebuild must not reintroduce? Routine coverage stays in the code.
- Mechanics: propose elevated claims to HITL as nlspec amendments (this is pump 2), classify each cell-local or integration, and record its provenance.
Run the converter routinely at slice exit (see slice-design) and mandatorily at harvest.