Cognitive ladder & disciplines¶
Alongside the habitat maturity model, the assessment places the team on the cognitive read — the six-level AI collaboration literacy ladder from The Sovereign Engineer. This measures what the team can think and do.
The six levels¶
| Level | Name | What's visible in the repo when a team is here |
|---|---|---|
| L0 | Aware of the landscape | AI tools may be used, but nothing in the repo encodes that fact. No instruction files, no AI-aware conventions, no captured prompts. |
| L1 | Communicating through prompts | Some AI-instruction file exists (.github/copilot-instructions.md, CLAUDE.md, .cursorrules, .windsurfrules, AGENTS.md). Usually thin — style hints, a few do/don't bullets. |
| L2 | Verification discipline | The above, plus deterministic guardrails: linting in CI, coverage thresholds, pre-commit hooks, PR review that catches AI drift. The team can detect when output has drifted. |
| L3 | Habitat design | A persistent collaboration environment: rich CLAUDE.md/AGENTS.md, an explicit constraint document (HARNESS.md), custom skills/commands/hooks, a reflection log, decision records, AI-aware onboarding. |
| L4 | Specification-led | Specs are first-class: a specs/ directory, plans separated from specs, spec-first ordering, adversarial spec review at a plan-approval gate, orchestrated multi-step agent workflows. |
| L5 | Sovereign engineering | Platform-level practice: cross-team templates or a published plugin, governance audit cadence, decision archaeology, fitness functions in CI, cost/model-routing discipline, portfolio-level artefacts. |
The cognitive ladder runs L0–L5 (six rungs); the model's dimensions run L1–L5 (five). L0 is "aware but nothing encoded" — on the model's dimensions, that is the L1 floor.
The three disciplines¶
Every cognitive level rests on three disciplines:
- Context Engineering — how the team encodes accumulated wisdom into the agent's context window (instruction files, skills, onboarding, the reflection-to-curation loop).
- Architectural Constraints — how the team makes structural rules machine-checkable (HARNESS.md, CI gates, fitness functions; deterministic vs agent-enforced vs by-convention).
- Guardrail Design — how the team designs the feedback loops that catch drift (test suites, pre-tool hooks, adversarial review, output validation, plan- and integration-approval gates).
The scoring heuristic¶
The assessed cognitive level is the highest level where the team has substantial evidence across all three disciplines. The weakest discipline is the ceiling.
So strong specs with weak verification is L2, not L4. Strong guardrails with no encoded context is L2, not L3. See Coherence, not level for the reasoning behind "the weakest discipline is the ceiling".
The Governance dimension of the model is the operational face of the Architectural Constraints discipline; the report keeps the two consistent — the dimension is the one-line placement, the discipline score is the deeper read.