cortex-x

Training data is frozen; about the current state of the world — framework and model versions, pricing, a11y standards — I know nothing reliable. So any answer or implementation that depends on external state must be verified before I state it as fact. AI is a tool, not an oracle.

R1 anchors verify-first as a mandatory step: the standard describes the protocol, cortex-goal.md embeds it into Phase 3 (Research) of the plan, and the sprint skill enforces it inside the pipeline. Findings are cited with URLs and cached, so research is reusable across sessions rather than one-shot.

A diff that wrote itself cannot judge itself. A non-trivial change passes six independent reviewers before the operator is allowed to merge — and a second round, the Pass-2 skeptic, is tasked with refuting the first round's findings.

r2-review.js is a dynamic workflow orchestrating six parallel agents; pre-commit-review-gate.cjs blocks the commit until a verdict arrives. Consensus HIGH findings are applied in-commit, not deferred to a backlog.

Arc 1: 23 HIGH bugs caught, zero refuted by the Pass-2 skeptic, all fixed in-commit before push — nothing escaped to main. In Sprint 2.46 six independent reviewers converged on a single bug at confidence 99/98/96/96/95/92.

Discipline that relies on human memory fails exactly when fatigue is highest. A rule you can bypass by forgetting is not a rule — it belongs outside the head, in a cryptographic artifact.

Signed verdict v2 carries an HMAC-SHA256 or Ed25519 signature over a payload containing commit_sha (cross-checked against HEAD), staged_tree (the index contents), workflow_run_id (a single-use nonce in a journal), and secret_tier (env > persisted random > host-derived). A replayed verdict is rejected, a stale verdict is rejected, and a host-derived secret is rejected under STRICT_SECRET=1. The signed verdict complements [skip-review] as a second unblock path beside the session marker — the session marker remains the highest-priority allow path.

When an agent writes code faster than a human reads, documentation diverges from reality within a single sprint. Stale docs are worse than none — they impersonate a truth they no longer hold.

cortex-doc-currency.cjs lints documentation against the measured repository state and flags stale numbers and claims; cortex-doc-regen.cjs regenerates derived passages from the source of truth. The documentation.md standard defines what is derived — and thus regenerable — versus hand-written.

Every new session starts with empty memory. What changes in weeks — current state — must not live in the same place as what is stable for years — institutional wisdom — otherwise it duplicates and drifts apart.

cortex-load.md is a mental cheat-sheet that brings a new session up to context; memory-decay.cjs governs memory aging so stale entries do not outweigh fresh ones; lessons.cjs manages the lessons-learned layer. One source of truth for each piece of information — never duplicate.

When there are 35 standards and all carry equal weight, under time pressure none gets applied. Rules need an explicit priority order so you can decide what yields when two budgets conflict.

Tier 0 (Ship-Ready) → 1 (SSOT/modularity) → 1.5 (coding behavior) → 2 (security/testing/correctness) → 3 (process) gives the standards a lexicographic order. RULE-1.md codifies the SSOT layer; action-kinds.cjs maps the 21 action_kinds to rules; code-review.md applies that same order during review.

A pattern that works in one project alone may be coincidence. Only when the same standard improves both RELO and the multi-tenant chatbot platform is it a demonstrably transferable method, not a local trick.

Everything shared lives in ~/.claude/shared/, so standards, hooks, and skills load into every project identically — README.md is the index of 35 standards and auto-orchestrate.cjs nudges parallel agents regardless of which project is running. Centralization is the mechanism of transferability.

One brilliant sprint proves nothing; a method you can run again and again at the same quality proves everything. The bottleneck is not model capability but the operator's consistent operational discipline.

/cortex-sprint wraps the whole cycle — plan → R1 → implementation → R2 → verdict → capture — into a single repeatable skill; sprint-pipeline.md defines the phases and gates. A sprint plan like sprint-2-44-plan.md is a documented instance of that same process, not a unique performance.