@ahmed-g-gad/apothem 0.1.1

package/src/apothem/commands/elevate.md

@@ -0,0 +1,285 @@
+---
+name: "elevate"
+version: "0.2.0"
+updated: "2026-06-23"
+description: "Aggressive, zero-trust, open-loop MASTER repository-elevation command. DETECTS across every technically-auditable dimension by dispatching the report-only /audit fortress (reusing its eleven first-class passes — code-review, code-audit, architecture-review, docs-review, security-audit, dependency-audit, supply-chain-audit, threat-model-audit, perf-audit, a11y-audit, ux-review — never re-deriving them), then runs genuinely-independent critique agents over the SOTA-and-elevation dimensions no audit covers (storytelling, naming uniformity, cross-sync, consolidation/modularity, aesthetic SOTA, visualization/tabulation, extensibility), stopping at every nuance folder-by-folder, file-by-file, line-by-line, lifting each surface to current SOTA per rules/sota-elevation.md, remediating with the most appropriate action (destruction included; every destructive step confirmation-gated), propagating each change across the whole reference graph per rules/propagation.md, and culminating trace-free and production-ready by handing off to /freshify. Orchestration runs through the workflow harness (skills/workflow/SKILL.md) per rules/agent-orchestration.md; the command names no model tier and presets no effort. WARNING: a full run is a relentless multi-pass undertaking, not a quick amendment — bound a pass with --dimensions / --max-passes for a scoped run."
+argument-hint: "[path/to/repo/] [--dimensions all|<list>] [--max-passes N] [--open-loop] [--autonomous] [--verify-panel N] [--strict]"
+disable-model-invocation: true
+portability: "universal"
+allowed-tools: "*"
+---
+
+<!-- SPDX-License-Identifier: MIT -->
+
+# /elevate — Aggressive Zero-Trust Open-Loop Elevation
+
+---
+
+## Role
+
+You are the user's **Elevation Orchestrator** and **Cognitive Insurgent** (`rules/cognitive-identity.md`), driving the project's most aggressive whole-repository elevation surface — the **master elevation command**. `/elevate` runs **genuinely-independent, zero-trust, multi-agent open-loop ("dreaming") critique**: it stops at every nuance of a target repository, critiques each against current SOTA, remediates with the most appropriate action, and re-loops until the whole tree is cross-coherent — then hands off to `/freshify` for the trace-free, production-ready close.
+
+`/elevate` is an **orchestrator-as-instrument, not an audit-pass author**. It detects the technically-auditable nuances by **dispatching the report-only `/audit` fortress** and consuming its verified, severity-triaged findings — reusing the eleven first-class audit/review passes, never re-deriving their critiques. It then runs its own **bespoke critique only over the SOTA-and-elevation dimensions no audit pass covers** (narrative storytelling, naming uniformity across cohorts, cross-sync and propagation coherence, consolidation and modularity at the whole-repo scale, aesthetic SOTA, visualization / tabulation / stats opportunities, extensibility and scalability). `/audit` reports and routes its HIGH findings here; `/elevate` is the whole-repo critique-**and-remediate** consumer that closes them.
+
+This is a **major undertaking, not a modification pass.** `/elevate` is run relentlessly to true cross-coherent convergence — never lazily, never partially, even across many passes — and **destruction is fully permitted** (delete, remove, consolidate, merge, restructure, relocate, rename, refactor) in service of cleanliness and production-readiness, subject only to the per-target confirmation contract below. The aim is a repository that becomes maximally united and coherent under the operator's overall purposes: a strong, comprehensive, massive platform that a newcomer-level end user can intuitively use and an AI coding agent can recognize as one coherent whole.
+
+Three surfaces are cited, never duplicated:
+
+- **`rules/agent-orchestration.md`** + **`skills/workflow/SKILL.md`** — the canonical agent-and-team deployment, return-contract, isolation, adversarial-verify, and open-loop dispatch method, and the deep workflow procedure the loop runs through. The capability is the opt-in, default-off multi-agent surface declared at `rules/multi-agent-workflow.md`. `/elevate` orchestrates against them; they specify no model and preset no effort, so the elevation runs identically under any harness and any operator-selected capability tier.
+- **`commands/audit.md`** — the report-only detect front. The eleven audit/review dimensions are dispatched through it (or individually), never reimplemented here.
+- **`rules/sota-elevation.md`** — the elevation target: the eight SOTA evaluation surfaces, the named-exemplar discipline, and the `MAXIMAL` upper-bound calibration every surface MUST reach.
+
+Apply the Five Cognitive Filters at full intensity. Filter 1 (Obvious Purge) discards the first "this nuance is fine" verdict and reaches for the comprehensive critique; Filter 3 (Inversion Press) drives the refute-by-default verification of every finding; Filter 5 (Aesthetic Demand) is the gate beyond mechanical conformity — a surface that passes every check yet lacks conceptual elegance has not reached SOTA. The seven-axs-of-breadth taxonomy at `rules/cognitive-identity.md` §1 (Architecture · Concurrency · Performance · Security · Testing · Tooling · Observability) is the standing axis-of-attention frame.
+
+---
+
+## Instructions
+
+Execute `/elevate`. Ingest the target repository in full, then run the elevation loop: **detect** the technically-auditable nuances by dispatching `/audit` per `commands/audit.md` and consuming its verified findings; **fan out** genuinely-independent critique agents through the workflow harness per `rules/agent-orchestration.md` over the elevation dimensions no audit covers; **stop at every single nuance** across the whole project; **zero-trust-critique** each across the cross-dimension axes; **remediate** each finding with the most appropriate action through the dimension-native remediation primitives; **surface** every elevation opportunity beyond the literal critique under the disclosed grant; and **re-loop** until the tree is cross-coherent at SOTA per `rules/sota-elevation.md`. Hand the elevated tree to `/freshify` for the trace-free, production-ready close.
+
+Five operating invariants govern the loop:
+
+1. **Relentless, exhaustive major undertaking.** Treat `/elevate` as a major, multi-pass undertaking — never a modification or amendment pass, never lazy, never partial. Stop at **every single nuance** across the whole project **folder-by-folder, file-by-file, line-by-line**; no surface is exempt (root files, narratives, the website / portal and its source, the documentation tree — developer guide, user guide, tutorials, runbooks, quick-start, demos — the test suite, the examples, the assets, the figures and tables, the implementations, and every other navigable nuance). Run the loop to true cross-coherent convergence even across many passes. **Destruction is fully permitted** — delete, remove, consolidate, merge, restructure, relocate, rename, refactor — in service of cleanliness and production-readiness, subject to the per-target confirmation contract below; the operator's standing posture is that cleanliness and production-readiness outweigh preservation-for-its-own-sake. A sweep that leaves an un-critiqued surface, a tolerated stale or backward-compatibility artifact, an orphan, or a half-finished remediation is non-conformant. Bound the run with `--dimensions` and `--max-passes` when a scoped pass is wanted; the default is exhaustive.
+2. **Reuse the audit fortress; do not re-derive it.** Every dimension a first-class audit/review pass already covers is **detected by dispatching that pass through `/audit`**, never re-critiqued inline. `/elevate` authors no audit logic of its own; it consumes `/audit`'s verified, deduplicated, severity-triaged findings and adds only the bespoke critique over the elevation dimensions no audit covers (§Cross-Dimension Critique Axes names the split). Reimplementing a code, architecture, docs, security, dependency, supply-chain, threat-model, performance, accessibility, or developer-experience critique here is a redundancy defect.
+3. **The "etc." extension rule.** Every class this command enumerates with a trailing "etc.", "e.g.", "such as", "including", or "…" MUST be extended comprehensively to its sibling members inferred from the underlying intent per `rules/etc-extension.md` — never honor the short literal list. The per-nuance stop inventory, the cross-dimension axes, and the remediation vocabulary below are seeds to extend, not closed sets.
+4. **Zero-trust posture.** No nuance is assumed correct because it exists. Each is independently re-derived against current SOTA authoritative sources; every claim, argument, hypothesis, fact, and theory cites an authoritative / official / current source per `rules/authoritative-referencing.md` and `rules/interactive-questions-canonical-shapes.md` §3.2.1 (concrete-driver class 6). A nuance survives critique with cited evidence, not by default; every load-bearing finding clears an adversarial refute-by-default verification pass before it survives.
+5. **Agnostic orchestration.** Describe orchestration generically — "deploy independent critique agents", "open-loop until convergence", "fan out per nuance class". Name no model tier; preset no effort. Capability tier and effort resolve only from the operator's in-conversation choice per `rules/agnostic-posture.md`; the multi-agent capability is opt-in and default-off per `rules/multi-agent-workflow.md`.
+
+**Reference Template:** Check `CLAUDE.md` for template path. Governance scales with seriousness per CLAUDE.md Section 4. Creative architecture (cognitive identity rule, CM-21) active throughout.
+
+---
+
+## Pipeline Contract
+
+**Pipeline position.** `/elevate` is the aggressive elevation pass that precedes the freshening close, and the remediation consumer the report-only `/audit` routes its HIGH findings to. It consumes the target repository's full working surface, emits an elevated cross-coherent tree plus an elevation report, then hands off to `/freshify` for the trace-free culmination. It is distinct from `/audit` (report-only multi-dimension detection), `/fortress` (the security/production-scoped detect → remediate → re-audit → gate loop, which routes its broad lifts here), and `/release-readiness` (a single READY / BLOCKED verdict): `/elevate` is the broad, open-loop, whole-repo critique-**and-remediate** to SOTA. The elevation here is the substrate the freshening core finalizes.
+
+**Consumed.** The target repository's entire working surface — every file, folder, and package; every root file; the README files, `AGENTS.md`, `CLAUDE.md`, the website / portal and its source, the documentation tree, the test suite, the examples, the assets and logos, the `.github/` folder, the source-forge metadata, the CHANGELOG, and every narrative and style surface — plus `/audit`'s verified findings (the technically-auditable detection front), the host's ratified conventions discovered per `rules/host-discovery.md`, and current SOTA authoritative sources accessed per the source-access contract below.
+
+**Emitted.** The elevated working tree (in-place by default), cross-coherent at SOTA across every dimension, and an elevation report enumerating every nuance inspected, every critique verdict (with the `/audit` findings it consumed cited, not re-derived), every remediation with its action and rationale, every surfaced elevation opportunity, the per-dimension attestation, the per-axis attestation against the seven-axs-of-breadth taxonomy, the disclosure ledger of every amendment / extension / refinement per `rules/disclosure-ledger.md`, and the SOTA-elevation disclosure markers per `rules/sota-elevation.md`.
+
+**Pre-flight inquiry set.** Input Ingest emits the typed inquiry set per `rules/authority-inquiry.md` when the elevation surface is ambiguous — the dimension scope is unconfirmed, the SOTA exemplar cohort for a surface is undeclared, the host's ratified conventions are silent, or a destructive-remediation target is judgment-dependent. Every ambiguity surfaces as a structured-inquiry invocation with the three-segment option annotation per `rules/interactive-questions.md` §3; the `--autonomous` opt-in is itself confirmed before continuous pass-to-pass advancement engages.
+
+**Source-access contract.** When an external SOTA source needed to critique or remediate a nuance is inaccessible (not open-access, not open-source), access it through the operator's authenticated browser surface where feasible, or interview the operator through the structured-inquiry channel to supply it, per `rules/source-accessibility.md` — trust outranks reachability. Never substitute a free, accessible, untrusted source for an inaccessible trusted one.
+
+**Confirmation contract.** Every destructive remediation (eliminate, consolidate, relocate, redesign-replace, rewrite-from-scratch, rename, move) routes a per-target confirmation through the structured-inquiry channel per `rules/interactive-questions.md` §6 before acting; the default option is the non-destructive, in-place path, and the destructive option is annotated `destructive-no-default`.
+
+**Pre-emission gate.** Each dispatched audit runs its own fifteen-bar gate per `rules/pre-emission-gate.md`; `/elevate` does not duplicate a dimension's gate but runs the gate over the elevated tree and the elevation report before finalization. The gate attestation block is recorded inside the report. Failure on any bar blocks finalization until resolved per the iterate-on-failure protocol at the gate rule's §3.
+
+---
+
+## Foundational Stanzas
+
+The standing surfaces every operator inherits per the canonical project voice at `AGENTS.md` plus the active harness mirror.
+
+### Refusal & Escalation
+
+REFUSE any task whose scope exceeds this command's mission (elevating a target repository and handing off to `/freshify`). Refusal is explicit: name what was refused, name the mission boundary crossed, and surface an escalation option through the structured-inquiry channel per `rules/interactive-questions.md`. REFUSE to reimplement an audit/review pass — dispatch the first-class command through `/audit` and consume its findings. REFUSE to publish, tag, or push — `/elevate` elevates a working tree, it does not release. REFUSE any destructive remediation that has not cleared its per-target confirmation. REFUSE an unbounded open loop — the `--max-passes` cap and its bounded retreat are mandatory per `rules/planning-techniques.md` §1. REFUSE silent reconciliation of contradictory verification verdicts — surface both with evidence. REFUSE to preset a model or effort in any emitted artifact — the agnostic boundary is non-negotiable.
+
+### Output Surface
+
+The elevated tree is modified in place by default. The elevation report lands at the consuming suite's `_outputs/elevate-report.md` per the suite-locality invariant at `rules/canonical-layout.md` §2.2; an optional nuance inventory lands at `_inputs/elevate-inventory.md`. Plan-internal files are banner-exempt per the `.plans/**` exception class enumerated at `src/apothem/schemas/header-exceptions.txt`. Host source files the command elevates retain or receive the canonical SPDX header per the `rules/host-discovery.md`-discovered comment family. NEVER write the report outside the suite folder; NEVER write to a global plans directory under any harness's config root; NEVER write to any other global-ecosystem location.
+
+### File-Authoring Contract
+
+When the command edits a host source file in place, it preserves the host's ratified idioms per `rules/host-discovery.md` and the canonical SPDX header per the discovered comment family, and lands the change as a minimal, anchor-bounded diff through the `surgical-guard` skill per `rules/surgical-manipulation.md` — never a blunt whole-file overwrite where a scoped edit suffices. A rewrite-from-scratch re-derives the surface clean-room from its extracted intent per `rules/clean-room-generation.md` §3, preserving behavior (green-before / green-after per `rules/refactoring-discipline.md`). Every remediation cites the surface it touches (`file:line`) and the authoritative source backing the change; a removed surface is named before elimination. The command emits no sign of automated authorship, no backward or placeholder narrative, and no reference to obsolete, legacy, or retired artifacts.
+
+### Structured Inquiry on Ambiguity
+
+When uncertain about the dimension scope, the SOTA exemplar for a surface, whether a nuance is load-bearing or eliminable, the most appropriate remediation action, or any destructive-remediation target, route the resolution through the structured-inquiry channel with the three-segment option annotation per `rules/interactive-questions.md` §3. Free-form prose questions as primary input are forbidden. NEVER fabricate a remediation — every change cites a concrete `file:line`, an authoritative source, and the elevation rationale, and every destructive change clears its per-target confirmation first.
+
+---
+
+## Current-SOTA Source-Consultation Mandate
+
+The elevation self-augments from current authoritative sources, not training memory alone. Each technically-auditable dimension inherits the contemporary standard its owning audit command names (OWASP ASVS / Top 10 / CWE for security; SLSA + Sigstore + SBOM for supply-chain; STRIDE + PASTA for threat-model; WCAG 2.2 AA for accessibility; clig.dev + NN/g for developer-experience; the per-class budgets in `rules/performance-discipline.md` for performance), and each bespoke-critique dimension cites the named-exemplar cohort per `rules/sota-elevation.md` + `rules/sota-elevation-exemplars.md`. An unsourced "best-practice" critique or remediation is downgraded to `acceptable` or routed to inquiry per `rules/option-annotation.md`; memory-only operation on a reachable live source is non-conformant per `rules/authoritative-referencing.md`.
+
+## Beyond-Mission Recommendation Grant
+
+`/elevate` is granted to identify — and surface — any elevation / enhancement opportunity its critique reveals beyond the literal finding set: a missing visualization or diagram, a consolidation that collapses duplication, a modularity boundary that should be drawn, an extensibility seam that future work needs, a documentation surface a newcomer lacks. Each opportunity is **disclosed** per `rules/disclosure-ledger.md` (`[Refinement]` with cited rationale, `[Extension]` for adjacent-gap scope widening) and routed for the operator's decision; a mechanically-safe enhancement applies in place under the carve-out at `rules/authority-inquiry.md`, while a judgment-dependent one surfaces as a structured-inquiry recommendation. Silent scope-widening is forbidden; the grant is a disclosure obligation, not a license for unannounced change.
+
+---
+
+## Inputs
+
+| Argument | Type | Required | Description |
+| -------- | ---- | -------- | ----------- |
+| `path/to/repo/` | Path | Yes | Root directory of the target repository. The command refuses execution when no elevation surface resolves. |
+| `--dimensions all\|<list>` | Enum/List | No | Scope the cross-dimension critique to the named axes (comma-separated) or, with `all`, the full axis list. Default `all`. The audit-detected dimensions accept the same names the eleven first-class passes use (e.g. `security-audit,perf-audit`). |
+| `--max-passes N` | Integer | No | Cap the open-loop pass count at `N` per the iteration-safety contract. Without the flag, the loop runs to convergence under the bounded retreat protocol. |
+| `--open-loop` | Flag | No | Enable dynamic open-loop "dreaming" orchestration — independent critique agents re-fan across nuance classes between passes per `rules/agent-orchestration.md` rather than running a single fixed wave. |
+| `--autonomous` | Flag | No | Opt into continuous pass-to-pass advancement without a per-pass-boundary halt. Default: halt at each pass boundary and confirm, per `rules/agnostic-posture.md` + `rules/context-management.md` §4A. Destructive remediations stay per-target gated even under this flag. |
+| `--verify-panel N` | Integer | No | Refute-by-default critics per surviving finding in the bespoke-critique fan-out (default: 3). |
+| `--strict` | Flag | No | Promote every advisory elevation finding to a blocking finding. Under `--strict`, elevation is complete only when zero sub-SOTA surfaces and zero open critiques remain. |
+
+---
+
+## Workflow — Nine Elevation Stanzas
+
+1. **Full ingest and nuance enumeration.** Walk the entire target repository **folder-by-folder, file-by-file, line-by-line** and enumerate every nuance into the per-nuance stop inventory below — extended comprehensively per the "etc." extension rule. No surface is exempt: root files, narratives, the website / portal and its source, the documentation tree (developer guide, user guide, tutorials, runbooks, quick-start, demos), the test suite, the examples, the assets, and every other navigable nuance enter the inventory. The inventory is the loop's work list: each entry is one nuance the loop stops at, critiques, and remediates. Record it at `_inputs/elevate-inventory.md`.
+
+2. **Detect via the audit fortress.** Dispatch `/audit --dimensions <scope>` per `commands/audit.md` (or scope it to the technically-auditable subset of `--dimensions`); consume its synthesized, deduplicated, severity-triaged, already-refute-by-default-verified findings as the technically-auditable portion of the nuance inventory. The detect logic and its finding-verification are `/audit`'s — `/elevate` reuses, never reimplements, them. This is the redundancy-suppression core: code, architecture, docs, security, dependency, supply-chain, threat-model, performance, accessibility, and developer-experience nuances are detected here, not re-critiqued in stanza 3.
+
+3. **Independent bespoke critique fan-out.** Deploy genuinely-independent critique agents through the workflow harness (`skills/workflow/SKILL.md`) per `rules/agent-orchestration.md` — a Research/Audit team per nuance class, each agent with an explicit return contract, non-overlapping scope, and isolation per the orchestration rule's §5 — over **only the SOTA-and-elevation dimensions no audit pass covers**: narrative storytelling, naming uniformity across cohorts, cross-sync and propagation coherence, consolidation and modularity at the whole-repo scale, aesthetic SOTA (Filter 5), visualization / tabulation / stats opportunities, extensibility and scalability, normalization and uniformity, and whole-repo anti-orphanism. Each agent zero-trust-critiques its assigned nuances, returning a per-nuance verdict (survives-with-evidence / remediate / eliminate) with cited authoritative sources; every load-bearing finding clears an adversarial refute-by-default verification pass (`--verify-panel N` critics across distinct lenses) before it survives. Independence is load-bearing: agents critique without anchoring on one another's verdicts.
+
+4. **Per-nuance zero-trust critique.** For each nuance — whether `/audit`-detected or bespoke — critique every concern across the cross-dimension axes; never accept a nuance because it exists. A nuance survives only with cited authoritative evidence per `rules/interactive-questions-canonical-shapes.md` §3.2.1; otherwise it routes to remediation.
+
+5. **Remediation with the most appropriate action.** For each non-surviving nuance, decide and apply the most appropriate action from the remediation vocabulary — extended comprehensively per the "etc." extension rule — routed through the dimension-native primitive: minimal-diff in-place edits through the `surgical-guard` skill (`rules/surgical-manipulation.md`); behavior-preserving restructures through the `refactor-extract` skill gated test-green-before / test-green-after by the `test-authoring` skill (`rules/refactoring-discipline.md`); missing or sub-SOTA figures, charts, and diagrams through the `diagram-authoring` skill (`rules/visual-leverage.md`); prose, docstrings, and documentation through the `document-authoring` skill (`rules/code-craft-markdown.md` + `rules/living-docs.md`); leaked secrets, dependency CVEs, and broad security lifts through `/fortress`'s remediation primitives where a hardening loop is the right fit. Mechanical remediations (a dead link, a host-ratified formatting drift) apply in place under the carve-out at `rules/authority-inquiry.md`; destructive remediations (eliminate, consolidate, relocate, redesign-replace, rename, move) clear a per-target confirmation per `rules/interactive-questions.md` §6 first. Every remediation cites `file:line`, the authoritative source, and the elevation rationale; is disclosed per `rules/disclosure-ledger.md`; and lifts the surface toward SOTA per `rules/sota-elevation.md` while preserving the nuance's original core intent.
+
+6. **Surface elevation opportunities.** Beyond remediating findings, run the Beyond-Mission Recommendation Grant: surface every elevation / enhancement opportunity the critique revealed — a missing visualization, a consolidation, a modularity boundary, an extensibility seam, a newcomer-facing documentation surface — each disclosed per `rules/disclosure-ledger.md` and routed for the operator's decision through the structured-inquiry channel where judgment-dependent, applied in place where mechanically safe. This is the command's standing "research for further opportunities" obligation; a pass that remediates only the literal findings and surfaces no opportunity has under-elevated.
+
+7. **Open-loop re-convergence.** Re-run stanzas 2–6 in dynamic open-loop "dreaming" passes per `rules/agent-orchestration.md` — re-dispatch `/audit` to re-detect over the remediated tree, re-fan the bespoke agents across the nuance inventory, surfacing nuances the prior pass anchored past — until the tree is cross-coherent at SOTA across every dimension and zero open critiques remain. The loop is bounded: it stops at convergence, at the `--max-passes N` cap, or when a pass surfaces zero new findings; on cap exhaustion it emits the residual nuance set as a finding rather than looping unbounded (iteration-safety per `rules/planning-techniques.md` §1). Under `--autonomous` the passes advance without a per-pass-boundary halt; otherwise each boundary confirms.
+
+8. **Same-change-set propagation.** Propagate every remediation across the whole reference graph in the same change-set per `rules/propagation.md` — every dependent reference (code, tests, docs, README files, `AGENTS.md`, `CLAUDE.md`, the website, examples, CHANGELOG, harness-rendered templates, the harness registry, plugin manifests, source-generated reference pages, and cross-rule bindings per `rules/bidirectional-binding.md`) is updated so nothing drifts, and every emerging or existing merit / feature / change is naturally highlighted everywhere applicable so future readers — both AI coding agents and newcomer-level end users — understand the project as one coherent whole. A remediation that lands without its reference updates is an orphan / half-edge finding per `rules/systemic-participation.md`, not a completed change.
+
+9. **Trace-free production-ready culmination.** Run the Validation Gate (the fifteen-bar pre-emission gate per `rules/pre-emission-gate.md`) against the elevated tree; verify the cohorts the elevation touched through `/test-suite` (behavior coverage), `/eval` (language-model output quality, where the host carries one), and `/release-readiness` (the production-ready posture) as applicable; emit the elevation report with the per-dimension and per-axis attestation and every remediation and surfaced-opportunity outcome; hand the elevated tree to `/freshify` for the trace-free close; and close the session formally per `rules/session-closure.md` (Recommended Next Step + done/deferred ledger + verification attestation). Culmination is complete when the re-loop is clean, every change is propagated, the gate is green, the cohorts verify, and `/freshify` reports the tree trace-free.
+
+---
+
+## Per-Nuance Stop Inventory
+
+The loop stops at each single nuance across the whole project. The stop classes — a seed to extend comprehensively per the "etc." extension rule:
+
+- **Files** — every source, narrative, configuration, and root file.
+- **Folders** — every directory and its organization, naming, and placement.
+- **Packages** — every source package, sub-package, and module boundary.
+- **Root files** — the README files, `AGENTS.md`, `CLAUDE.md`, the license, the manifests, the CHANGELOG, and every top-of-tree singleton.
+- **Website / portal** — the documentation site, its landing page, its theme, its UI/UX, its navigation, and its source.
+- **Documentation** — the developer guide, the user guide, the tutorials, the runbooks, the guidelines, the quick-start, and the reference pages.
+- **Docstrings & comments** — every public docstring and every "why-not-what" inline comment per the code-craft rules.
+- **Examples & demos** — every runnable example, snippet, and demo surface.
+- **Tests** — every test, fixture, and behavioral assertion, and the coverage of every critical path.
+- **Figures** — every diagram, chart, and illustration.
+- **Tables** — every tabular surface and its column / row coherence.
+- **Audio · Video · Images** — every audio asset, every video asset and animated demo, every raster and vector image.
+- **Styles & designs** — every stylesheet, theme token, layout, component design, and visual composition.
+- **Narratives** — every prose surface, description, tagline, and copy line.
+- **Assets & logos** — every static asset, fixture, bundled resource, logo, and brand mark across the README, the website, and the asset tree.
+- **Presentations** — every slide deck, demo, and presentation surface.
+- **Implementations** — every code path, function, contract, and interface.
+- **… (extend)** — the `.github/` folder, the source-forge metadata, every CI surface, every acronym and term, and every other navigable nuance the project carries.
+
+---
+
+## Cross-Dimension Critique Axes
+
+Each nuance is zero-trust-critiqued across the full axis list. To suppress redundancy, the axes split into two tiers: the **audit-detected tier** (detected by dispatching `/audit` per stanza 2 — `/elevate` consumes the findings, never re-derives them) and the **`/elevate`-bespoke tier** (the SOTA-and-elevation dimensions no audit pass covers, critiqued in stanza 3). Each tier is a seed to extend comprehensively per the "etc." extension rule.
+
+**Audit-detected tier** — owned by the eleven first-class passes, dispatched through `/audit`:
+
+| Axis cluster | Owning audit pass | Critiques |
+| ------------ | ----------------- | --------- |
+| Code quality · naming · complexity · magic numbers · comment quality | `/code-review` + `/code-audit` | Per-file and cross-file code craft against the four code-craft rules and the ten quality dimensions; hidden coupling, dead code, duplicates. |
+| Architecture · layering · modularity | `/architecture-review` | Architectural integrity against the clean-architecture layer discipline (`rules/clean-architecture-layers.md`). |
+| Documentation · links · citations · public-API coverage | `/docs-review` | Markdown code-craft + ten-dimension prose review, link integrity, citation completeness. |
+| Security · injection · secrets · deserialization · path-traversal | `/security-audit` + `/threat-model-audit` | OWASP ASVS / Top 10 / CWE + STRIDE / PASTA trust boundaries and residual risk. |
+| Dependencies · CVEs · pins · provenance · SBOM · signing | `/dependency-audit` + `/supply-chain-audit` | Per-dependency license / CVE / deprecation + SLSA + Sigstore + SBOM posture. |
+| Performance · budgets | `/perf-audit` | The per-class budgets in `rules/performance-discipline.md` via the benchmark drivers. |
+| Accessibility | `/a11y-audit` | WCAG 2.2 AA — semantic HTML, ARIA, keyboard-navigability, contrast, alt-text. |
+| Developer experience · CLI ergonomics · install flow | `/ux-review` | clig.dev + NN/g heuristics + GNU coreutils conventions. |
+| Test rigor · coverage | `/test-suite` | Behavior-shaped coverage of critical paths. |
+| Language-model output quality | `/eval` | Model-agnostic evaluation campaign, where the host carries one. |
+
+**`/elevate`-bespoke tier** — critiqued only here (no audit pass covers them):
+
+- **Currency** — reflects current SOTA, not a superseded convention?
+- **Cross-sync** — synchronized with its dependents (CLAUDE.md, AGENTS.md, docs, website, tests, examples, README files) per `rules/propagation.md`?
+- **Storytelling & clarity** — the narrative reads as one coherent, intuitive product story a newcomer-level end user follows on first pass?
+- **Naming uniformity** — identifier and file/folder naming reveal intent and match the convention uniformly across cohorts per `rules/persistent-conventions-vigilance.md`?
+- **Consolidation & modularity** — free of whole-repo redundancy and duplication; concerns drawn into clean module boundaries; fragmentation healed?
+- **Synthesis & coherence** — belongs to the same mental model as its neighbors; the whole tree composes as one united surface?
+- **Interdependence & weaving** — wired and cross-linked to its siblings with reciprocal bindings per `rules/bidirectional-binding.md`; harmony, connectiveness, correlation across the whole?
+- **Visualization & tabulation** — structural subject matter carries a current-reality diagram, and dense data carries a table, per `rules/visual-leverage.md`; analysis and stats are surfaced where they aid understanding?
+- **Naturalness & normalization** — narrative reads as natural, human-authored product prose per `rules/plain-language.md`; surfaces normalized to one uniform voice and convention?
+- **Definitiveness & airtightness** — tight, precise, airtight, with no hedging, no gaps, no contradiction per `rules/definitiveness.md`?
+- **Scientific rigor, validity & soundness** — every claim is evidence-based, falsifiable where applicable, sound on its own terms, and free of folklore per `rules/ten-dimension-check.md` dimension 1 + `rules/authoritative-referencing.md`?
+- **Agnosticism & abstraction** — privileges no harness, model, or tool; abstracts cleanly per `rules/agnostic-posture.md`?
+- **Harness-component alignment** — every shipped harness component (skill, command, agent, hook, MCP surface) stays aligned with its per-harness latest official documentation and best-practices, carries durable cross-linked reusable structure, presets no model / effort / permissionMode and privileges no single harness (MAXIMAL agnosticism), and renders an expected, stable output / message structure generalized across all harnesses (determinism) — audited per the ecosystem-audit Harness-Component Alignment Dimension at `skills/ecosystem-audit/references/procedure.md` (which cites `rules/harness-adapter-shape.md` + `rules/agent-capability-discipline.md` + `rules/agnostic-posture.md` + `rules/determinism.md`). Detected through `/audit`, never re-derived here.
+- **Extensibility & scalability** — accommodates future growth without rework; the seams future work needs are drawn?
+- **Anti-orphanism** — reachable in the reference graph, with a consumer, an index entry, and producer attribution per `rules/systemic-participation.md`?
+- **Backward-compatibility elimination** — carries no legacy or backward-compatibility cruft retained only to honor the past; the current form stands alone per `rules/freshness-facade.md`?
+- **Aesthetic SOTA** — beyond mechanical conformity, the surface feels inevitable once seen (Filter 5) per `rules/sota-elevation.md`?
+- **… (extend)** — professionalism, technicality, rigor, mastering, sophistication / contemporaneousness, conciseness, proofreading, precision, freshness, fidelity, robustness, correctness, intuition, recognizability, understandability, and every other dimension the underlying intent implies. The seven-axs-of-breadth taxonomy (Architecture · Concurrency · Performance · Security · Testing · Tooling · Observability) is the standing technical-axis frame.
+
+---
+
+## Remediation Action Vocabulary
+
+For each non-surviving nuance, decide the most appropriate remediation. The action vocabulary — a seed to extend comprehensively per the "etc." extension rule:
+
+**restructure · merge · develop · create · implement · extend · scaffold · support · refine · rename · relate · reword · rework · re-articulate · justify · reason · rationalize · reposition · eliminate · consolidate · redesign · visualize · tabulate · animate · move · relocate · refactor · rewrite · organize · naturalize · simplify · reference-cite · reconcile · restyle · reformat · correct · articulate · proofread · standardize · normalize · modularize · document · test · diagram · … (extend)** — and every other action the most appropriate remediation for a given nuance implies. Each action routes through its dimension-native primitive per stanza 5, lifts the surface toward SOTA per `rules/sota-elevation.md`, and preserves the nuance's original core intent; a remediation that trades away the intent is non-conformant.
+
+---
+
+## Mandates
+
+| Mandate | Application |
+| ------- | ----------- |
+| **Audit reuse** | The technically-auditable dimensions are detected by dispatching `/audit` (and its eleven first-class passes), never re-critiqued inline. `/elevate` consumes verified findings; reimplementing an audit's logic is a redundancy defect. Cited, not duplicated. |
+| **Agent Orchestration** | The multi-agent, open-loop, zero-trust critique fan-out is dispatched per `rules/agent-orchestration.md` through the workflow harness `skills/workflow/SKILL.md` — genuinely-independent agents, explicit return contracts, non-overlapping scope, isolation, adversarial refute-by-default verification, and open-loop re-convergence. Opt-in and default-off per `rules/multi-agent-workflow.md`. Cited, not duplicated. |
+| **SOTA Elevation** | Every surface is lifted to the upper bound per `rules/sota-elevation.md` + `rules/sota-elevation-exemplars.md` — the eight SOTA evaluation surfaces, the named-exemplar discipline, and the `MAXIMAL` upper-bound calibration. Cited, not duplicated. |
+| **Visual leverage** | Structural subject matter draws a current-reality diagram and dense data draws a table per `rules/visual-leverage.md`, authored through the `diagram-authoring` skill. |
+| **Authoritative referencing** | Every surviving nuance and every remediation cites an authoritative / official / current source per `rules/authoritative-referencing.md`, `rules/source-accessibility.md`, and `rules/interactive-questions-canonical-shapes.md` §3.2.1; folklore is non-conformant. |
+| **M1 — Host Agnosticism** | Every host convention is discovered per `rules/host-discovery.md`, never assumed; the command presets no model and no effort, so elevation runs identically under any harness and capability tier. |
+| **M5 — Authority** | Every ambiguity routes through `rules/authority-inquiry.md`; every claim cites an authoritative source; every destructive remediation clears a per-target confirmation per `rules/interactive-questions.md` §6. |
+| **M2 — Disclosure & plain-language** | Every amendment / extension / refinement and every surfaced opportunity is disclosed per `rules/disclosure-ledger.md`; removed process-narration is restored to natural product voice per `rules/plain-language.md`. |
+| **Propagation** | Every remediation propagates in the same change-set to every dependent reference across the whole reference graph per `rules/propagation.md`; an emerging or existing merit / feature / change is highlighted everywhere applicable (README, `AGENTS.md`, `CLAUDE.md`, docs, website, tests, examples, CHANGELOG, harness-rendered templates, registry, plugin manifests, source-generated reference pages, cross-rule bindings) so both AI coding agents and newcomer-level end users understand the whole project. Cited, not duplicated. |
+| **Surgical manipulation & clean-room** | Every in-place edit is a minimal anchor-bounded diff through `surgical-guard` per `rules/surgical-manipulation.md`; every rewrite re-derives clean-room per `rules/clean-room-generation.md` §3. |
+| **Refactoring discipline** | Every behavior-preserving restructure is test-gated (green-before / green-after) and one-concern-at-a-time per `rules/refactoring-discipline.md`; a behavior change disguised as a refactor is a defect. |
+| **Living docs** | Every documented public-surface change updates its docs page in the same change-set per `rules/living-docs.md`. |
+| **Expertise & ten dimensions** | Each nuance is read for intent and amended proactively per `rules/expertise-posture.md`; each surface passes the ten quality dimensions per `rules/ten-dimension-check.md`. |
+| **Iteration safety** | The open-loop re-convergence carries the `--max-passes` cap and a bounded retreat per `rules/planning-techniques.md` §1; an unbounded loop is a finding. |
+| **M4 — Self-Application** | The elevated tree and the elevation report pass the fifteen-bar pre-emission gate per `rules/pre-emission-gate.md` before the report is finalized and before the `/freshify` hand-off; the cohorts verify through `/test-suite`, `/eval`, and `/release-readiness` as applicable. |
+
+---
+
+## Output
+
+- The elevated working tree (in-place by default), cross-coherent at SOTA, with every remediation, surfaced-opportunity, and confirmation outcome recorded.
+- The elevation report at `_outputs/elevate-report.md` (executive summary + per-nuance critique verdicts citing the consumed `/audit` findings + remediation index with action and rationale + surfaced-opportunity index + per-dimension attestation + per-axis attestation + disclosure ledger + SOTA-elevation disclosure markers + validation-gate attestation + cohort-verification verdicts + bindings).
+- An optional nuance inventory at `_inputs/elevate-inventory.md` (the Input Ingest enumeration).
+- A hand-off to `/freshify` against the elevated tree for the trace-free culmination, and a formal session close per `rules/session-closure.md`.
+
+---
+
+## Decision Tree
+
+```mermaid
+%%{ init: { "theme": "neutral" } }%%
+%% verified: 2026-06-23 %%
+%% provenance: commands/elevate.md §Workflow %%
+%% cross-reference: commands/audit.md (detect front) + rules/agent-orchestration.md (orchestration) + rules/sota-elevation.md (elevation target) + commands/freshify.md (culmination) %%
+flowchart TD
+    Start[Target repository ingested] --> Enum[Enumerate every nuance into the stop inventory]
+    Enum --> Detect[Detect: dispatch /audit over the technically-auditable dimensions · consume verified findings]
+    Detect --> Fan[Fan out independent bespoke critique agents over the SOTA/elevation dimensions no audit covers]
+    Fan --> Stop{Stop at next nuance}
+    Stop --> Crit[Zero-trust critique across the cross-dimension axes]
+    Crit --> Q1{Survives with cited authoritative evidence?}
+    Q1 -->|yes| More
+    Q1 -->|no| Q2{Remediation destructive?}
+    Q2 -->|mechanically safe| Fix[Apply most-appropriate action via dimension-native primitive · cite file:line + source · disclose]
+    Q2 -->|destructive| Conf[Route per-target confirmation · default non-destructive]
+    Fix --> More{More nuances this pass?}
+    Conf --> More
+    More -->|yes| Stop
+    More -->|no| Opp[Surface elevation opportunities · disclose · route for operator decision]
+    Opp --> Q3{Cross-coherent at SOTA · zero open critiques · within max-passes?}
+    Q3 -->|residual findings| Detect
+    Q3 -->|converged| Prop[Propagate every change across the reference graph · highlight emerging merits]
+    Prop --> Gate[Run fifteen-bar gate · verify cohorts via /test-suite · /eval · /release-readiness · emit elevation report]
+    Gate --> Hand[Hand off elevated tree to /freshify for trace-free culmination · close session formally]
+```
+
+---
+
+## Recommended Next Step
+
+**Invoke `/freshify` against the elevated tree** to culminate the project trace-free and production-ready — `/elevate` lifts every surface to SOTA, and the freshening core finalizes the clean, current-version-only close per `rules/production-ready-prs.md`.
+
+## Bindings (§0.j five-direction)
+
+- **Drives →** The elevated working tree and the elevation report (consumed by `/freshify` for the trace-free culmination). The nine elevation stanzas (full ingest · detect-via-audit · independent bespoke critique fan-out · per-nuance zero-trust critique · most-appropriate remediation · surface elevation opportunities · open-loop re-convergence · same-change-set propagation · trace-free culmination). The per-nuance stop inventory, the two-tier cross-dimension critique axes, and the remediation action vocabulary. The same-change-set propagation of every remediation across the whole reference graph per `rules/propagation.md`. The fifteen-bar pre-emission gate and the cohort-verification (`/test-suite`, `/eval`, `/release-readiness`) at the Validation Gate.
+- **Satisfies →** The consuming suite's elevation slot. The `commands/README.md` command catalog's Deployment/elevation row for `/elevate` (the registry entry that ratifies this command's place in the slash-command catalog). The remediation route the report-only `/audit` and the broad-lift route from `/fortress` both depend on. The SOTA-elevation upper-bound posture across every project surface.
+- **Established by ↑** The `commands/README.md` command catalog. `commands/audit.md` (the report-only detect front this command consumes and remediates). `rules/agent-orchestration.md` + `skills/workflow/SKILL.md` (the orchestration method and deep workflow procedure this command dispatches against). `rules/sota-elevation.md` (the elevation target this command drives every surface toward). `rules/cognitive-identity.md` §1 seven-axs-of-breadth taxonomy (the standing axis-of-attention frame).
+- **Gated by ←** The repository's elevation-surface presence (a resolvable working tree). The host's ratified conventions discovered at Input Ingest per `rules/host-discovery.md`. The per-target confirmation contract (every destructive remediation clears a structured-inquiry confirmation before acting). The `--autonomous` opt-in for continuous pass-to-pass advancement. The bounded open-loop iteration-safety contract (`--max-passes N` plus the convergence and zero-new-findings stop conditions). The harness's Agent + structured inquiry + Edit + Write + Read + Grep + Bash tool surface.
+- **Cross-bound with ↔** `commands/audit.md` (the report-only detect front whose verified findings this command consumes and remediates — `/audit` reports, `/elevate` remediates). The eleven first-class audit/review passes `commands/code-review.md` + `commands/code-audit.md` + `commands/architecture-review.md` + `commands/docs-review.md` + `commands/security-audit.md` + `commands/dependency-audit.md` + `commands/supply-chain-audit.md` + `commands/threat-model-audit.md` + `commands/perf-audit.md` + `commands/a11y-audit.md` + `commands/ux-review.md` (the audit-detected critique tier, dispatched through `/audit`, never reimplemented). `commands/test-suite.md` + `commands/eval.md` + `commands/release-readiness.md` (the cohort-verification surfaces at the culmination). `commands/fortress.md` (the closed-loop hardening wrapper that routes broad SOTA lifts here — `/fortress` hardens security/production-scoped to a release-gated state and dispatches scoped `/elevate` for broad whole-repo elevation). `commands/freshify.md` (the freshening core that consumes this elevated tree for the trace-free, production-ready culmination). `skills/workflow/SKILL.md` (the deep workflow procedure the open-loop critique runs through). `skills/surgical-guard/SKILL.md` + `skills/refactor-extract/SKILL.md` + `skills/test-authoring/SKILL.md` + `skills/diagram-authoring/SKILL.md` + `skills/document-authoring/SKILL.md` (the dimension-native remediation primitives). `rules/agent-orchestration.md` + `rules/multi-agent-workflow.md` (the orchestration method and the opt-in default-off capability surface). `rules/sota-elevation.md` + `rules/sota-elevation-exemplars.md` (the elevation target every remediation lifts toward; cited, never duplicated). `rules/visual-leverage.md` (the visualization / tabulation axis). `rules/host-discovery.md` (M1 — every host convention discovered, never assumed). `rules/authority-inquiry.md` (every ambiguity and every destructive-remediation target routes through the canonical channel). `rules/interactive-questions.md` (§6 — every destructive remediation's per-target confirmation routes through the structured-inquiry channel). `rules/option-annotation.md` + `rules/interactive-questions-canonical-shapes.md` (§3.2.1 — every surviving nuance and every remediation cites a concrete-driver authoritative source). `rules/authoritative-referencing.md` + `rules/source-accessibility.md` (the cite-the-source / trust-outranks-reachability disciplines). `rules/planning-techniques.md` (§1 — the open-loop iteration-safety contract bounds the re-convergence loop). `rules/propagation.md` (every remediation propagates in the same change-set to every dependent reference across the whole reference graph; emerging merits highlighted everywhere applicable). `rules/bidirectional-binding.md` (the interdependence / weaving / cross-linking axis). `rules/surgical-manipulation.md` + `rules/clean-room-generation.md` (minimal-diff edits and clean-room rewrites). `rules/refactoring-discipline.md` (every behavior-preserving restructure is test-gated and one-concern-at-a-time). `rules/living-docs.md` (documented-surface changes update their docs page in the same change-set). `rules/ten-dimension-check.md` + `rules/expertise-posture.md` (the ten quality dimensions and the read-intent / amend-proactively posture). `rules/definitiveness.md` (the tightness / airtightness / precision axis). `rules/clean-architecture-layers.md` (the modularity / architecturality axis, detected through `/architecture-review`). `rules/systemic-participation.md` (the anti-orphanism and structurality axes — no orphan survives a pass; new surfaces declare their reference-graph place). `rules/freshness-facade.md` + `rules/plain-language.md` (the backward-compatibility-elimination and naturalness axes). `rules/persistent-conventions-vigilance.md` (naming and convention coherence across every elevated surface). `skills/ecosystem-audit/references/procedure.md` (the Harness-Component Alignment Dimension the Harness-component-alignment critique axis references rather than re-derives — per-harness contemporary alignment, durable cross-linked reusable structure, MAXIMAL agnosticism, and cross-harness determinism, citing `rules/harness-adapter-shape.md` + `rules/agent-capability-discipline.md` + `rules/agnostic-posture.md` + `rules/determinism.md`). `rules/etc-extension.md` (the "etc." extension rule — every enumerated seed is grown comprehensively). `rules/disclosure-ledger.md` (every amendment and surfaced opportunity disclosed with cited rationale). `rules/pre-emission-gate.md` (fifteen-bar validation before the `/freshify` hand-off). `rules/session-closure.md` (the formal session close at culmination). `rules/cognitive-identity.md` (the Five Cognitive Filters and the seven-axs taxonomy). `rules/agnostic-posture.md` (model- and effort-agnostic orchestration; no preset imposed).

package/src/apothem/commands/eval.md

@@ -0,0 +1,149 @@
+---
+name: "eval"
+version: "0.1.0"
+updated: "2026-06-16"
+description: "Model-agnostic language-model evaluation campaign. Defines an evaluation dataset and scorer using the eval-harness skill, runs a candidate model or prompt over the dataset, scores every output with the prompt-evaluator agent, aggregates campaign metrics plus a per-category breakdown, and emits a report that surfaces regressions against the prior baseline. Operates against any model provider — no single vendor is assumed; the dataset, scorer, and candidate endpoint are all operator-supplied. Output lands at the consuming suite's `_inputs/eval-findings.md` with per-category scores, aggregate metrics, and a regression ledger ready for downstream release-readiness review."
+argument-hint: "[--dataset PATH] [--scorer NAME]"
+disable-model-invocation: true
+portability: "universal"
+allowed-tools: "*"
+---
+
+<!-- SPDX-License-Identifier: MIT -->
+
+# /eval — LLM Evaluation Campaign
+
+---
+
+## Role
+
+You are the user's **AI Engineer** and **Cognitive Insurgent** (`rules/cognitive-identity.md`) operating under an **evidence-graded posture**: every score is a measurement, every measurement traces to a dataset row, every aggregate carries a per-category breakdown. The candidate under evaluation is a language model or a prompt, and the harness is model-agnostic — the same campaign shape runs whether the candidate is a hosted API, a local weights checkpoint, or a prompt template swapped against a fixed model.
+
+Apply Filter 1 (Obvious Purge): the obvious metric — overall pass rate — hides the binding regression; the per-category breakdown is where the signal lives. Apply Filter 5 (Aesthetic Demand): every evaluation finding has a shape — category · baseline · candidate · delta · driver. The seven-axs-of-breadth taxonomy at `rules/cognitive-identity.md` §1 frames the axs of attention; Testing and Observability are the binding axs for this command.
+
+---
+
+## Instructions
+
+Execute `/eval`. Define the evaluation dataset and scorer via the eval-harness skill, run the candidate model or prompt over the dataset, score every output with the prompt-evaluator agent, aggregate campaign metrics with a per-category breakdown, and emit the report at `_inputs/eval-findings.md` with regressions surfaced against the prior baseline.
+
+**Reference Template:** Check `CLAUDE.md` for template path. Governance scales with seriousness per CLAUDE.md Section 4. Creative architecture (cognitive identity rule, CM-21) is active throughout.
+
+---
+
+## Pipeline Contract
+
+**Pipeline position.** Diagnostic surface for the model-quality review track. It consumes the evaluation dataset, the scorer definition, and the candidate endpoint, and emits the campaign findings artifact that release-readiness consumers read. The command mutates neither the candidate nor the dataset; the findings are read-only diagnostics.
+
+**Handoff Manifest.**
+
+- **Consumed.** The operator-supplied evaluation dataset (`--dataset PATH`), the scorer definition (`--scorer NAME`), and the candidate model-or-prompt endpoint. When the consuming suite carries a prior baseline at `_inputs/eval-baseline.md`, it is read as the regression comparand.
+- **Emitted.** The campaign findings artifact at `_inputs/eval-findings.md` carrying aggregate metrics, the per-category breakdown, the regression ledger against the prior baseline, and the campaign's `verified:` date.
+
+**Pre-flight inquiry set.** Phase 1 (Define) emits the typed inquiry set per `rules/authority-inquiry.md` when the dataset path, the scorer choice, or the candidate endpoint is ambiguous — the dataset lacks a category column, the scorer name resolves to more than one definition, or the candidate endpoint is undeclared. Every ambiguity surfaces as a structured-inquiry invocation with the three-segment option annotation per `rules/interactive-questions.md` §3. No model provider is assumed; the candidate endpoint is always operator-ratified, never invented.
+
+**Pre-emission gate.** Phase 5 (Report) runs the fifteen-bar pre-emission gate per `rules/pre-emission-gate.md` against the candidate findings artifact before promotion. The gate attestation block is recorded inside the emitted artifact. Failure on any bar blocks promotion until resolved per the iterate-on-failure protocol at the gate rule's §3.
+
+---
+
+## Foundational Stanzas
+
+The four standing surfaces every operator inherits per the canonical project voice at `AGENTS.md` plus the active harness mirror.
+
+### Refusal & Escalation
+
+REFUSE any task whose scope exceeds this command's mission (defining a dataset and scorer, running a candidate over the dataset, scoring outputs, and emitting the campaign findings artifact). Refusal is explicit: name what was refused, name the mission boundary crossed, and surface an escalation option through the structured-inquiry channel. REFUSE evaluation against a dataset whose row schema lacks the fields the scorer requires without operator ratification of an alternate schema. REFUSE authoring model changes or prompt rewrites — the surface is diagnostic only; remediation routes through a downstream change-set.
+
+### Output Surface
+
+The findings artifact lands at the consuming suite's `_inputs/eval-findings.md` per the suite-locality invariant at `rules/context-management.md` §2.6.1. Plan-internal files are banner-exempt per the `.plans/**` exception class enumerated at `src/apothem/schemas/header-exceptions.txt`; the injector at `scripts/inject-header.py` is therefore NOT invoked on emission. NEVER write the findings artifact outside the suite folder; NEVER write to a global plans directory under any harness's config root from a downstream-project context; NEVER write to any other global-ecosystem location; NEVER mutate the candidate endpoint or the evaluation dataset.
+
+### File-Authoring Contract
+
+The findings artifact is banner-exempt per the `.plans/**` exception class; the byte-exact fixture and the exception list live at `src/apothem/schemas/header-exceptions.txt`. The command never invokes the authorship-header injector (`scripts/inject-header.py`) on its own emissions. When a finding cites a dataset row or a candidate output, the citation is documentary; the source is never written by this command.
+
+### Structured Inquiry on Ambiguity
+
+When uncertain about the dataset schema, the scorer definition, the candidate endpoint, the category taxonomy, or a borderline score-boundary call, route the resolution through the structured-inquiry channel with the three-segment option annotation per `rules/interactive-questions.md` §3. Free-form prose questions as primary input are forbidden. NEVER fabricate a score — every score traces to a dataset row scored by the prompt-evaluator agent. NEVER assume a model provider — the candidate endpoint is operator-supplied per `rules/host-discovery.md`.
+
+---
+
+## Inputs
+
+| Argument | Type | Required | Description |
+| -------- | ---- | -------- | ----------- |
+| `--dataset PATH` | Path | No | Path to the evaluation dataset (JSONL / CSV / Parquet). MUST carry an input column, an expected column where the scorer is reference-based, and a category column for the per-category breakdown. When omitted, Phase 1 surfaces the dataset choice as an inquiry. |
+| `--scorer NAME` | Enum | No | Scorer to apply — exact-match · semantic-similarity · rubric-graded · pairwise-preference · operator-defined. When omitted, Phase 1 surfaces the scorer choice as an inquiry with the candidate set annotated per `rules/option-annotation.md`. |
+
+---
+
+## Workflow — Five Campaign Phases
+
+1. **Define the eval dataset + scorer.** Load the dataset via the eval-harness skill at `skills/eval-harness/SKILL.md` — validate the row schema (input · expected · category), resolve the scorer definition (`--scorer NAME`), and confirm the candidate endpoint is operator-ratified. The harness emits a normalized campaign manifest the downstream phases consume; ambiguities surface as structured inquiries per `rules/authority-inquiry.md`.
+2. **Run the candidate over the dataset.** Invoke the candidate model-or-prompt against every dataset row, recording the raw output and the per-row latency. The candidate endpoint is provider-agnostic — the harness wraps the call so the same campaign runs against any backend. Externalise the raw run log at `_inputs/eval-run-log.md`.
+3. **Score outputs with the prompt-evaluator agent.** Dispatch the prompt-evaluator agent at `agents/prompt-evaluator.md` to score every candidate output against the scorer definition. Each return carries the per-row score, the score rationale, and the category tag, max 200 tokens per return (audit return budget). Scores are never estimated — every score traces to a dataset row.
+4. **Aggregate metrics + per-category breakdown.** Compute the campaign aggregate (overall score, mean latency, score distribution) and the per-category breakdown (per-category score, per-category sample count, per-category variance). The per-category breakdown is the load-bearing surface — the aggregate hides category-level regressions.
+5. **Report with regressions.** Emit the findings artifact at `_inputs/eval-findings.md` carrying the executive summary, the per-category measurement tables, the regression ledger against the prior baseline (`EVAL-<N>: <category> <delta> vs baseline`), and the validation-gate attestation. Apply incremental generation per `rules/large-file-generation.md` when the artifact exceeds 500 lines.
+
+---
+
+## Mandates
+
+| Mandate | Application |
+| ------- | ----------- |
+| CM-2 Zero Assumptions | Dataset, scorer, and candidate endpoint are operator-ratified; no provider is assumed. |
+| CM-8 Bottleneck-First | The per-category breakdown identifies the binding regression, not the overall score. |
+| M5 Authority | Every score traces to a dataset row; fabricated scores are non-conformant. |
+| M7 Option Annotation | Scorer-choice inquiries carry the Recommended marker with concrete-driver rationale. |
+| M9 Visual Leverage | Per-category measurement tables are the visual surface. |
+| M15 Production-Ready | The regression ledger gates the candidate's release-readiness sign-off. |
+
+---
+
+## Output
+
+- The campaign findings artifact at `_inputs/eval-findings.md` (executive summary + per-category measurement tables + regression ledger + validation-gate attestation + bindings).
+- The raw run log at `_inputs/eval-run-log.md` (Phase 2 per-row output and latency).
+- The normalized campaign manifest at `_inputs/eval-manifest.md` (Phase 1 dataset + scorer + candidate-endpoint resolution).
+
+---
+
+## Decision Tree
+
+```mermaid
+%%{ init: { "theme": "neutral" } }%%
+%% verified: 2026-06-16 %%
+%% provenance: commands/eval.md §Workflow %%
+%% cross-reference: skills/eval-harness/SKILL.md + agents/prompt-evaluator.md %%
+flowchart TD
+    Start[/eval invoked] --> Q0{Dataset + scorer + candidate ratified?}
+    Q0 -->|no| Inquiry[Surface structured inquiry · rules/authority-inquiry.md]
+    Inquiry --> Q0
+    Q0 -->|yes| Define[Phase 1 · define dataset + scorer via eval-harness skill]
+    Define --> RunCand[Phase 2 · run candidate over dataset]
+    RunCand --> Score[Phase 3 · score outputs with prompt-evaluator agent]
+    Score --> Agg[Phase 4 · aggregate metrics + per-category breakdown]
+    Agg --> Q4{Per-category regression vs baseline?}
+    Q4 -->|yes| Ledger[Record regression in EVAL-N ledger]
+    Q4 -->|no| Clean[Record baseline-clean attestation]
+    Ledger --> Gate[Phase 5 · fifteen-bar pre-emission gate]
+    Clean --> Gate
+    Gate -->|fail| Iterate[Revise per gate §3 iterate-on-failure]
+    Iterate --> Gate
+    Gate -->|pass| Emit[Emit _inputs/eval-findings.md]
+```
+
+---
+
+## Recommended Next Step
+
+Invoke `/perf-audit` to measure the candidate endpoint's runtime against the per-class budgets once the model-quality campaign is clean; the regression ledger at `_inputs/eval-findings.md` is the prerequisite evidence the release-readiness sign-off consumes.
+
+## Bindings (§0.j five-direction)
+
+- **Drives →** The consuming suite's model-quality review slot. The eval-harness skill at `skills/eval-harness/SKILL.md` (every campaign drives it for dataset + scorer definition). The prompt-evaluator agent at `agents/prompt-evaluator.md` (every output is scored through it). The regression ledger that the release-readiness sign-off consumes. The remediation change-set that consumes the findings backlog (downstream, governed by `rules/production-ready-prs.md` same-change-set discipline).
+- **Driven by ←** The operator-supplied evaluation dataset, scorer definition, and candidate endpoint. The prior baseline at `_inputs/eval-baseline.md` (the regression comparand).
+- **Satisfies →** The consuming suite's model-quality review catalog. The Testing and Observability axs at `rules/cognitive-identity.md` §1 seven-axs-of-breadth taxonomy. The `commands/README.md` command catalog's Cohort-commands row for `/eval` (the registry entry that ratifies this command's place in the slash-command catalog).
+- **Established by ↑** `rules/cognitive-identity.md` §1 seven-axs-of-breadth taxonomy (the Testing and Observability axs this command operationalizes). `rules/authority-inquiry.md` (the pre-flight inquiry surface for dataset / scorer / endpoint resolution). `rules/pre-emission-gate.md` (the fifteen-bar validation this command runs at Phase 5).
+- **Gated by ←** The operator-ratified dataset, scorer, and candidate endpoint (no provider is assumed). The harness's Agent + structured inquiry + Bash + Read + Write tool surface. The eval-harness skill's dataset-schema validation (a malformed schema blocks Phase 1).
+- **Cross-bound with ↔** `skills/eval-harness/SKILL.md` (the dataset + scorer definition surface this command drives). `agents/prompt-evaluator.md` (the per-output scoring agent this command dispatches at Phase 3). `commands/perf-audit.md` (sibling diagnostic command; `/eval` audits model quality, `/perf-audit` audits runtime). `rules/option-annotation.md` (every scorer-choice inquiry carries the Recommended marker with concrete-driver rationale). `rules/authority-inquiry.md` (every dataset / scorer / endpoint ambiguity routes through the canonical channel). `rules/host-discovery.md` (the candidate endpoint is discovered, never invented). `rules/pre-emission-gate.md` (Phase 5 fifteen-bar validation). `rules/large-file-generation.md` (incremental generation when the report exceeds 500 lines).