Evaluate asset

Asset Eval — Orchestrator

<!-- BUDGET_EXCEPTION: Eval skills carry artifact-schema-as-contract (8 body sections + 8-col results row + cross-stack consumer contract) that is load-bearing and cannot move to references/. Asset-eval also surfaces the return-leg gate (score the re-ingested render, never the prompt) + brief-fidelity-vs-render-quality discrimination, which are the point of a render eval. Cycle ledger discipline requires the schema be visible in the SKILL.md body. ~800 tokens over the standard cap is the legitimate cost (matches the evaluate-content / evaluate-ad siblings). -->

Evaluation skill. Converts a re-ingested rendered asset into a cycle snapshot + ledger row + narrowly-scoped next action inside an existing eval loop. One asset (or one picked variant) per cycle; the loop closes only when the real render — not the prompt — is scored (CLOSED-LOOP.md §6 return-leg).

Core Question: "Did the rendered asset, scored against its brief's acceptance criteria and brand tokens, realize the brief well enough to keep / discard / watch / block — and what should the next brief-graphic / produce-asset cycle know?"

Why, methodology, history: references/playbook.md [PLAYBOOK]. Capability metadata (route triggers, prerequisites, load map): routing.yaml.

Critical Gates

1. Existing eval loop required. program.md + context.md absent → NEEDS_CONTEXT, recommend /run-pipeline. This skill does not create loops.
2. Re-ingested asset required (the return-leg). Score the real render, never the prompt. The asset must be attached to its manifest via the return-leg (asset_picked / assets, per references/_shared/execution-fork.md + CLOSED-LOOP.md §6). Only a prompt/brief present → NEEDS_CONTEXT, route to produce-asset + re-ingest (forsvn-preview attach).
3. Source brief required. The brief-graphic / produce-asset artifact carrying the acceptance criteria to score against. Absent or unreadable → BLOCKED.
4. Static-visual lane only. Image / graphic / carousel frame. Video → evaluate-shortform; landing page → evaluate-landing-page; the asset's live-post engagement → evaluate-content / evaluate-ad. Otherwise → NEEDS_CONTEXT, route to sibling.
5. One asset (or picked variant) per cycle. When the option-picker returned a variant set, score the picked variant (asset_picked); the rest archive. A 6-variant render = one cycle on the chosen variant, not six blended.
6. No fabricated quality claims. Judge only what is present in the attached render. Hallucinated visual detail, invented dimensions, or scoring an un-ingested asset are fabrication.
7. Attribution confidence must be explicit. Every verdict states the render engine + execution_mode, render settings, known-engine failure modes (text mangling, hand artifacts, aspect-ratio drift) as confounders, and confidence: high | medium | low | blocked.
8. Evaluation does not produce assets. Recommend changes; route re-rendering to produce-asset, brief/spec changes to brief-graphic, brand-token questions to create-brand.

Responsibility Split

/run-pipeline owns loop setup + program.md / context.md / results.tsv schema + durable learnings. This skill owns post-render brief-fidelity + brand-fit snapshots scored against one re-ingested asset. /produce-asset owns rendering. /brief-graphic owns the brief. /evaluate-content + /evaluate-ad own the asset's live-post performance lanes.

This skill is the formal return-leg gate (execution-fork.md → "Closing the loop"): a brand-critical render is not "done" until scored here against its brief and its realized surface. Inside an eval loop this is the required closing step; without a loop, the lighter fallback is an explicit squint-test against the cited realized surface (realized-surface-grounding.md). Either way, off-brief output is not accepted or committed — score the real re-ingested render, never the prompt.

Inputs

Required: loop slug/path · re-ingested asset path/id (the asset_picked / assets attachment on docs/forsvn/artifacts/marketing/produced-assets/[slug]/manifest.md — image / graphic / carousel frame) · source brief artifact (docs/forsvn/artifacts/marketing/design-briefs/[slug].md) · render engine + execution_mode (brief-only / assisted / direct) · primary metric (brief-fidelity score · acceptance-criteria pass rate · or a downstream metric if the asset is live).

Recommended: brand tokens (brand/BRAND.md + brand/DESIGN.md — palette, type, logo safe-zone, Leaf <10%) · the brief's explicit acceptance criteria (composition, required copy slots, aspect ratio/dimensions, art direction) · variant set with the picked variant flagged · baseline/prior-cycle render · downstream usage context (where the asset will ship).

Outputs

.forsvn/loops/[slug]/evals/YYYY-MM-DD-cycle-N.md + append one row to results.tsv via bun scripts/append-loop-result.ts (8-col helper) + update learnings.md ONLY for high-confidence keep/discard lessons (critic-gated) + run bun scripts/manifest-sync.ts.

Agent Manifest + Dispatch

4 sub-agents: Layer 1 parallel (Metric Ingest + Diagnosis) → Layer 2 (Recommendation) → Layer 3 (Critic). Critic FAIL → revise once; still FAIL → no ledger row + BLOCKED. Full agent table + per-layer dispatch + 7-dim rubric: references/agent-manifest.md. Domain rubric: references/rubric.md. Shared frame: references/_shared/evaluation-loop-rubric.md.

Pre-Dispatch

Canonical: references/_shared/pre-dispatch-protocol.md + references/_shared/eval-loop-spec.md. Hard-blocks (BEFORE Cold Start): missing program.md/context.md → NEEDS_CONTEXT + /run-pipeline; asset not re-ingested (only a prompt) → NEEDS_CONTEXT + /produce-asset + re-ingest; asset is video / landing page → route to evaluate-shortform / evaluate-landing-page; no source brief OR missing asset id → BLOCKED; custom 10+ col results.tsv → warn + hand-edit. Cold Start: 6 bundled questions (loop · asset path/id · source brief path · render engine + execution_mode · primary metric value/baseline · brief acceptance criteria + brand tokens). Full read-order + templates + --fast behavior: references/procedures/pre-dispatch.md [PROCEDURE]. Session execution profile (single-vs-multi): inherit per references/_shared/execution-policy.md.

Artifact Contract

• Path: .forsvn/loops/[slug]/evals/YYYY-MM-DD-cycle-N.md. Lifecycle: evaluation.
• Frontmatter (10 fields): skill / version / date / status / summary / purpose / lifecycle / use_when / do_not_use_when / upstream / downstream + provenance (input_artifacts = source brief-graphic artifact + the re-ingested asset + brand/BRAND.md; output_eval: null).
• Body sections (8): Title · Verdict · Evidence (6-col table) · What Changed This Cycle · Diagnosis (Brief-Fidelity Check + Render-Quality Signals + Brand-Fit Signals + Confounders) · Next Cycle Recommendation · Results Row (8-col TSV) · Learning Promotion. Verdict block must name the asset id + render engine explicitly (Gate 7); Evidence table scopes to that one asset/variant.
• Results Row (8 cols): cycle date artifact primary_metric value baseline status description. status ∈ keep|discard|watch|blocked (Critic Hard Fail otherwise); description includes the asset id + engine tag.
• Cross-stack contract: consumed by future asset-eval cycles (fidelity trend) + brief-graphic --rev=N+1 / produce-asset (re-render seeding) + downstream write-social / write-ad (which reference the kept asset) + human reviewers. Schema changes require atomic update across references/_shared/eval-loop-spec.md + downstream callers.

Full template + Evidence/Brief-Fidelity/Results/Learning formats + helper invocation: references/format-conventions.md [PROCEDURE].

Results Row Helper

bun scripts/append-loop-result.ts "<loop slug>" \
  --artifact evals/YYYY-MM-DD-cycle-N.md \
  --metric "<primary metric>" --value "<current>" --baseline "<baseline>" \
  --status "<keep|discard|watch|blocked>" --description "<one sentence — include asset id + engine>"

Do not append on Critic FAIL — return BLOCKED instead.

Critic Override Protocol

Operator ships despite critic FAIL (or accepts pass-with-concerns) — log BEFORE writing artifact or ledger row: bun scripts/log-critic-override.ts --skill evaluate-asset …. Three overrides → rubric-revision escalation. An override never promotes a contested cycle to keep; a no-override FAIL still returns BLOCKED. Full protocol: references/_shared/critic-override-protocol.md [PROCEDURE].

Anti-Patterns

references/anti-patterns.md [ANTI-PATTERN] — asset-eval rows + 4 cross-cutting marketing-stack rows. Most common: scoring the prompt instead of the re-ingested render (Gate 2 + Critic Hard Fail), a pretty render that misses a hard brief constraint read as keep (Gate 5 + Critic "Brief-Fidelity Discrimination"), off-brand palette / Leaf overrun ignored (Critic "Render-Quality & Brand-Fit"), scope drift to produce-asset re-rendering (Gate 8 + Critic "Decision Discipline").

Worked Example

End-to-end cycle (return-leg scored, hard-miss → discard, Critic PASS): references/examples/asset-eval-cycle-walkthrough.md.

Durable Rules (protected)

<!-- SLOW_UPDATE_START --> <!-- No pinned rules yet. Populate via the slow-update workflow (see references/slow-update-fence.md). Each pinned rule must (a) be procedural not instance-specific, (b) be earned from a regression or critic-flagged failure, (c) cite the artifact / decision record that justified pinning. --> <!-- SLOW_UPDATE_END -->

Completion Status

• DONE — eval artifact written, ledger row appended, critic PASS.
• DONE_WITH_CONCERNS — artifact + row written, but confidence low/medium or confounders material.
• NEEDS_CONTEXT — missing loop, un-ingested asset, source brief, or asset id; OR the asset is video / a landing page / a live-post-engagement question (route to right sibling).
• BLOCKED — contradictory data, no re-ingested asset, filesystem failure, or critic failed after revision.

References

• references/{playbook, agent-manifest, rubric, format-conventions, anti-patterns}.md + procedures/{pre-dispatch, dispatch-mechanics}.md
• references/_shared/{eval-loop-spec, evaluation-loop-rubric, pre-dispatch-protocol, critic-override-protocol, quality-dashboard-spec, execution-fork}.md
• Siblings: produce-asset + brief-graphic (upstream/downstream), run-pipeline (loop scaffolding), evaluate-{content, ad, shortform, landing-page} (sibling lanes)