eval-outcomes
Grade agent or model output against Outcomes for holdout-safe evals and runtime comparisons.
What this skill does
# /eval-outcomes — Outcomes as a Projection of the Locked Eval Substrate Grade an agent's output through Anthropic **Outcomes** (or any model) without ever forking the bar: the rubric is *compiled from* the locked eval substrate (`~/.agents/evals/SCHEMA.md`), the grader runs holdout-safe, and the score is ingested back as the **one** council verdict record. Outcomes is a *projection*, never an alternate authority. ## Overview Outcomes (Anthropic, May 2026) closes the loop "agent runs → grader scores against a rubric → agent retries." AgentOps already has a locked, council-reviewed eval substrate (Suites/Tasks, deterministic-or-LLM scorers, dev/holdout splits, a global Dolt-synced holdout burn ledger, validity gates 1–9, `judge_content_hash` drift detection). This skill makes Outcomes a **grading transport** for that substrate rather than a second bar: - **`ao eval outcomes compile <input.json>`** — projects a locked Task + criteria into a holdout-safe Outcomes rubric payload. It **refuses** to emit any `target`/`ground_truth` field (Managed Agents are not ZDR). - A grader runs the rubric against the candidate output — cloud Managed Agents in the bushido self-hosted sandbox (Claude path), or the local Inspect AI / llama.cpp Qwen grader (Codex/NTM path). Both grade against the *same* compiled rubric. - **`ao eval outcomes ingest <score.json> --json`** — converts the returned score into one council verdict record (`skills/council/schemas/verdict.json` shape), closing the Outcomes → Knowledge Flywheel loop. It **extends** two existing skills, rewriting neither: - [validate](../validate/SKILL.md) — Outcomes is an additional grading transport for the existing per-criterion rubric; the verdict output stays the council verdict schema. - [ratchet](../ratchet/SKILL.md) — Outcomes-derived deltas feed the same Brownian-Ratchet gate records; no new gate semantics. ## ⚠️ Critical Constraints - **Managed Agents are NOT ZDR.** Never send a holdout `target`, `ground_truth`, or PII into an Outcomes payload. **Why:** any leak ships answers to Anthropic permanently and breaks holdout isolation — the load-bearing eval invariant. `ao eval outcomes compile` enforces deny-by-default (allowlist-only field copy + a re-parse guard + schema `additionalProperties:false`); a CI check verifies it, not a runtime hook. - **SCHEMA.md is the authority; the rubric is a derived artifact.** **Why:** the compiled rubric carries `judge_content_hash`, so a stale cloud rubric self-invalidates exactly like a stale local judge (gate #2 parity). Never hand-edit a cloud rubric. - **Every holdout grade registers a global Dolt burn.** **Why:** routing all holdout grades through `ao eval outcomes ingest` keeps gate #3 (the global burn ledger) unconditional — a cloud/Codex run against holdout cannot silently bypass quota. - **Do not relitigate the locked substrate.** **Why:** this skill is a projection; changes to splits/scorers/gates belong to the `ao-evals-schema-v1` council process, not here. ## Workflow ### Phase 1: Compile a holdout-safe rubric ```bash ao eval outcomes compile task-input.json > rubric.json ``` Projects the locked Task + per-criterion rubric into an Outcomes payload, stripping every holdout field. If a holdout value would leak, compile refuses with a non-zero exit. **Checkpoint:** `rubric.json` carries `judge_content_hash` and contains no `target`/`ground_truth` keys. ### Phase 2: Grade against the rubric - **Claude path:** submit the rubric + candidate output to a Managed Agents Session in the bushido sandbox (`100.109.17.108`); the grader sees criteria + candidate only, never the holdout target. - **Codex/NTM path:** grade locally via the Inspect AI runner (dev split) or dispatch to the bushido llama.cpp Qwen grader over tailnet. Same rubric, byte-identical verdict. **Checkpoint:** the score payload carries the same `judge_content_hash` as the rubric (parity), else discard it as stale. ### Phase 3: Ingest into the one verdict record ```bash ao eval outcomes ingest score.json --json ``` Emits a council verdict record (PASS/WARN/FAIL + per-criterion breakdown) on the existing `eval-verdict-compiler` pipeline. A holdout grade registers a global Dolt burn here. **Checkpoint:** the verdict matches `skills/council/schemas/verdict.json`, identical in shape to a local `ao eval run` verdict. ## Output Specification **Format:** JSON. **Filename:** caller-chosen (`<run>.verdict.json`). **Structure:** one council verdict record — `result` (PASS|WARN|FAIL), `satisfaction_score`, per-criterion `breakdown`, `judge_content_hash`. ## Quality Rubric - [ ] Compiled rubric contains zero `target`/`ground_truth` keys (deny-by-default held). - [ ] Rubric + score share the same `judge_content_hash` (no stale-bar drift). - [ ] Holdout grades registered a global Dolt burn via ingest. - [ ] Ingested verdict matches the council verdict schema — one bar, not a fork. ## Examples ```bash # Project a locked Task into a holdout-safe rubric, grade, then ingest the score ao eval outcomes compile task-input.json > rubric.json # (grade rubric.json against the candidate via cloud or local grader -> score.json) ao eval outcomes ingest score.json --json ``` ## Troubleshooting | Problem | Cause | Solution | |---------|-------|----------| | `compile` exits non-zero "would leak into the rubric payload" | A holdout `target`/`ground_truth` reached the allowlist boundary | Working as intended — never bypass; grade against the dev split or fix the input shape | | Verdict differs from a local `ao eval run` | Stale rubric (`judge_content_hash` mismatch) | Recompile from the current Task; discard the stale score | | Holdout quota unexpectedly consumed | A cloud/Codex holdout grade registered a burn (correct) | Expected — gate #3 is unconditional; use the dev split for iteration | ## See Also - [validate](../validate/SKILL.md) — the per-criterion rubric Outcomes transports - [ratchet](../ratchet/SKILL.md) — gate records Outcomes deltas feed - [council](../council/SKILL.md) — the verdict schema ingest emits - [skill-auditor](../skill-auditor/SKILL.md) — audit this skill before declaring stable
Related in AI Agents
skill-development
IncludedComprehensive meta-skill for creating, managing, validating, auditing, and distributing Claude Code skills and slash commands (unified in v2.1.3+). Provides skill templates, creation workflows, validation patterns, audit checklists, naming conventions, YAML frontmatter guidance, progressive disclosure examples, and best practices lookup. Use when creating new skills, validating existing skills, auditing skill quality, understanding skill architecture, needing skill templates, learning about YAML frontmatter requirements, progressive disclosure patterns, tool restrictions (allowed-tools), skill composition, skill naming conventions, troubleshooting skill activation issues, creating custom slash commands, configuring command frontmatter, using command arguments ($ARGUMENTS, $1, $2), bash execution in commands, file references in commands, command namespacing, plugin commands, MCP slash commands, Skill tool configuration, or deciding between skills vs slash commands. Delegates to docs-management skill for official documentation.
reprompter
IncludedTransform messy prompts into well-structured, effective prompts — single or multi-agent. Use when: "reprompt", "reprompt this", "clean up this prompt", "structure my prompt", rough text needing XML tags and best practices, "reprompter teams", "repromptception", "run with quality", "smart run", "smart agents", multi-agent tasks, audits, parallel work, anything going to agent teams. Don't use when: simple Q&A, pure chat, immediate execution-only tasks. See "Don't Use When" section for details. Outputs: Structured XML/Markdown prompt, quality score (before/after), optional team brief + per-agent sub-prompts, agent team output files. Success criteria: Single mode quality score ≥ 7/10; Repromptception per-agent prompt quality score 8+/10; all required sections present, actionable and specific.
adaptive-compaction
IncludedAdaptive add-on policy and recovery layer that decides WHEN to compact, prune, snapshot, or fork -- replacing fixed-percent auto-compaction across Claude Code, Codex, and MCP-capable hosts. Trigger on auto-compact timing or damage: "when should I compact", "is it safe to compact now or start a fresh session", "auto-compact fires too early/mid-task", "switching to an unrelated task but the window still has space", "context rot", "answers get worse the longer the session runs", "the agent forgot the plan or my decisions after it summarized", "add a layer on top that manages context without changing the agent", raising autoCompactWindow to give the policy room, or installing/tuning a cross-tool compaction policy or PreCompact hook -- even when "compaction" is never said but the problem is context-window pressure or post-summarization memory loss. Do NOT use to summarize a conversation, build RAG, write a summarization prompt (decides WHEN not HOW), or answer max-context-length trivia.
agent-skill-creator
IncludedCreate cross-platform agent skills from workflow descriptions. Activates when users ask to create an agent, automate a repetitive workflow, create a custom skill, or need advanced agent creation. Triggers on phrases like create agent for, automate workflow, create skill for, every day I have to, daily I need to, turn process into agent, need to automate, create a cross-platform skill, validate this skill, export this skill, migrate this skill. Supports single skills, multi-agent suites, transcript processing, template-based creation, interactive configuration, cross-platform export, and spec validation.
llm-wiki
IncludedUse when building or maintaining a persistent personal knowledge base (second brain) in Obsidian where an LLM incrementally ingests sources, updates entity/concept pages, maintains cross-references, and keeps a synthesis current. Triggers include "second brain", "Obsidian wiki", "personal knowledge management", "ingest this paper/article/book", "build a research wiki", "compound knowledge", "Memex", or whenever the user wants knowledge to accumulate across sessions instead of being re-derived by RAG on every query.
skill-master
IncludedAgent Skills authoring, evaluation, and optimization. Create, edit, validate, benchmark, and improve skills following the agentskills.io specification. Use when designing SKILL.md files, structuring skill folders (references, scripts, assets), ingesting external documentation into skills, running trigger evals, benchmarking skill quality, optimizing descriptions, or performing blind A/B comparisons. Keywords: agentskills.io, SKILL.md, skill authoring, eval, benchmark, trigger optimization.