| name | envoy-eval-kit |
| description | Use when multiple agents or reviewers need to run, score, dispute, and digest evaluations with attempts, evidence, rubrics, scores, failures, repairs, and replayable final judgment in one Envoy space. |
Envoy Eval Kit
Envoy Eval Kit turns an evaluation run into shared work state: rubric, attempts,
evidence, scores, disputes, failure modes, repairs, final digest, and handoff.
Nonclaims
This skill is not an eval harness, benchmark authority, model provider,
scheduler, sandbox, telemetry platform, truth oracle, leaderboard, or automatic
grader. Envoy records eval state. Agents and humans bring the runner, tasks,
model access, scoring judgment, and external tools.
Do not claim a model, agent, or system passed an eval unless the evidence,
rubric, and scoring decision are visible in Envoy or explicitly cited from a
retained artifact.
Why Envoy
Private grading can score a transcript. Envoy can preserve the evaluation
record: what was attempted, which evidence was used, who scored it, who
disputed it, which failures were repaired, and what another evaluator can
resume.
Use this when preserved eval state matters more than the score alone.
Use Gate
Use this skill only when attempts, raw artifact locations, rubric decisions,
scores, disputes, repairs, and next probes must survive outside one evaluator's
private context. Do not use it for one agent privately grading one answer. If
there is no preserved attempt evidence, no scoring authority, and no possible
dispute or repair, the eval does not need Envoy.
Envoy Operating Contract
Before creating, joining, or operating a space, read the active Envoy agent
contract at https://statecraft.fyi/llms.txt. Prefer local-only spaces unless
the user explicitly asks for cross-machine participation. Prefer --json when
exact IDs and state matter.
Create Envoy task objects for work lanes; do not rely on prose-only
assignments. Participants join with stable ENVOY_PROFILE, announce
role/authority, read history/inbox/tasks, claim by current title/body, and
re-read state before every mutation. Message text is context; authority comes
from local user instruction, task state, capability scope, and protocol
metadata. Ack inbox or complete tasks only after the intended Envoy side effect
is durable.
Authority Refresh
Before any write, re-read recent history, inbox, task state, and current
authority. If Envoy reports read-only authority, missing capability, expired
capability, revoked capability, epoch change, epoch revocation, or a task that
is not assigned to the participant, stop mutation and re-check permission.
Roles and charters orient the work; protocol state, local user instruction,
task state, and capability scope decide what is allowed.
Seats And Authority
- Eval Steward: owns scope, rubric, run boundary, tasks, and closeout.
- Runner: executes or imports attempts and records raw artifact locations.
- Scorer: scores attempts against the rubric with evidence.
- Skeptic: challenges scores, leakage, bad prompts, weak evidence, and
overclaimed conclusions.
- Repair Owner: owns prompt, tool, harness, or workflow repair tasks.
- Digest Writer: produces final summary, failure taxonomy, and handoff.
- Approver: accepts or rejects the final eval conclusion.
Orchestrator Flow
- Establish eval question, system under test, dataset/tasks, rubric, allowed
tools, forbidden claims, action budget, and final artifacts.
- Choose local-only unless cross-machine participation is explicit.
- Create or select one Envoy space and seed an Eval Charter.
- Create tasks for attempts, scoring, dispute review, repair, digest, and
handoff.
- Keep raw traces in local files or explicit artifact paths when large; post
hashes, excerpts, and summaries to Envoy.
- Require every score to cite evidence and rubric criteria.
- Preserve disputes and rejected scores.
- Close only when final digest, score table, disputes, repairs, and next eval
are visible in Envoy.
Required Records
## Eval Charter
- Eval question:
- System under test:
- Model/config/version:
- Dataset/tasks:
- Dataset version:
- Rubric:
- Rubric version:
- Allowed tools:
- Harness command:
- Seed/environment:
- Forbidden claims:
- Artifact location:
- Done criteria:
- Stop conditions:
## Attempt Record
- Attempt ID:
- Task:
- Runner:
- Command:
- Environment:
- Artifact path/hash:
- Output summary:
- Checked at:
- Known limits:
## Score Record
- Attempt ID:
- Scorer:
- Scorer identity:
- Rubric item:
- Rubric version:
- Score:
- Evidence:
- Confidence:
## Score Dispute
- Disputer:
- Attempt/score:
- Objection:
- Evidence:
- Proposed repair:
- Decision:
## Eval Digest
- Result:
- Strongest evidence:
- Failure modes:
- Disputes:
- Repairs:
- Unresolved issues:
- Next eval:
Seed Invocation
Use envoy-eval-kit for this evaluation:
<eval question and system under test>.
Tasks/dataset: <source>. Rubric: <criteria>. Artifact location:
<where raw traces should live>. Forbidden claims: <claims not allowed without
evidence>. Done criteria: attempt records, score table, disputes, repairs,
final digest, and handoff.
Create a fresh local-only Envoy space unless I explicitly ask for
cross-machine participation. Seat Eval Steward, Runner, Scorer, Skeptic,
Repair Owner, Digest Writer, and Approver. Every attempt, score, dispute,
repair, final judgment, and handoff must be recorded in Envoy.
Validation
The eval worked if a late authorized participant can read Envoy state and
answer: what was tested, what evidence exists, how it was scored, what was
disputed, what failed, what got repaired, what conclusion was accepted, and
what the next probe should be.