Octocode Graph Eval Loop

Octocode Graph Eval

Evaluate outcomes and run improvement loops with evidence, not vibes — for one loop or a graph of loops. Flow: ERROR-ANALYZE → FRAME(goal→KPI) → BASELINE → LOOP → JUDGE → CAPTURE → VERIFY → SUITE-EVOLVE. Modes: ErrorAnalyze · Define · Run · Suite · Benchmark · Audit.

Lobby rules

• No goal→KPI link → STOP. No measurable primary → STOP. No runnable sensor → build one before looping.
• Narrative-only accept → REJECT. Editing harness/cases/graders to pass → REJECT.
• ACCEPT only if primary moves on held-out and guardrails hold.
• Prefer deterministic graders; binary/LLM next; humans calibrate. Grade outcomes over paths.
• TDD for agents: write or select a failing case / KPI check before mutating the subject; green only after the change (red → green → keep|discard).
• Public benches orient; private failure suites gate ships. Distrust saturated/contaminated boards.
• Freeze the harness during an experiment; evolve the suite only between experiments.
• Graph check: before evaluating a multi-agent workflow, run edge detection — if no two nodes are independent (every step reads the prior step's output), it is a loop, not a graph. Build a loop.
• Goodhart guard: every primary KPI must have a counter-metric guardrail the agent cannot tune. Primary improving + guardrail degrading → reframe the goal, not the loop.
• Verifier independence: a verifier sharing the executor's context is not independent. Require fresh context before calling a result verified.
• Anchor requirement: every graph must have at least one node whose output cannot be argued with (tests that ran, build exit codes, type errors). No anchors → build one before trusting the graph.

Workflow

1. Error-analyze traces into a failure taxonomy; frame success, primary/leading metrics, guardrails, and decision rule.
2. Measure a fixed-budget baseline; make the smallest subject change; keep or discard from comparable results.
3. Judge grader quality, fairness, capability versus regression, and contamination; capture one durable lesson.
4. Verify held-out results and required checks; then add new failure cases between experiments. Stop when goal/KPI is undefined, checks did not run, the harness changed to pass, or another loop cannot change the verdict.

Smart routes — load only what the current step needs

• When deriving failures, load references/error-analysis.md; when connecting intent to measures load references/goal-kpi-cascade.md, then fill references/kpi-contract.md — make success and budget explicit.
• When choosing experiment, suite, or meta scope, load references/nested-loops.md; before the first iteration load references/feedback-loops.md, then for the inner keep/discard cycle load references/agent-loop.md — no workable sensor, no loop.
• When the subject is a multi-agent workflow (graph of loops), load references/graph-of-loops.md — run edge detection first, require anchor nodes, check verifier independence, name Goodhart guardrails, then set primary KPI at the graph boundary with per-node sensors.
• When auditing that graph for structural failure risk before trusting its green lights — shared context, opaque state, no checkpoint/resume, unbounded tool permissions, missing human gates — load references/graph-failure-modes.md; add a suite case on a mode's first trace appearance.
• When managing or measuring subagents under eval, load references/subagent-cookbook.md first for the ownership split; spawn mechanics stay in octocode-subagent.
• When running an evaluated multi-agent iteration, load references/subagent-protocol.md for the frozen FRAME→verdict protocol; when choosing worker and graph-boundary metrics, load references/subagent-kpis.md so spawn cost is measured, not invisible.
• When defining how parent and workers talk during an evaluated run, load references/subagent-communication.md — bad channels create false certainty and unattributable failures; when choosing the topology itself, load references/subagent-approaches.md because the pattern decides which KPIs and checks matter.
• When inner loop is flat and no new hypothesis exists, suspect stuck search priors — load references/nested-loops.md for bilevel escalation, then references/karpathy-patterns.md for the Bilevel Autoresearch pattern.
• When selecting graders or statistical checks, load references/eval-techniques.md; when grading agent tool-call sequences or multi-turn trajectories load references/trajectory-grading.md; when trusting public/private suites load references/benchmarking.md — match evidence strength to the decision.
• When creating cases and runners, load references/eval-harness.md; before acceptance load references/held-out-and-guards.md — prevent leakage, overfitting, and greenwashing.
• When grounding methods in primary patterns, load references/karpathy-patterns.md — anchor techniques in proven loops.
• When a result needs another skill or durable capture, load references/routing.md; when closing a meta improvement cycle load references/improve-loop.md — transfer ownership without losing the decision rule.
• When reporting, load references/output.md and run scripts/loop-report.mjs — require goal, baseline, result, and verdict.