feat(evals): add LLM-as-judge scoring by sudoKrishna · Pull Request #8515 · simstudioai/sim

sudoKrishna · 2026-10-01T14:09:17Z

Summary

Substring checks measure phrasing, not correctness — every live failure so far
was a valid paraphrase or an over-specific assertion. This adds a judge model
that scores an answer against a weighted rubric and returns structured numbers,
so the suite can assert behavior instead of wording.

Stacked on #8409 (the eval harness) — base branch is feat/agent-tool-use-evals.

Closes #8514

What changed

judge.ts — rubric + prompt, judgeAnswer, JSON parsing/clamping, verdict
judge.test.ts — key-free coverage: parsing, fenced JSON, clamping, weights,
missing criteria, non-JSON
judge.live.test.ts — opt-in: a grounded answer outscores an invented one
harness.ts — runScenario accepts an optional judge and adds a judge
check; deterministic checks are unchanged
test:evals:judge script; README documents it

How it works

const verdict = await judgeAnswer({
  completion, model: 'deepseek-chat',
  userMessage, answer,
  evidence: JSON.stringify(toolResults),
  rubric: { criteria: [
    { id: 'grounding', description: 'every claim is supported by the evidence' },
    { id: 'completeness', description: 'answers the user request' },
  ], minScore: 0.7 },
})
// => { scores, rationale, weightedScore, passed }

Because the judge transport is an ordinary OpenAI-compatible completion, it can
be recorded and replayed deterministically with the record/replay layer.

Test plan

judge.test.ts → 7/7 (parsing, weights, clamping, failures)
bun run test:evals → 19/19
bun run check:test-patterns passes
judge.ts and the harness change type-check against the real loop signature
bun run test:evals:judge live run (grounded > invented)
Full bun run type-check — run in CI

Follow-up

Wire the judge into selected scenarios (replace brittle finalContent regexes
where a rubric is more honest) once the live judge run confirms the prompt.

Add a deterministic eval layer for the agent harness. Scenarios script the OpenAI-compatible streaming tool loop with model turns and stub tool results, then score tool selection, planning, retrieval, and recovery without a provider key. - apps/sim/evals/agent-tool-use: 8 scenarios, scoring, JSON+Markdown report - `bun run test:evals` from apps/sim runs the suite and writes the report - picked up by the normal vitest run so a regression fails CI - README documents the contract and how to add a case

Replay the same scenarios against a real model. The model is the only thing that changes: runScenario now takes an optional completion transport and a live mode that relaxes exact assertions (ordered subsequence, minimum successes) and skips scripted-only recovery cases. - live.ts: OpenAI-compatible transport + DeepSeek factory - agent-tool-use.live.test.ts: K trials per scenario, gated on EVAL_LIVE=1 and DEEPSEEK_API_KEY, never runs in CI - live report with pass rates, avg iterations, latency, failed checks - test:evals:live script and README knobs

…ve mode The first live DeepSeek run exposed brittle assertions, not harness bugs: the model chained the tools correctly but the checks were case-sensitive and required an internal order id. Match the retrieved value case-insensitively and let live runs accept the grounded status rather than the internal id.

Add an executor-level harness: a real Start -> Agent workflow on DAGExecutor, with only executeProviderRequest mocked at the provider boundary. This covers agent-block input wiring, variable resolution from Start outputs, and executor run/error handling, which the direct loop harness cannot see. - executor-harness.ts: workflow builder + runExecutorScenario - shares the scorer (scoreExpectations) and report with the loop suite - two scenarios: Start->Agent output, and <start.message> resolution - README documents adding an executor-level scenario

Add executor-retries-failed-block: the first provider call rejects, the Agent block has retry enabled, and the executor replays it. The run must complete with the second response. Verifies providerCalls === 2, and fails without the retry policy (checked locally: expected 2, got 1).

Add executor-falls-back-to-secondary-model: the primary call rejects, the Agent block has a fallback model, and the handler serves the answer from gpt-4o-mini. Asserts providerCalls === 2 and lastRequestModel, and fails without the fallback row (checked locally: got gpt-4o, run errored).

Substring checks measure phrasing, not correctness. judgeAnswer scores an answer against a weighted rubric with a judge model and returns structured scores; runScenario gains an optional judge that adds a judge check. The judge transport is an injectable OpenAI-compatible completion, so a recorded transcript can replay it deterministically. - judge.ts: rubric, prompt, JSON parsing/clamping, verdict - judge.test.ts: parsing/weighting/clamping (key-free) - judge.live.test.ts: grounded answer outscores an invented one (opt-in) - test:evals:judge script; README documents it

vercel · 2026-10-01T14:09:27Z

@sudoKrishna is attempting to deploy a commit to the Sim Team on Vercel.

A member of the Team first needs to authorize it.

greptile-apps · 2026-10-01T14:14:19Z

Confidence Score: 2/5

[Medium risk] Adds evaluation framework for agent tool-use behavior.

The PR is not ready to merge because the judge lacks grounding evidence and the scripted and live harnesses can report misleading results.

Findings

Judge lacks tool results ▶
Scripted turns ignore feedback ▶
Live calls receive fabricated results ▶
Live checks reject paraphrases ▶
Invalid rubrics skew verdicts ▶
Relative imports violate app convention ▶

Summary

The PR adds scripted and opt-in live agent-tool-use evaluations, an executor harness, report generation, and an injectable LLM judge.

The judge lacks tool-result evidence for grounding, while scripted scenarios do not verify result feedback.
Live fixture handling and wording checks can distort pass rates.
Rubric validation and app import conventions need attention.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  S[Scenario] --> H[Tool-loop harness]
  H --> T[Mocked tool results]
  T --> L[Streaming tool loop]
  L --> A[Final answer]
  A --> D[Deterministic checks]
  A --> J[Optional judge]
  H --> J
  D --> R[Eval report]
  J --> R

_{Reviews (1) · Last reviewed commit: "feat(evals): add LLM-as-judge scoring"}

greptile-apps · 2026-10-01T14:14:26Z

+        rubric: options.judge.rubric,
+        userMessage: scenario.userMessage,
+        answer: finalContent,
+        evidence: JSON.stringify(invocations),


Judge lacks tool results When a scenario enables the judge, it receives tool names, arguments, and success flags, but not the returned outputs. If the answer says the rate limit is 100 requests per minute, the judge cannot see the retrieved rate limit and cannot reliably score whether that claim is grounded. Pass the tool results as evidence.

greptile-apps · 2026-10-01T14:14:26Z

+function createScriptedCompletion(scenario: AgentToolUseScenario): OpenAICompatCreateCompletion {
+  let turnIndex = 0
+  return async () => {
+    const turn = scenario.script[turnIndex]
+    turnIndex += 1
+    if (!turn) {
+      throw new Error(
+        `Scenario "${scenario.id}" requested model turn ${turnIndex} but only ${scenario.script.length} are scripted`
+      )
+    }
+    return (async function* () {
+      for (const next of turnToChunks(turn)) yield next
+    })()
+  }
+}


Scripted turns ignore feedback The scripted completion advances by turn number without reading the messages sent to it. If the loop stops feeding back a tool result or error, the script still emits its planned retry or answer, so retrieval and recovery scenarios can pass despite that regression. Check the tool messages at the completion boundary.

greptile-apps · 2026-10-01T14:14:26Z

+  toolsMockFns.mockExecuteTool.mockImplementation(
+    async (toolId: string, params: Record<string, unknown>): Promise<ToolResponse> => {
+      const startedAt = Date.now()
+      const response = resultQueues.get(toolId)?.shift() ?? { success: true, output: {} }


Live calls receive fabricated results Live tool calls consume results queued for the scripted calls, without matching the model’s actual arguments. If the model makes an extra retry, the exhausted queue returns a successful empty result. That fabricated result reaches the model, so its answer and the reported pass rate no longer reflect the intended tool evidence. Match results to live calls or fail when a fixture runs out.

greptile-apps · 2026-10-01T14:14:26Z

+    expect: {
+      toolCallSequence: ['get_weather', 'get_news'],
+      finalContent: /12°C[\s\S]*transit strike ends/i,
+      maxIterations: 3,
+      successfulToolCalls: 2,
+    },
+    /** The two tools are independent; a real model may emit them in either order. */
+    liveExpect: { toolCallSequence: undefined, requiredTools: ['get_weather', 'get_news'] },


Live checks reject paraphrases The live expectations keep this exact weather-and-news answer pattern. A correct answer such as “12 degrees Celsius” with a paraphrased headline fails the check, making the pass rate depend on wording rather than correctness. The weather scenario likewise retains an exact 17°C check; relax or judge these live answers.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

greptile-apps · 2026-10-01T14:14:26Z

+    const weight = criterion.weight ?? 1
+    scores[criterion.id] = score
+    weighted += score * weight
+    totalWeight += weight
+  }
+
+  const weightedScore = totalWeight === 0 ? 0 : weighted / totalWeight
+  return {
+    scores,
+    rationale: typeof parsed.rationale === 'string' ? parsed.rationale : '',
+    weightedScore,
+    passed: weightedScore >= (rubric.minScore ?? 0.5),


Invalid rubrics skew verdicts Weights and minScore are not range-checked. A negative weight can put the weighted score outside the stated 0–1 range, while a negative threshold can pass a zero-scored answer. Validate positive finite weights and a threshold within 0–1 to avoid misleading verdicts.

greptile-apps · 2026-10-01T14:14:26Z

+import { adaptOpenAIChatToolSchema } from '@/providers/tool-schema-adapter'
+import type { ProviderToolConfig, TimeSegment } from '@/providers/types'
+import type { ToolResponse } from '@/tools/types'
+import { type JudgeRubric, judgeAnswer } from './judge'


Relative imports violate app convention This module imports ./judge, although the repository requires absolute @/... imports in apps/sim. The same pattern appears in report.ts and executor-harness.ts. These imports need to use the app alias before merging to satisfy that requirement.

Context Used: CLAUDE.md (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

sudoKrishna added 7 commits September 29, 2026 14:47

sudoKrishna requested a review from a team as a code owner October 1, 2026 14:09

greptile-apps Bot reviewed Oct 1, 2026

View reviewed changes

Navigation Menu

feat(evals): add LLM-as-judge scoring - #8515

feat(evals): add LLM-as-judge scoring#8515
sudoKrishna wants to merge 7 commits into
simstudioai:mainfrom
sudoKrishna:feat/evals-llm-judge

sudoKrishna commented Oct 1, 2026

Uh oh!

vercel Bot commented Oct 1, 2026

Uh oh!

greptile-apps Bot commented Oct 1, 2026 •

edited

Loading

Uh oh!

greptile-apps Bot Oct 1, 2026

Uh oh!

greptile-apps Bot Oct 1, 2026

Uh oh!

greptile-apps Bot Oct 1, 2026

Uh oh!

greptile-apps Bot Oct 1, 2026

Uh oh!

greptile-apps Bot Oct 1, 2026

Uh oh!

greptile-apps Bot Oct 1, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

sudoKrishna commented Oct 1, 2026

Summary

What changed

How it works

Test plan

Follow-up

Uh oh!

vercel Bot commented Oct 1, 2026

Uh oh!

greptile-apps Bot commented Oct 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Confidence Score: 2/5

Findings

Summary

Uh oh!

greptile-apps Bot Oct 1, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot Oct 1, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot Oct 1, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot Oct 1, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot Oct 1, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot Oct 1, 2026

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

greptile-apps Bot commented Oct 1, 2026 •

edited

Loading