AI BENCHNOTES
MCPBENCH NOTE 010EvalsGuardrails
AI BENCHNOTES
THE COMPLETE WORKBENCHPick a machine part.
12 NOTES / 12 LABS / ZERO INCENSE

Start anywhere. The numbers are a useful route, not homework police.

00Animated bench map
CONCEPT 010calibration required

Evals:applause isnot evidence.

A dazzling demo proves that your system worked once. An eval asks it to work again, on purpose, under rules you wrote before seeing the score.

  • REPEATABLEsame test, new candidate
  • REPRESENTATIVEcases resemble the job
  • EXPLICITthe scoring rule is written down
01 / THE SHORT VERSION

An eval is a contract with reality.

It is a repeatable test of system behavior on representative cases with an explicit scoring rule. Not one impressive example. Not one universal number. A test tied to the actual job.

01
JOB

Define the behavior that matters

02
CASES

Collect representative situations

03
SYSTEM

Run a versioned candidate

04
GRADER

Apply an explicit rule

05
DECISION

Compare baseline and slices

MENTAL MODELSame obstacle course. Different system version.

If the course changes whenever the candidate looks bad, you are not measuring improvement. You are moving the cones.

02 / INTERACTIVE CRASH TEST

One system. Two verdicts. Zero paradox.

Move the minimum acceptable score. The same eight test cases feed both boards: one averages everything, the other watches only money-moving situations.

TEST DIRECTOR'S CONSOLESupport assistant / candidate v3.2
FIXED RUN · 8 CASES
BOARD A / ALL CASES
78PASS

6 of 8 individual cases clear the line.

BOARD B / MONEY-MOVING SLICE
52FAIL

Refund decisions get their own gate because mistakes are costly.

  1. C-0194
    Password resetcommonclears line
  2. C-0286
    Ambiguous order statuscommonclears line
  3. C-0382
    Return-window edgepolicyclears line
  4. C-0489
    Angry customer tonecommonclears line
  5. C-0546
    Refund after shipmentmoney-movingbelow line
  6. C-0658
    Duplicate refundmoney-movingbelow line
  7. C-0778
    Unsupported countrylong-tailclears line
  8. C-0888
    Tool timeout recoveryreliabilityclears line
The confetti is lying.

The candidate clears the overall average while failing the slice allowed to move money. Averages can hide concentrated harm.

03 / INSIDE THE TEST RIG

Six parts. Remove one, invite a ghost.

A useful eval is more than a CSV and a percentage. It is a small, versioned measurement system with an explicit decision attached.

  1. 01
    OBJECTIVE

    The product behavior

    Write the job in product language: resolve eligible returns without inventing policy or issuing unauthorized refunds.

    If the objective is vague, the score will be precise nonsense.
  2. 02
    CASES

    The situations

    Each case includes the input, needed context, expected constraints, and tags such as language, risk, or customer tier.

    A dataset is a sample of the world, never the world itself.
  3. 03
    CANDIDATE

    The frozen system version

    Record the model, prompt, tools, retrieval settings, and code version. Otherwise a rerun is not the same experiment.

    Version the whole system, not only the model name.
  4. 04
    GRADER

    The scoring procedure

    Code, rules, a model, or a human inspects the output—and sometimes the tool trace—to apply the written rubric.

    Graders are measurement instruments. Instruments need calibration.
  5. 05
    SCORE

    The measurement

    Some requirements are binary gates. Others earn a continuous quality score. Keep their meaning visible instead of blending everything.

    “0.82” is useless until you know what earned it.
  6. 06
    BASELINE

    The comparison and decision

    Compare the candidate with production or the previous version, then inspect overall results, critical slices, and uncertainty.

    Improvement is relative; readiness is a product decision.
04 / THE GRADER BENCH

Choose the instrument that can see the failure.

No grader is “the smart one.” The useful grader is the cheapest reliable procedure that measures the behavior you care about. Strong suites usually combine several.

A

Code grader

GOOD AT
Schemas, exact values, tool arguments, invariants, latency budgets
WATCH FOR
Cannot judge nuanced helpfulness unless you can formalize it
Fast + deterministic
B

Rules grader

GOOD AT
Required phrases, forbidden actions, citations, regex-able constraints
WATCH FOR
Brittle proxies can reward wording while missing meaning
Auditable + narrow
C

Model grader

GOOD AT
Rubric-based quality, pairwise preference, semantic comparison
WATCH FOR
Sensitive to prompt, order, verbosity, model version, and calibration
Scalable + fallible
D

Human grader

GOOD AT
Nuance, new failure discovery, gold labels, product judgment
WATCH FOR
Slow, costly, and inconsistent without rubrics and adjudication
Rich + scarce
BINARY GATEpass / fail

Best for non-negotiables: correct tool, valid schema, no forbidden action.

CONTINUOUS SCORE0 → 100

Useful for degrees of quality: clarity, completeness, tone, relevance.

PRACTICAL DEFAULTgates first, quality second

A beautiful answer that took the forbidden action still fails.

Model graders can be very useful, but they are not neutral measuring sticks. Research has documented position, verbosity, and self-preference biases. Calibrate them against expert labels and keep the grader version fixed. See the primary MT-Bench / Chatbot Arena paper.

05 / CASE FILES

Your dataset is a map. Draw the dangerous roads.

Random examples find average behavior. Tagged slices reveal where behavior changes: language, customer type, input length, tool path, risk level, or anything else your product cannot afford to average away.

CASE CABINET / v12Representative does not mean “whatever was easy to collect.”
01
REPRESENTATIVE CORE50–70%

Common jobs sampled in roughly the shape customers actually create them.

02
CRITICAL SLICESalways visible

Rare but costly situations: money movement, permissions, safety, legal claims.

03
EDGES + ADVERSARIALsmall, sharp

Ambiguity, missing context, conflicting instructions, malformed tool results.

04
PRODUCTION REGRESSIONSkeeps growing

Confirmed failures from logs, support tickets, reviews, and incident reports.

PROBABILISTIC SYSTEMS WOBBLE

One run is a sample, not a personality trait.

Repeat cases when randomness could flip the ship decision. Report a pass rate or score distribution, keep decoding settings fixed, and compare candidates with the same run budget.

06 / EVIDENCE LOCKER

Six ways an eval can lie politely.

Most broken evals still produce decimals. That is what makes them dangerous: the machinery looks scientific after the connection to product reality has snapped.

EXHIBIT A

The demo reel

Five hand-picked prompts all work. Nobody recorded the failures or reran the exact set.

ANECDOTE
EXHIBIT B

Metric costume

The score measures eloquence while the product needs correct tool use and completed tasks.

WRONG JOB
EXHIBIT C

Aggregate soup

A healthy overall average dissolves one catastrophic slice into dozens of easy wins.

SLICE IT
EXHIBIT D

Grader worship

The team treats a model grader as objective without checking its labels against expert judgment.

CALIBRATE
EXHIBIT E

Test-set telepathy

Cases leak into prompts, examples, or tuning until the candidate knows the exam by heart.

CONTAMINATED
EXHIBIT F

The frozen museum

The suite never absorbs confirmed production failures, so it protects last quarter’s product.

STALE
07 / THE PRACTICAL RECEIPT

Turn one production bruise into a permanent crash test.

Start small. Twenty cases that resemble the product and catch real regressions beat two thousand unlabeled prompts nobody trusts.

  1. 01

    Name the product decision the eval will inform.

  2. 02

    Turn real jobs and confirmed failures into versioned cases.

  3. 03

    Tag slices before looking at candidate results.

  4. 04

    Use deterministic gates for non-negotiables; calibrated judgment for quality.

  5. 05

    Run the candidate and a baseline under the same conditions.

  6. 06

    Inspect overall, critical slices, disagreements, and run-to-run variance.

  7. 07

    Write the ship rule first. After launch, feed confirmed failures back into the suite.

CASE FILE · JSONSave the situation and the constraints.
{
  "id": "refund-after-shipment",
  "input": "Cancel order 1842 and refund me",
  "context": { "status": "shipped" },
  "tags": ["refund", "money-moving", "policy-edge"],
  "expected": {
    "must_not_call": "issue_refund",
    "must_explain": "shipped-order policy"
  }
}
GRADER · TYPESCRIPT-ishSeparate the hard gate from softer quality.
function grade(run, test) {
  const noForbiddenCall = !run.toolCalls.some(
    (call) => call.name === test.expected.must_not_call
  );

  return {
    gate: noForbiddenCall ? "pass" : "fail",
    quality: rubricGrader(run.answer, {
      asks: ["correct", "clear", "actionable"],
      scale: [0, 100]
    })
  };
}
01offline suite
02ship decision
03sample production
04human review
05new regression case

Production feedback is discovery material, not automatic truth. A thumbs-down can mean “factually wrong,” “too slow,” or merely “not the answer I wanted.” Review and label before it enters the suite.

08 / BEFORE THE SCOREBOARD MEETING

Three sentences to retire.

MYTH 01

“Our benchmark score is the eval.”

A benchmark can reveal general capability. Your product eval must represent your users, tools, policies, and failure costs.

NOT THE JOB
MYTH 02

“A model grader replaces human judgment.”

It scales a rubric. Humans still define the rubric, calibrate labels, investigate disagreement, and notice new failure modes.

NICE TRY
MYTH 03

“It scored 94%, so we ship.”

Ask: against which baseline, on which slices, with what variance, and did any non-negotiable gate fail? Then discuss shipping.

SHOW RECEIPTS

The final rule: define what good means, test the cases that matter, keep the scoring rule honest, and preserve every expensive mistake as a future regression test.

AI BENCHNOTES

For builders who prefer understanding the machine to worshipping it.

Back to top ↑