Evals Console

Every number shows its denominator, interval and rows. Hover a metric for its definition.

Metrics include flagged synthetic demo history (a four-week sample for illustration, not a measurement). Every card shows its real/synthetic split; the trust level below uses real decisions only. Show real only

Trust level 0: Review everything

Real decisions only

Every draft needs a rep's approval. Every action still needs a rep's approval in this prototype.

To reach level 1

  • Real decisions ≥ 10: 1 (needs ≥ 10)
  • Clean approval rate: not enough data (0/1) (needs ≥ 50%)
  • Critical errors in last 20 = 0: 1 (needs = 0)shield block, wf_b730a5fb

To reach level 2

  • Real decisions ≥ 20: 1 (needs ≥ 20)
  • Clean approval rate: not enough data (0/1) (needs ≥ 70%)
  • Fact correction rate: not enough data (0/0) (needs ≤ 20%)
  • Rubber-stamp rate: not enough data (0/0) (needs ≤ 30%)

To reach level 3

  • Real decisions ≥ 40: 1 (needs ≥ 40)
  • Clean approval, lower 95% bound: 0% (needs ≥ 75%)
  • Completion rate: not enough data (0/0) (needs = 100%)
  • Write fidelity: not enough data (0/0) (needs = 100%)

Rep decisions

Clean approval rate

Show rows

44% (26/59)

95% CI 32%–57%

1 real · 58 synthetic

Fact correction rate

Show rows

9% (4/47)

95% CI 3%–20%

0 real · 47 synthetic

Rubber-stamp rate

Show rows

13% (4/30)

95% CI 5%–30%

0 real · 30 synthetic

Reject reason mix

Show rows

33% (4/12)

95% CI 14%–61%

1 real · 11 synthetic

  • wrong fact4/12 · 33%
  • bad timing4/12 · 33%
  • too risky3/12 · 25%
  • wrong action1/12 · 8%

Completion rate

Show rows

98% (46/47)

95% CI 89%–100%

0 real · 47 synthetic

Critical errors (last 20)

Show rows

1of 2 recent events

2 real · 0 synthetic

Execution and grounding

Write fidelity

Show rows

99% (140/141)

0 real · 141 synthetic

Grounding pass rate

Show rows

72% (52/72)

95% CI 61%–81%

10 real · 62 synthetic

Replay pass rate

Show rows

0% (0/2)

95% CI 0%–66%

0 real · 2 synthetic

  • generate@v10/2 · 0%
  • generate@baseline0/2 · 0%

Unsupported claim rate

Show rows

Not enough data (n=0)

0 real · 0 synthetic

Routing accuracy

Show rows

Not enough data (n=0)

0 real · 0 synthetic

Graph corrections

Show rows

Not enough data (n=0)

0 real · 0 synthetic

Classifier accuracy

Show rows

Not enough data (n=0)

0 real · 0 synthetic

Human burden

Show rows

Not enough data (n=0)

0 real · 0 synthetic

Critical actions auto-run

Show rows

0of 0 recent events

0 real · 0 synthetic

Judge agreement (verify.ts)

Show rows

Not enough data (n=0)

0 real · 0 synthetic

Edit classifier

No classifier results yet. Run the feedback experiment on /evals/experiments.

Replay (2 active tests)

Testgenerate@v1generate@baseline
Follow-up states renewal date as December 31, 2026
fail · diff

Expected a sentence containing: December 31, 2026

The draft said:

  • Great to hear all 140 seats are live.
  • As Luis asked, I'll send over the SOC 2 Type II report and API rate-limit doc today.
  • Next step on our side: Security review call, week of October 12, 2026.
  • I have your renewal down for December 31, 2026.
  • I understand the budget owner for the expansion is not decided yet.

live generation failed; fell back to fixture

fail · diff

Expected a sentence containing: December 31, 2026

The draft said:

  • Great to hear all 140 seats are live.
  • As Luis asked, I'll send over the SOC 2 Type II report and API rate-limit doc today.
  • Next step on our side: Security review call, week of October 12, 2026.
  • I have your renewal down for December 31, 2026.
  • I understand the budget owner for the expansion is not decided yet.

live generation failed; fell back to fixture

Keeps the rep's addition: "week, including Jana Svoboda's maintenance savings"
fail · diff

Expected a sentence containing: week, including Jana Svoboda's maintenance savings

The draft said:

  • Congratulations on the pilot: 18% faster line changeover.
  • Thanks for confirming you're about 450 people after the spin-off; I'll size the proposal on that.
  • I'll put together the one-page ROI summary this week.
  • I'll address it to Tomas Hruby, who you mentioned has final sign-off.

live generation failed; fell back to fixture

fail · diff

Expected a sentence containing: week, including Jana Svoboda's maintenance savings

The draft said:

  • Congratulations on the pilot: 18% faster line changeover.
  • Thanks for confirming you're about 450 people after the spin-off; I'll size the proposal on that.
  • I'll put together the one-page ROI summary this week.
  • I'll address it to Tomas Hruby, who you mentioned has final sign-off.

live generation failed; fell back to fixture

Decision log (8)

WorkflowAccountOutcomeReview timeEdits
wf_03c6c571Brightline Freight rejected: too risky15.9s0
wf_b730a5fbBrightline Freight held by the shield–0
wf_3fcf69a2Walkthrough · Ostrava Robotics SYNTHETICapproved edited31.0s1
wf_23ca674dWalkthrough · Ostrava Robotics SYNTHETICrejected: wrong fact (Final approver)9.0s0
wf_6f300246Walkthrough · Halcyon Clinics SYNTHETICapproved clean18.0s0
wf_beca6a8dWalkthrough · Halcyon Clinics SYNTHETICheld by the shield–0
wf_2d4f8e98Walkthrough · Brightline Freight SYNTHETICapproved clean12.0s0
wf_6383231bWalkthrough · Brightline Freight SYNTHETICapproved edited24.0s1

Edits (2)

BeforeAfterCategorySeverityDecided byFix goes toWorkflow
week.week, including Jana Svoboda's maintenance savings.missing contextmajorrulegraphwf_3fcf69a2
March 31, 2027.December 31, 2026.fact correctioncriticalrulegraphwf_6383231b

Regression tests (2)

TestLatest resultWhyCreated from
Keeps the rep's addition: "week, including Jana Svoboda's maintenance savings"faillive generation failed; fell back to fixtureMaya Reyes's edit on wf_3fcf69a2
Follow-up states renewal date as December 31, 2026faillive generation failed; fell back to fixtureMaya Reyes's edit on wf_6383231b