Research v1.1 scoring · 120 attempted · 119 completed · Synthetic diagnostic

Open any case. Follow every score.

Inspect submitted facts, eligibility decisions, original comp grids and prospect lists, source IDs and costs.

Refinance Leads Alder

Complete input packet · Public answer key

Lev Agent

Verified / targetEligible / returnedScreeningInference cost
6/66/620/20$0.44 estimate
Read the findings and original evidence

Original structured answer · Every scored check · Full run record

No material selected-field errors or ineligible selections under the v1.1 scoring rules.

All twenty candidate eligibility decisions were correct.

Original deliverables

refinance_leads_alder_charlotte.csv

Acceptance checkResult
Readable Required FormatPass
Identity And OrderPass
Selected Facts And UnknownsFail
Source IdentifiersPass
  • R1514 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R4598 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R7567 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R8446 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R8446 / contact_email: artifact differs from answer or lacks explicit unknown
  • R2663 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R3927 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown

Inspect file-consistency checks. These assess consistency with the answer; they do not replace the truth score.

Observed native tools: get_generated_file. Internal file execution is not visible; data-access equivalence is unverified. Models: claude-opus-4-7. Agent version: 7.9. Sanitized cost and access audit.

GPT-5 + tools

Verified / targetEligible / returnedScreeningInference cost
6/66/620/20$0.53
Read the findings and original evidence

Original structured answer · Every scored check · Full run record

No material selected-field errors or ineligible selections under the v1.1 scoring rules.

All twenty candidate eligibility decisions were correct.

Original deliverables

prospects-v1.csv
prospects-v2.csv
prospects-v3.csv
prospects-v4.csv
prospects-v5.csv
prospects-v6.csv
prospects-v7.csv

Acceptance checkResult
Readable Required FormatPass
Identity And OrderPass
Selected Facts And UnknownsFail
Source IdentifiersPass
  • 7 blank unknown annotations were caused by the benchmark CSV writer after explicit model-supplied JSON nulls; these are harness delivery failures, not model factual errors.
  • R6774 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R4598 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R7567 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R8446 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R8446 / contact_email: artifact differs from answer or lacks explicit unknown
  • R2663 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R3927 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown

Inspect file-consistency checks. These assess consistency with the answer; they do not replace the truth score.

Opus 5 + tools

Verified / targetEligible / returnedScreeningInference cost
6/66/620/20$0.30
Read the findings and original evidence

Original structured answer · Every scored check · Full run record

No material selected-field errors or ineligible selections under the v1.1 scoring rules.

All twenty candidate eligibility decisions were correct.

Original deliverables

prospects-v1.csv

Acceptance checkResult
Readable Required FormatPass
Identity And OrderPass
Selected Facts And UnknownsFail
Source IdentifiersPass
  • 7 blank unknown annotations were caused by the benchmark CSV writer after explicit model-supplied JSON nulls; these are harness delivery failures, not model factual errors.
  • R1514 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R4598 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R7567 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R8446 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R8446 / contact_email: artifact differs from answer or lacks explicit unknown
  • R2663 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown
  • R3927 / outstanding_balance_usd: artifact differs from answer or lacks explicit unknown

Inspect file-consistency checks. These assess consistency with the answer; they do not replace the truth score.

GPT-5 direct

Verified / targetEligible / returnedScreeningInference cost
5/65/618/20$0.09
Read the findings and original evidence

Original structured answer · Every scored check · Full run record

RecordFindingDetail
R6607ineligible selectionSource published after as-of date
RecordReturned eligibilityExpectedReason
R6607truefalseSource published after as-of date
R4598falsetrueMeets the supplied rules

Original deliverables

Direct API control: file creation not requested or scored.

Opus 5 direct

Verified / targetEligible / returnedScreeningInference cost
6/66/620/20$0.21
Read the findings and original evidence

Original structured answer · Every scored check · Full run record

No material selected-field errors or ineligible selections under the v1.1 scoring rules.

All twenty candidate eligibility decisions were correct.

Original deliverables

Direct API control: file creation not requested or scored.