Raycaster / evalsBack to AI Agents for M&A Legal Due Diligence

APEX-Agents · Investment Banking

World226_RM_04

1/3Fail

APEX-Agents task World226_RM_04 in AI Agents for M&A Legal Due Diligence. Compare dual-harness agent runs across models — rubric criteria, scores, and public traces.

AI Agents for M&A Legal Due DiligenceInvestment Banking World 226Dual harnessGrader: rubric
task_06a33a12bddc482fbbb01ae1752e1907
Investment Banking World 226
message_in_console
4 models · dual config

Task prompt

What the agent was asked to do

Update the LBO model to include an incentive payment structure of PLTF management post-transaction. Assess the impacts on the 5-year LBO analysis. Management is eligible for these payments each year of the forecast based on 3 levels of performance targets: - Minimum: Meets Currently modeled EBITDA projections - Midpoint: Exceeds EBITDA projections by 10% - Maximum: Exceeds EBITDA projections by 20% The Payout for each level: - Minimum: $2mm - Midpoint: $3mm - Maximum: $5mm Here are some assumptions: - For EBITDA outcomes that surpass one threshold but not the next, management will receive the pro-rata proportion of EBITDA in excess of the threshold, calculated linearly between the two thresholds - Create 2 new cases (in addition to the "base" case currently in the model) where revenue exceeds the base case forecast by 5% and 10% per year, respectively - For the 5% revenue outperformance case, assume capex in this scenario scales faster than revenue and as a % of revenue increases by 100bps above the base case - For the 10% revenue outperformance case, assume capex in this scenario scales faster than revenue and as a % of revenue increases by 100bps above the base case In the final results, round all % values to 1 decimal point. Write back to me with your findings here as a short message.

Published trajectories

Agent runs on this task

Curated dual-harness runs (parsed + original sandbox). Best scored run per model.

ModelHarnessScoreResultLinks
GPT-5.5showcasedual1/3Fail
Gemini 3.1 Produal0/3Fail
GPT-5.4 minidual1/3Fail
GPT-5.4 nanodual0/3Fail

Grading rubric

Criteria and grader verdict (showcase run)

  1. States Base IRR is 16.2%

    Pass

    Evidence: TEXT_RESPONSE says, “Base + incentive case: ... 5-year IRR declines slightly from 16.2% to 16.2%...” Assessment: Criterion “States Base IRR is 16.2%” is met because the response explicitly states 16.2%.

  2. States +5.0% revenue outperformance IRR is 19.2%

    Fail

    Evidence: TEXT_RESPONSE says, “Revenue +5% case: ... 5-year IRR increases to 17.3%...” Assessment: Criterion “States +5.0% revenue outperformance IRR is 19.2%” is not met; the response states 17.3%, not 19.2%.

  3. States +10.0% revenue outperformance IRR is 21.8%

    Fail

    Evidence: TEXT_RESPONSE says, “Revenue +10% case: ... 5-year IRR increases to 18.5%...” Assessment: Criterion “States +10.0% revenue outperformance IRR is 21.8%” is not met; the response states 18.5%, not 21.8%.