FLASHRSI / RESULTS MATRIX

Source-task and transfer performance (%)

Rows are methods × target harnesses. Columns separate the source benchmark, evaluation type, and benchmark route so the full study stays readable without repeating one route per row.

positive Δ ppzeronegative Δ ppFinal score shown above the change from the initial harness
MethodTarget
harness
SourceIn-domainNear-domainFrozen transferFar-OODFrozen transfer
SearchQAIn-domainHotpotQAIn-domainTerminal-Bench 2.0In-domainAppWorldIn-domainSearchQANQSearchQATriviaQAHotpotQAMuSiQueTerminal-Bench 2.0SWE-benchAppWorldBFCL Multi-TurnSearchQAOfficeQA ProHotpotQADeepSearchQA
Initial harnessClaude Code73.9357.3263.3364.8856.6768.6767.5076.6765.0044.1962.00
Mini-Agent76.4358.5466.6765.4859.3374.3365.8380.0070.0027.9152.00
SkillOptClaude Code83.14+9.2160.98+3.6670.00+6.6784.52+19.6469.00+12.3382.00+13.3376.67+9.1780.00+3.3368.75+3.7546.51+2.3236.00-26.00
Mini-Agent82.93+6.5059.76+1.2266.670.0091.67+26.1962.00+2.6784.33+10.0074.17+8.3486.67+6.6768.75-1.2537.21+9.3046.00-6.00
OPROClaude Code81.43+7.5062.20+4.8863.330.0066.07+1.1969.67+13.0080.33+11.6673.33+5.8380.00+3.3366.25+1.2551.16+6.9730.00-32.00
Mini-Agent80.71+4.2860.98+2.4466.670.0059.52-5.9664.67+5.3481.33+7.0078.33+12.5080.000.0068.75-1.2541.86+13.9544.00-8.00
AHEClaude Code79.36+5.4360.98+3.6670.00+6.6791.67+26.7966.00+9.3381.67+13.0074.17+6.6780.00+3.3371.25+6.2546.51+2.3246.00-16.00
Mini-Agent78.57+2.1457.32-1.2260.00-6.6771.43+5.9566.33+7.0079.00+4.6773.33+7.5083.33+3.3372.50+2.5034.88+6.9746.00-6.00
ACEClaude Code83.93+10.0063.41+6.0966.67+3.3491.07+26.1964.00+7.3385.67+17.0072.50+5.0073.33-3.3473.75+8.7541.86-2.3336.00-26.00
Mini-Agent84.71+8.2863.41+4.8770.00+3.3391.67+26.1962.67+3.3485.33+11.0075.83+10.0070.00-10.0067.50-2.5030.23+2.3254.00+2.00
EvoPromptClaude Code74.07+0.1458.54+1.2263.330.0061.90-2.9859.67+3.0069.00+0.3370.00+2.5083.33+6.6666.25+1.2544.190.0044.00-18.00
Mini-Agent76.71+0.2859.76+1.2276.67+10.0064.29-1.1958.67-0.6674.00-0.3367.50+1.6786.67+6.6763.75-6.2534.88+6.9740.00-12.00
GEPAClaude Code82.43+8.5062.20+4.8866.67+3.3467.86+2.9865.33+8.6683.67+15.0078.33+10.8376.670.0070.00+5.0041.86-2.3338.00-24.00
Mini-Agent84.64+8.2159.76+1.2270.00+3.3375.00+9.5261.67+2.3486.33+12.0075.83+10.0070.00-10.0062.50-7.5051.16+23.2534.00-18.00
ADASClaude Code81.36+7.4360.98+3.6673.33+10.0089.88+25.0063.67+7.0083.33+14.6677.50+10.0073.33-3.3470.00+5.0051.16+6.9742.00-20.00
Mini-Agent84.86+8.4360.98+2.4466.670.0079.76+14.2862.33+3.0084.33+10.0074.17+8.3473.33-6.6770.000.0032.56+4.6534.00-18.00
DarwinXClaude Code80.50+6.5760.98+3.6663.330.0070.83+5.9565.67+9.0082.00+13.3377.50+10.0066.67-10.0072.50+7.5048.84+4.6536.00-26.00
Mini-Agent79.29+2.8660.98+2.4466.670.0069.05+3.5764.67+5.3483.00+8.6775.83+10.0083.33+3.3367.50-2.5041.86+13.9534.00-18.00
MetaHarnessClaude Code85.07+11.1462.20+4.8870.00+6.6790.48+25.6064.00+7.3386.33+17.6680.00+12.5080.00+3.3366.25+1.2553.49+9.3028.00-34.00
Mini-Agent79.14+2.7159.76+1.2273.33+6.6688.10+22.6265.67+6.3481.00+6.6772.50+6.6776.67-3.3373.75+3.7541.86+13.9538.00-14.00
ModularRSIClaude Code81.07+7.1459.76+2.4460.00-3.3356.55-8.3363.00+6.3384.00+15.3372.50+5.0080.00+3.3365.000.0044.190.0046.00-16.00
Mini-Agent78.64+2.2159.76+1.2270.00+3.3369.05+3.5767.00+7.6781.00+6.6765.830.0073.33-6.6775.00+5.0053.49+25.5840.00-12.00
RRSIClaude Code82.79+8.8660.98+3.6676.67+13.3492.86+27.9864.00+7.3383.67+15.0079.17+11.6780.00+3.3372.50+7.5053.49+9.3052.00-10.00
Mini-Agent81.64+5.2159.76+1.2276.67+10.0084.52+19.0468.67+9.3483.00+8.6776.67+10.8476.67-3.3366.25-3.7553.49+25.5836.00-16.00
Evo-HarnessClaude Code75.36+1.4360.98+3.6666.67+3.3481.55+16.6758.67+2.0071.33+2.6677.50+10.0080.00+3.3365.000.0034.88-9.3160.00-2.00
Mini-Agent80.64+4.2159.76+1.2266.670.0077.38+11.9065.33+6.0080.00+5.6775.00+9.1776.67-3.3372.50+2.5048.84+20.9358.00+6.00

Reading: each method has two rows, one for Claude Code and one for Mini-Agent. Source columns are in-domain; Near-domain and Far-OOD columns show frozen transfer routes. Bold marks the highest final score per benchmark and harness, including ties.