forced per-field routing audit

Rich evidence does not make routing reliable

The LLM router receives deploy-visible evidence cards and chooses among five correction actions. The full 30-setting grid is complete, but oracle-action agreement remains low, motivating abstention rather than forced action selection.

Router summary

Gemma4 five-action grid, 30 settings

The five-action router is evaluated as a diagnosis of forced action selection. Low oracle-action agreement and action collapse support the shift to abstention.

Tool-audit router

v3 rich evidence grid across crop, protocol, and sample ratio

The tool-audit grid asks the model to override a tool default only when evidence justifies it. Low override is desirable only if the default is already appropriate; here it mostly confirms that the router is cautious rather than reliably corrective.

Crop-protocol summary

median values across five sample ratios
CropProtocolSettingsCasesOracle match medianAction collapse medianLLM vs q_agro median

Action distribution by setting

selected action shares in the 30-setting grid
CropProtocolSRBase onlyCal fullq_agro fullOracle matchCollapseLLM vs q_agro