Contextual tuning

These settings live under [tool.mcmr.contextual]. Defaults favor bounded routine checks.

Setting	Default	Effect
reasoning_effort	medium	Provider reasoning level
timeout_seconds	180	Deadline for one model operation
minimum_confidence	0.6	Answers below this become uncertain
batch_size	32	Candidate slice for process and fallback batches
candidate_budget	512	Maximum answer rows in one OpenRouter pack
prompt_token_budget	128000	Maximum estimated input tokens per pack
max_output_tokens	32000	Provider output cap and planning reserve

Pack related rules once

OpenRouter groups rules that depend on shared repository evidence. When all candidates and input tokens fit, MCMR sends one schema-constrained request instead of repeating the same context for each rule. Larger repositories split at the candidate or prompt budget.

Keep prompt_token_budget below the model context window after reserving output. Increase candidate_budget only when the answer schema and expected failure details still fit the output cap. Passing judgments omit detailed prose, which keeps grouped responses small.

Choose output headroom

Reasoning models may count hidden reasoning against output capacity. MCMR reserves part of max_output_tokens according to reasoning_effort, then sizes each pack from the remaining answer budget. A higher effort can therefore produce smaller packs even when the input limit is unchanged.

Tune safely

Start with the defaults and inspect one run. Raise timeout_seconds for slow providers. Raise the prompt budget only for a model whose context window is known. Lower minimum_confidence only when reviewed labels show that lower-confidence answers are useful.

[tool.mcmr.contextual]
timeout_seconds = 600
candidate_budget = 768
prompt_token_budget = 640000
max_output_tokens = 384000

Token usage, cached input, reasoning effort, and model identity remain attached to findings and DataHub run records. See Cost provenance.