A bounded agent for paleomicrobiome data processing passes 23 of 24 frozen runs, and every run can be replayed
Synopsis
Ancient microbial data are scarce and contamination-prone, and a language model left to plan an analysis may cite files that do not exist, apply operations a format forbids, or answer questions the data cannot settle. PaleoRigor confines the model's planner to predefined local modules, validates file references, formats and step dependencies before anything runs, refuses out-of-scope requests with a named reason code, and saves the plan, intermediate files, checksums and decisions. In a preregistered frozen held-out evaluation it passed 23 of 24 runs (95.8%), including all 12 boundary decisions; the same model without the planning contract passed 19.
Interpretation
The first step of paleomicrobiome analysis, choosing the right files, checking local copies, selecting operations and recording changes, can now be drafted by a constrained language-model agent as an explicit workflow that a paleobiologist reviews before execution. Generic workflow engines require users to specify inputs and operations themselves, while a language-model agent can turn a request into those choices directly but may reference nonexistent files, apply unsupported operations, or assert conclusions the data cannot support. In the frozen held-out evaluation, PaleoRigor passed 23 of 24 runs (95.8%; Wilson 95% confidence interval 79.8–99.3%), including 11 of 12 supported runs and 12 of 12 boundary decisions; an ablated arm using the same model without the planning contract passed 19 of 24 (79.2%).
The constraint layer catches errors before the plan runs: a validator checks every file reference, format and step dependency, the executor calls only predefined local modules, and the planning contract directs the agent to decline out-of-scope requests with a named reason code instead of answering them. Without that layer the plan is generated rather than authored and errors are hard to trace afterwards, and generic workflow engines do not record why each step was chosen. In the frozen evaluation the single supported-run failure was a planner that omitted the .tsv suffix from an uploaded-file reference, which the validator rejected before execution; boundary decisions were 12 of 12 for PaleoRigor and 10 of 12 for the ablated arm.
Each run exports its plan, intermediate files, checksums and decisions, so results can be recomputed from the record without calling the model again. A run that merely completes shows only that the software did something, not which public record it acted on or what changed. All 96 run labels across two model configurations were independently recomputed from retained records and matched; six public sequencing records matched ENA/SRA read counts and compressed-file sizes (Δreads = 0, Rbytes = 1), and a peptide table went from 5,810 rows to 5,696 unique records after removing 114 duplicates.
Perspective
The evaluation covers workflow execution, recorded data transformations and agreement with public repository metadata; ancient-DNA authentication and biological interpretation are outside its scope and remain with the paleobiologist. All reported results were produced on macOS, and the Windows build was verified only by automated build and tool-identity checks.
The comparison is an ablation of PaleoRigor's own planning contract on the same model, with 24 runs per arm, and neither paired comparison reached statistical significance (McNemar p = 0.219 and 0.0625), so it supports no statement about unfamiliar task types, other providers, or performance relative to established workflow practice. Metadata-level agreement cannot exclude files with identical read counts and compressed lengths but different content, and public sequencing files were not retained, so checksums were not recomputed. Whether the packaged applications improve usability for paleobiologists remains a question for independent user testing.