Skip to content
AI Knowledge MapSearch / Text directory

Which intermediate conclusions need verification before becoming inputs to later steps?

D003 · D Organizing complex work · Task decomposition and dependencies

Key answer

Prioritize premises whose errors change later direction, calculations or operation targets. Not every transitional sentence needs a checkpoint; identify what subsequent work depends on.

Full explanation

Key premises include complete extracted requirements, correct units/filters, citation support, target-file identity and successful prior writes. Treating ten-thousand-unit figures as single units invalidates even sophisticated budget analysis; good formatting cannot repair that foundation.

Tyen et al. separated finding reasoning errors from correcting located errors. Across five reasoning tasks, true error locations improved correction, while models struggled to locate errors themselves. Supply checkable information rather than merely asking for more careful review. Study

Pair key premises with verification targets: data to source tables, claims to papers, calculations to formulas/scripts, operations to tool receipts/current state. Anthropic also recommends environmental feedback rather than relying on an agent's own narrative. Official guide

For example:

Before charting, verify source-table units, dates, missing values and filters. Use checked data for charts and analysis. Resolve verifiable issues and continue; ask me only when multiple reasonable interpretations change conclusions and sources cannot decide. Include verification evidence at delivery.

Checks have different scope. Verify every amount when none may be wrong; sample passages to detect a layout pattern. Unverified premises may support clearly labeled trial assumptions, but must not silently become confirmed facts in the final conclusion.

References

1. Tyen et al. (2024), LLMs cannot find reasoning errors, but can correct them given the error location, Findings of ACL; localization and backtracking experiments

2. Anthropic (2024), Building effective agents, Agents: environmental feedback and actual results