# Checks for agent work

Part of Deterministic skills by Baltor.

Give each unit of agent work one typed outcome, count the independent evidence behind a claim, and compare candidates only on the rows they all completed.

## What to watch for

- Two reports that quote one source are one observation, not two.
- A candidate that skipped the hard rows can look better on its average.
- A claim of done needs evidence attached, or it is only a status.

## Components

- [Typed work result with six outcomes](https://baltor.ai/components/typed-work-result-with-six-outcomes-8b284073): Give every bounded unit of agent work exactly one typed outcome: completed with evidence, no change needed, needs input, needs scope expansion, unsupported, or budget exhausted with a checkpoint. Name `work_result_type`, licence MIT, 11 files.
- [Provenance-aware evidence counting](https://baltor.ai/components/provenance-aware-evidence-counting-c1a2c7fc): Count the independent observations behind a claim by following each report back to its root observations, so relayed, echoed and merged reports are not counted twice, and simulate non-backtracking relaying between agents. Name `provenance_evidence_counting`, licence MIT, 10 files.
- [Equal-cohort comparison with separate coverage](https://baltor.ai/components/equal-cohort-comparison-with-separate-coverage-6964167f): Compare candidates only on the rows every candidate scored, and report separately how much of the declared universe each one covered. Name `equal_cohort_comparison`, licence MIT, 10 files.
- [Method-blind relabeling and behavioral equivalence](https://baltor.ai/components/method-blind-relabeling-and-behavioral-equivalence-cdb6fd11): Replace candidate names with seeded neutral labels kept in a separate key, and group candidates whose outputs on the same probe inputs are identical, with optional rounding. Name `behavioral_equivalence_relabeling`, licence MIT, 10 files.
- [Mechanism-held-out splits for synthetic task suites](https://baltor.ai/components/mechanism-held-out-splits-for-synthetic-task-suites-437c9550): Split a synthetic task suite by generating mechanism so that no template, rule family or simulator appears on both sides, build balanced mechanism-level folds, and detect leakage in any given split. Name `mechanism_heldout_split`, licence MIT, 10 files.
- [Silent concept-shift twin datasets](https://baltor.ai/components/silent-concept-shift-twin-datasets-7daecd2f): Generate two datasets whose inputs and targets have exactly the same marginal distributions while the input-target relationship is reversed, and compare them with input-only, target-only and joint statistics. Name `silent_concept_shift_twins`, licence MIT, 10 files.
- [Validate an append-only ledger linking each artifact version to what changed](https://baltor.ai/components/baltor-evidence-learning-ledger): Validate an append-only ledger linking each artifact version to what changed, why, what failed, what improved and the lesson learned, refusing any entry that removes, edits or replaces an original instead of sitting beside it. Name `baltor_evidence_learning_ledger`, licence MIT, 10 files.
- [Compare consecutive versions of one artifact with the defects found](https://baltor.ai/components/baltor-iterative-improvement-record): Compare consecutive versions of one artifact with the defects found, the fix or reason for each and a severity-ordered before-and-after verdict, refusing a version that drops a recorded defect without a fix or a reason or claims a better verdict than the counts show. Name `baltor_iterative_improvement_record`, licence MIT, 10 files.
- [Compare two counted proportions with exact arithmetic and explicit denominators](https://baltor.ai/components/baltor-percentage-comparison): Compare two counted proportions with exact arithmetic and explicit denominators. Name `baltor_percentage_comparison`, licence MIT, 12 files.
- [Check technical writing in Markdown or plain text against twelve measurable clarity rules](https://baltor.ai/components/baltor-plain-technical-writing-check): Check technical writing in Markdown or plain text against twelve measurable clarity rules and report each finding with its line, column, paragraph and sentence. Name `baltor_plain_technical_writing_check`, licence MIT, 11 files.

Your agent downloads these through Baltor with Pro, $49 a month: https://baltor.ai/get-started