Research paper 03
Evidence Ledger Data Descriptor: A Case-Record Package for AI Workflow Evaluation
AI workflow evaluation needs data descriptors that do more than list artifacts. The evaluated objects are often operational records: change packets, route boards, scenario runs, evidence receipts, claim packets, source-unit reviews, and proof manifests. These records are meaningful only if their denominator, provenance, review state, and claim ceiling remain attached. This paper describes an evidence ledger data package used across local portfolio research drafts. The package centers on a case_record schema that binds the operating object, product lane, source context, protocol, verifier, review state, outcome, provenance, and nonclaims. The companion methods paper reports a validated corpus of 193 records across three lanes: 161 Signal Box change-packet records, 17 Interlock route-board records, and 15 Test Stand scenario-run records. Of those, 175 were accepted and 18 were failed Signal Box records retained in the denominator. This descriptor explains what the data object is, how it should be interpreted, how failures and held states are preserved, and what must happen before a public data release. Its contribution is a reusable data-card structure for agentic workflow evidence where the record of what cannot be claimed is as important as the record of what passed.
- Paper
- 03
- Authors
- A.G. Mauro and C.A. Harris
- Date
- 2026-07-19
- Collection
- Standing Framework Research
Abstract
AI workflow evaluation needs data descriptors that do more than list artifacts. The evaluated objects are often operational records: change packets, route boards, scenario runs, evidence receipts, claim packets, source-unit reviews, and proof manifests. These records are meaningful only if their denominator, provenance, review state, and claim ceiling remain attached. This paper describes an evidence ledger data package used across local portfolio research drafts. The package centers on a case_record schema that binds the operating object, product lane, source context, protocol, verifier, review state, outcome, provenance, and nonclaims. The companion methods paper reports a validated corpus of 193 records across three lanes: 161 Signal Box change-packet records, 17 Interlock route-board records, and 15 Test Stand scenario-run records. Of those, 175 were accepted and 18 were failed Signal Box records retained in the denominator. This descriptor explains what the data object is, how it should be interpreted, how failures and held states are preserved, and what must happen before a public data release. Its contribution is a reusable data-card structure for agentic workflow evidence where the record of what cannot be claimed is as important as the record of what passed.
← Back to research papers