Provenance-aware evidence counting

What it takes and gives

Runs with

Python 3.10 or later

Entry points

count_independent_evidence, relay_non_backtracking

Known answers

8 cases its tests reproduce

Limits and scope

Limits

Counts are only as good as the provenance links: a report that hides its sources looks like a new root observation. Root observations are treated as independent even when they share an unrecorded cause. Reports may only cite reports about the same claim. The relay is a synchronous round-based simulation for small graphs (thousands of edges) and does not model message loss or delay.

Guarantees

The independent count is the number of distinct root observations reachable through sources; relaying, echoing and merging never increase it. Reports whose sources never reach a root are listed as unsupported and contribute nothing. In the relay, an agent ends up holding exactly the observations whose origin can reach it along directed edges, given enough rounds.

Assumes

Every report either is a root observation or names the reports it rests on; the provenance links are complete and honest. Two root observations are independent of each other; the tool cannot detect a shared hidden cause.

Where it is shown

• Checks for agent workDeterministic skills

Related components

• Typed work result with six outcomes
• Equal-cohort comparison with separate coverage
• Method-blind relabeling and behavioral equivalence
• Mechanism-held-out splits for synthetic task suites
• Silent concept-shift twin datasets
• Validate an append-only ledger linking each artifact version to what changed