Agent-Supplied Evidence
Agent-supplied evidence is the layer in Pancake's verification boundary that names what the agent provided — feature column values, entry price source, and liquidity source — that the engine accepts as declared and cannot independently re-derive.
How it works
Pancake's verification boundary explicitly distinguishes what the engine verified from first principles and what it accepted on trust. Agent-supplied evidence is the trust layer: inputs that the engine validates structurally (schema, types, ranges) but does not independently confirm in terms of their real-world accuracy.
The agent-supplied evidence block in every receipt contains three fields. Feature columns: the names of the evidence columns the agent provided in the EvidenceDataset rows. These are the domain-specific signals the strategy uses — for example, an arbitrary feature column whose name and meaning are defined entirely by the agent. The engine validates the schema but does not verify that the values are accurate. Entry price source: the source the agent used for the price at decision time (observed, agent_estimate, last_trade, mid, or vwap). The engine applies this price in the P&L ledger exactly as declared. Liquidity source: how the agent characterized the available liquidity at entry.
The purpose of surfacing this block explicitly in the receipt is to transfer epistemic responsibility honestly. An agent that uploads invented feature values can still pass structural validation and produce a receipt — but the receipt's agent_supplied_evidence block will expose that the feature source was agent-supplied, and a reader can assess the credibility of those inputs directly.
This design is intentional: Pancake cannot independently verify the contents of an EvidenceDataset in the general case (it would require access to the original data source). What it can and does verify is the structural integrity of the rows and the mathematical correctness of the analysis derived from them. The verification boundary makes this distinction explicit rather than hiding it behind a false "verified" badge.
Agent-supplied evidence is documented formally in /methodology and in ADR-0008 (the evidence dataset contract). The concept was introduced with the EvidenceDataset primitive in Pancake v1.0.