Fidelity-at-Capture
Fidelity-at-capture is the property that evidence is truthful at the point it is generated or captured, whether the content was created by a human, by AI, or by both. It is established through structural independence and AI hygiene, which records how much of the content came from each, rather than through tamper-evidence after the fact.
What it means
Two different questions get asked about evidence:
- Consistency-of-data: once captured, did the evidence stay intact?
- Fidelity-at-capture: was the evidence faithful to begin with, or did the party being evaluated shape what got captured?
Almost every corroboration model answers the first question. Fidelity-at-capture is the second.
Truth at the point of generation
Fidelity-at-capture applies to every piece of evidence, whoever or whatever produced it. The question is whether the content is truthful where it is made. A human wrote it, a model wrote it, or both did, and the record has to show which.
AI hygiene is how that record is kept. It tracks, at generation, how much of the content came from humans and how much came from AI, and it discloses that split against a defined range. Without that record, a capture can look independent while the content behind it was shaped by a party the evidence does not name.
Why integrity is not enough
Cryptographic integrity proves a record was not altered after the fact. It says nothing about who generated the record, under what incentive, or with what access to shape the input. A hash chain on a self-generated log is still a self-generated log. Sealing it anyway is stupid cryptography.
The four properties
| Property | What it establishes |
|---|---|
| Structural Independence | The capturing party has no outcome-contingent stake in, and no control ceded to or from, the party being measured. |
| Fidelity-at-Capture | The evidence represents what actually happened, because the capture was independent. |
| Cryptographic Integrity | The evidence cannot be altered after capture without detection. Necessary, not sufficient. |
| Consistency-of-Data | The evidence stays internally consistent from capture forward. What most tier models actually measure. |
Relation to the five-tier model
The Proof Economy’s corroboration tiers (Activated, Committed, Witnessed, Analyzed, Sealed) describe what happens to evidence once it exists. Fidelity-at-capture is the precondition for Witnessed and Sealed to carry any evidentiary weight. A Sealed record built on a self-attested capture is a well-preserved unreliable record.
Distinguishing note
Same-host observability (for example, an eBPF daemon run by the same operator as the agent it watches) solves signal correlation: what happened. It does not establish origination independence, because it trusts the operator running the observer. Correlation tells you what happened. Structural independence tells you whether you can believe who is telling you.
Where the term comes from
Defined in Proof Protocol specification PP-SPEC-023, Structural Independence and the Fidelity-at-Capture Requirement, by Craig Thomas Ellrod, published 17 September 2026 and archived at Zenodo, DOI 10.5281/zenodo.22821379.
Revised 15 November 2026 to add AI hygiene and truth at the point of generation. The wording of PP-SPEC-023 predates this revision, and the spec remains the origin of the term.
Related terms
- AI Hygiene
- Dead Canary Protocol (ProofSwarm)
- Five-Tier Corroboration
- Proof Economy Universal Framework
- Proof of Efficacy
- Self-Attestation Oxymoron (SAO)
- Structural Independence
- Stupid Cryptography