# HXP-005 — Intrinsic Evidence Capture and Correspondence

Status: Validation Protocol 0.1
Public article edition: 2026-09-09

## What intrinsic evidence can establish

Intrinsic evidence concerns observations of the work itself: a photographed surface, a recorded texture, or another measurement made under a declared capture method. It complements the credential's digital integrity and the optional physical carrier. A signature can establish that particular evidence was committed to a record. Determining whether a new observation corresponds to the same physical work is a separate measurement problem.

This article defines how that measurement should be specified, evaluated, and reported. It is a validation protocol, not a statement that arbitrary phone photographs can already authenticate artwork. The distinction matters for both artists and collectors. A carefully preserved enrollment file is useful evidence, but its existence does not establish a reliable correspondence detector. Equally, a different photograph of a genuine work is not automatically evidence of a counterfeit.

The intended outcome is a bounded, reproducible statement: an observation made within a named capture profile did, did not, or could not be shown to correspond to an enrolled template under a declared comparison rule.

## Three different kinds of comparison

| Comparison | Question answered | Important limitation |
| --- | --- | --- |
| Exact file digest | Are these the same recorded bytes? | Recompression, metadata edits, and a new capture usually change the digest. |
| Canonical geometry digest | Do supported point data produce the same named canonical representation? | Geometric normalization is not a universal same-object detector. |
| Intrinsic feature correspondence | Does a fresh observation meet a validated similarity rule for this object class? | Requires capture controls, thresholds, and empirical evaluation. |

The `exact-byte-inventory-v1` profile belongs to the first category. It can establish consistency with an enrolled file and help detect substitutions in a package. It cannot support a fresh-capture correspondence claim. The geometry commitment described in [HXP-011](/standards/scan-binding) supplies another specific comparison rule, with its own limitations. Neither mechanism should be relabeled as a validated visual recognition system.

## The initial public extractor

The initial public extractor is `helix-public-tiles-v1`. It produces a deterministic 8×8 RGB feature vector from a controlled image. This deliberately simple representation makes the enrollment concept inspectable and reproducible: a declared processing method turns an image into a compact template, and the template can be committed to the signed evidence record.

Its simplicity also constrains what can be inferred. Coarse color structure can be shared by different works, while lighting and framing can substantially affect observations of the same work. The extractor therefore provides a signed enrollment template, not a production correspondence claim. A demonstration that reproduces its output is a software consistency check. It is not evidence of resistance to deliberate lookalikes, capture variation, or adversarial reproduction.

Any replacement extractor needs a distinct method identity, reproducible processing rules, and appropriate test vectors. A more sophisticated model does not inherit validation merely by occupying the same place in the interface.

## Define the capture profile first

A capture profile states which object class and observation conditions the comparison addresses. It identifies lighting, framing, calibration, resolution, permitted preprocessing, extractor identifier and digest, quality rules, and resource limits. These details are part of the meaning of the result. They should not be silently inferred from a filename or the presence of an uploaded image.

For example, a profile designed for controlled photographs of a flat painted surface should not automatically apply to a reflective sculpture, a framed work photographed through glass, or a textured object captured at an oblique angle. Each case introduces different sources of ambiguity. A usable profile tells the operator how to obtain an admissible observation and tells the verifier when the observation falls outside its scope.

Public documentation should describe the applicable conditions and decision semantics. It need not disclose protected training images, private object inventories, proprietary calibration procedures, or confidential experimental materials.

## Enrollment and fresh observation

Enrollment establishes the reference against which later observations will be compared. A defensible enrollment record associates the work, capture profile, processing method, template, and evidence commitment. Operator and session context may be needed for a controlled study, but unnecessary personal information should not enter a public credential.

A fresh observation is processed by the same authorized extractor and evaluated under the matching profile. Changing preprocessing, model weights, or normalization while retaining the same method label undermines reproducibility. Where methods change, the comparison needs an explicit compatibility rule or a new enrollment; it should not quietly reinterpret an older template.

Capture quality is assessed before a similarity decision. Severe glare, incomplete framing, insufficient detail, or an unsupported acquisition method may prevent a meaningful comparison. Such observations should be recorded as unusable or indeterminate according to the profile, rather than forced into the nearest positive or negative answer.

## Scores, thresholds, and honest outcomes

Matching thresholds are versioned by object class and capture profile. A report identifies the score, threshold, relevant quality assessment, device, operator, session, and result: `match`, `no-match`, or `indeterminate`. Identifying information can be retained in an authorized evaluation record while the public report carries only the context necessary to understand its scope.

A score is not inherently a probability that an artwork is authentic. Its interpretation depends on the metric, the population tested, and how the threshold was selected. A match supports the named correspondence claim within the evaluated conditions. It does not independently establish the creator's identity, legal ownership, current registry status, or physical-token continuity.

A no-match requires equally careful language. It may indicate a different object, but it may also reflect treatment, an out-of-profile capture, or natural changes. The result should preserve the distinction between a valid negative measurement and a measurement that cannot answer the question.

## Design the evaluation to prevent leakage

Training, threshold selection, and final evaluation sets must be separated by physical object, device, operator, and session as required by the study design. Splitting individual photographs at random can place nearly identical observations of the same object in both development and evaluation sets. That can make apparent generalization much stronger than the evidence supports.

Threshold selection belongs in the development stage. The final evaluation should use the frozen method and decision rule, with documented exclusions established in advance. Repeatedly adjusting a threshold after inspecting test failures converts the test set into another development set. A subsequent independent evaluation is then needed.

Sampling should cover both genuine repeated observations and different-object comparisons. The number of pairwise comparisons is not automatically the number of independent samples: many comparisons may reuse the same work, capture session, or device. Reports should describe those dependencies and choose uncertainty estimates that respect the study design.

## Variations and challenging cases

The validation dossier includes natural changes, conservation treatment, glare, focus, scale, crop, contamination, and deliberate lookalikes. These categories represent different failure mechanisms and should remain distinguishable in the results. A method that tolerates crop changes may still fail under reflective glazing; a method robust to modest lighting changes may confuse similar compositions.

| Evaluation group | Evidence needed |
| --- | --- |
| Genuine repeat captures | Repeatability across permitted sessions and equipment |
| Different works | Separation across the intended object population |
| Similar or reproduced surfaces | Behavior on challenging negatives, not only easy examples |
| Changed or treated works | Scope of continuity after documented physical changes |
| Poor-quality observations | Reliable abstention and clear recapture guidance |

Aggregate accuracy alone can hide these differences. Report false acceptance, false rejection, and indeterminate outcomes with denominators, uncertainty, and the applicable conditions. Unsupported categories remain outside the claim even if a demonstration happens to return a favorable score.

## A practical example

An artist enrolls a controlled close-up of a painting and later supplies another photograph taken through reflective glass. The new file's digest differs from the enrolled file. That is expected and does not contradict the signed record. If the fresh photograph fails the capture profile's quality requirements, the appropriate correspondence outcome is indeterminate, accompanied by guidance to recapture under suitable conditions.

If a qualified method later evaluates a suitable observation, that result becomes an additional evidence signal. It does not replace the original file commitment or silently revise the historical credential. This separation allows the system to retain useful records while improving measurement methods over time.

## Qualification and related reading

No correspondence claim advances from the existence of an extractor alone. Qualification requires a documented method, representative evaluation, calibrated decision rules, reproducible results, and a clearly bounded statement of performance. The current public extractor remains an enrollment demonstration until that evidence exists.

Read [HXP-001](/standards/core-capsule) for evidence commitments, [HXP-003](/standards/open-verifier) for independent verification signals, [HXP-009](/standards/conformance) for qualification gates, and [HXP-011](/standards/scan-binding) for exact capture and geometry commitments. Together they separate recorded evidence, measured correspondence, and the broader authenticity claim.