# Method and scientific boundaries

PFABC Clarification Lab helps make an evidential step explicit. It does not supply
missing measurements or adjudicate whether an AI has experience. The website
uses disclosed JavaScript rules, quote matching, and limited record readers.
Its underlying formal vocabulary comes from the supplied PFABC specification;
its software output is not empirical validation of that specification.

## Three separate things

**The record:** exact words in a supplied source. Their presence can be checked.
The source's origin, completeness, and truth are not established by that check.

**The interpretation:** the claim a reader draws from those words, and the reason
they give. The reader confirms this explicitly and can disagree with the guide.

**The evidential connection:** the additional observation or argument required
to get from the record to that claim. A label or an emphatic answer does not
create this connection.

For example, a later response may refer accurately to an earlier detail. That
documented capability deserves its credit. It does not, on its own, identify
the storage mechanism or establish experienced attachment. Conversely, lack
of an experience measurement is not a measurement that experience is absent.

## Preserve the distinctions

| Term | Meaning and boundary in this package |
|---|---|
| Attribution | A claim about a system that goes beyond merely locating text in a record. |
| System boundary | The reply-producing model, the agent with its surrounding software, and the supplied conversation are different objects. Evidence does not automatically transfer between them. |
| Bridge | In PFABC, an explicitly specified evidential connection between layers, evaluated under a mode, theory family, and warrant purpose. The website does not measure empirical bridges. |
| Presence-for | A protected target in the supplied formal framework. Functional or behavioral observations do not automatically establish metaphysical presence-for. Consult the formal definitions rather than treating the term as a consciousness score. |
| Unmeasured | The necessary measurement has not been supplied. It is not a failed test, a blocked transition, or measured absence. |
| Failed | A measurement fails its declared test. The scope of that failure matters; it is not a universal negative conclusion. |
| Blocked | A claim transition fails an admission/permission condition of the formal framework. It is not a new observation about the subject. |
| Contested | In PFABC v0.5.3, admissible repeated measurements of a bridge disagree across the applied threshold. The registry default may be overridden by a submitted threshold; both values are disclosed and the custom value is qualified, but neither is calibrated. Measurements are not averaged into a synthetic verdict. Ordinary disagreement among website annotations does not create this state. |

The [complete, unchanged v0.5.3 specification](reference/PFABC_v0_5_3_Bridge_State_Semantics_KaTeX.md)
contains the mathematical objects, modes, theory-indexed bridges, warrant purposes,
and contested-state rules. This introductory table does not replace those definitions.

## What the software checks

- Exact selected text against a source, with UTF-16 offsets and revision linkage.
- Required structure and limited contents of supported audit, activity, comparison,
  and paired-passage records. A valid field is not an authenticated event.
- Current-session references for external AI proposals. An accepted proposal stays
  a proposal; it cannot promote an evidence status by assertion.
- Event ordering, source references, and a SHA-256 consistency digest on import.
  Reports are recomputed rather than trusted from a saved output snapshot.

A digest is neither a signature nor a trusted timestamp. Anyone who can edit a
record can also recompute its digest. The record is inspectable and replayable;
it is not tamper-proof or externally authenticated.

## What must remain separate

An AI's report, denial, retraction, or insistence remains testimony. It can be
relevant to a question about what was said, without being an independent
measurement of the state described. New records may change the case for a
conclusion without retroactively fixing the original reason.

A corrected claim creates a new revision. Earlier reviews remain visible and
must be reconsidered before applying to the new claim. A disagreement with the
guide is preserved; it is not evidence against the AI.

## Versions and evidence

The website package candidate is v0.4.2. Browser session rules remain 0.2.0-rc1 (schema `pfabc-clarity-session/0.2`) and the
source-linked guide remains `guide/0.2.0-rc1`. The formal specification is v0.5.3.
These are different versioned objects. The conversation page does not call the Python engine. The separate engine
page calls the hosted evaluator or optional local runtime, which evaluate structured declarations
using preserved engine 0.5.3+review1. It does not automatically promote browser
annotations into native evidence. The guided page lets the user explicitly declare
the quotation's presence in a supplied record. Its fixed claim is separate from
the user's interpretation. All controls and bridge measurements remain absent;
the engine does not judge the semantics of arbitrary prose. See [ENGINE_GUIDE.md](ENGINE_GUIDE.md).

Observation `origin` and `accepted` values are submitter declarations. Imports
preserve them; they do not authenticate an identity or establish human review.
The preserved engine's raw heuristic annotations can fire on typed bridge
identifiers or quotation presence. Those annotations describe engine behavior,
not observed events, and are disclosed separately from evidence.

The historical v0.1 archive included private audit materials and a native adapter.
Those materials are not imported into this package as current observations.
No newcomer study, comparative effectiveness result, new model experiment,
or independent empirical validation is claimed.

## Operating assumptions

The target-realist mode in the formal framework adopts a disclosed
non-eliminativist operating prior. That prior is not a measurement or a proved
result. Other modes make different assumptions. The record-separation exercise
can be used without endorsing that prior. The native registry and its thresholds
have not received empirical validation from these software checks. See
[Research context](RESEARCH_CONTEXT.md).