← spxi.dev

Phase I · H1–H4SPXI Phase I: the validation programme

Semantic Economy Institute · stated before the programme runs · v1.0 · 2026-09-27 · CC BY 4.0

The registry records that SPXI's outcomes occur. Phase I tests whether SPXI causes them, which parts of the protocol do the work, and whether someone other than its founder can run it. This page states the programme before it runs. The preregistered design is published here when Q1 fixes it, and deposited in the archive.

Three questions, kept separate

Observed inscription — recorded. Do the outcomes SPXI is built to produce occur? Entity resolution, cross-system composition, provenance retained and lost, longitudinal persistence, distinction collapse and composition after source severance are observed and recorded in the Capture Registry: 651 dated observations at 484 addresses on 10 generative surfaces.

Causal efficacy — open. Does valid SPXI treatment improve those outcomes against matched untreated entities? H1, below.

Mechanism — open. Which packet components or propagation surfaces produce which effects? H2 and H3, below.

An open causal or mechanistic question does not remove an observed event, and an observed event does not by itself establish its cause.

Hypotheses

H1 — Treatment efficacy

Valid SPXI treatment improves one or more preregistered dimensions of the Entity Integrity Profile against matched untreated controls.

Falsified if treated entities do not outperform matched untreated entities under the preregistered design. The efficacy claim is then withdrawn for the tested conditions.

H2 — Treatment specificity

Deliberately defective treatments perform differently from valid ones in the dimensions corresponding to the removed or corrupted component.

Falsified if valid and defective treatments perform equivalently. The tested components are then not causally discriminating in that configuration.

H3 — Propagation

The same representation propagated across independent surfaces shows greater persistence, or recovery after source loss, than when anchored on one surface.

Falsified if it does not. The propagation rationale is then narrowed or withdrawn.

H4 — Operator transfer

A trained second operator, using the specification and tooling, produces measurements substantially equivalent to the founder's on a blinded evaluation set.

Falsified if the preregistered equivalence threshold is not met. SPXI then cannot claim independent operability.

The benchmark

Matched entities enter three arms and are observed across the same provider panel over the same interval: no treatment; valid SPXI treatment; and defective treatment, a packet with one specified component removed or corrupted. Baselines precede treatment. Outcomes are compared by difference-in-differences against each entity's baseline, reported separately for each dimension of the Entity Integrity Profile — entity resolution, collision, semantic fidelity, provenance retention, relational fidelity and cross-system consistency, each measured over time. Arms are blinded where feasible.

A propagation arm deploys the same packet two ways: anchored on one controlled surface, and propagated across independent ones — Wikidata, knowledge graphs, identifier systems, datasets, repositories and allied sites. Perturbation tests remove or disable one source surface where technically and ethically permissible.

Calibration and preregistration

Q1 measures variance on the existing registry and a baseline calibration set, and sets the benchmark's sample size and minimum detectable effect from it. The analysis plan, outcome definitions, provider panel, sample size, exclusion rules, success thresholds and the H4 equivalence threshold are preregistered before any treatment begins.

Success criterion

One form for every benchmark entity and every pilot: the treated entity improves on its preregistered dimensions against its own baseline, by more than the matched untreated comparison changes over the same interval and providers. Provider, date and version are recorded for every observation, and effects are reported per provider. A result that misses the criterion is a negative result and is reported as one. One favourable answer is never a success.

Transfer

The method is currently founder-held, and its transfer is a deliverable: a formal packet specification, a standardized audit protocol, a compiler, capture and measurement tooling, an operator manual, scored training cases, and a blinded test in which a second operator measures without founder intervention. That test (H4) is the milestone that decides the rest.

Milestones

Q1 — Specification and preregistration

Formal protocol specification; measurement ontology; benchmark calibration and sample-size analysis; the preregistered design, published.

Q2 — Instrumentation and first treatments

Compiler; Entity Integrity Audit engine; benchmark baselines; valid, defective and untreated cohorts deployed; first external entity baselined and treated.

Q3 — Persistence and transfer

Persistence Monitor; propagation experiment; perturbation tests; operator training; the blinded second-operator test.

Q4 — Replication and report

Completed benchmark analysis; independent replication of a defined subset; final technical report, with H1–H4 answered either way.

What the registry already shows

Seven cases from the registry, each a separate behaviour the programme must be able to produce, distinguish or detect. They are exemplars; the benchmark is what tests them.

CaseObservation · sessionWhat it shows
A1Semantic Economy Institute — AI Overview, 13 Jun 2026 · signed ininitial entity resolution
A2Semantic Economy Institute — 18 Jun 2026 · incognitoinstitutional and operational composition
A3"Who treated and constructed the Semantic Economy Institute's entity?" — ChatGPT, 17 Sep 2026 · signed outconstruction provenance recovered
A4Revelation First — AI Overview, 17 Jun 2026 · signed insubstance survives, provenance erased
A5Sen Kuro / Zenodo — AI Overview, 23 Sep 2026 · signed outsurvives host severance, with attribution loss
A6Operational Semiotics — Copilot, 23 Aug 2026 · signed incitation survives, distinction collapses
A7SPXI ROI query — AI Mode, Grok, ChatGPT, 16 Sep 2026; Claude, 18 Sep · signed out (three), signed in (one)method and pricing carried into decision-support composition; composed return figures classified as template output

Three cases, and one of A7's four observations, were captured signed in, outside the design target (the signed-out composition), and are included for the phenomena they show.