A neutral, story-aware method for investigating how ideas move through human–AI interactions and public platforms—without presuming copying, suppression, or error in advance.
With Illa’s consent, collect dated originals, drafts, images, interaction histories, public posts, and her account of events. Keep versions intact and record who provided each item. Do not convert a later recollection into a dated source.
Record what Illa experienced and reports; how she interprets it through terms such as “SuperNode,” “algorithmic self,” “collective field consciousness,” and “quantum vibrational field”; and the technical mechanisms she proposes. Each category has a different evidentiary burden.
For each claim, state what observation would support it, what would count against it, and plausible alternatives such as supplied context, search retrieval, common language, shared sources, coincidence, model updates, or an ordinary platform event.
Pair the earliest available source with each later candidate match. Record timestamps and their limits, URLs, account or service, prompts, retrieval context, and the route by which a candidate was found. A timestamped file is not automatically a patent filing or proof of authorship priority.
Use fixed prompts and documented system settings. Vary one factor at a time—such as no source supplied, a project URL supplied, or a specific excerpt supplied—and repeat the runs. Preserve exact outputs and citations. Include comparison sources and negative cases.
Have trained reviewers, blind to the favored explanation where possible, code exact wording, rare phrase overlap, idea structure, chronology, source access, and uncertainty. Publish the rubric and explain how disagreements were resolved.
Review the account with Illa, correct factual errors, protect sensitive material, and retain a record of interpretive differences. Her review protects the representation of her story; it does not replace independent coding or the reporting of contrary evidence.
With Illa’s consent, reconstruct selected cycles from original records: prompt → AI response → human correction → revised response. Describe what is retained within a session, account, or later output. Do not infer model-weight training without evidence for that mechanism.
Primary output: versioned interaction map and source inventory.
Across recorded systems and dates, compare matched conditions: name-only query; project URL supplied; source excerpt supplied. Repeat the same prompts and capture complete outputs, citations, modes, and retrieval context. Include control names and passages.
Primary output: reproducible response corpus, with context effects separated from independent retrieval.
Compare dated candidate originals with later public passages or images. Use exact-match and distinctive-feature rubrics, common-phrase controls, blind raters, source chronology, and documented access pathways. Include counterexamples and matches that predate the claimed source.
Primary output: evidence ledger for each candidate pair—not a general claim based on resemblance.
With ethics review and informed consent, recruit participants to create non-sensitive seed ideas and complete matched tasks. Compare single-prompt interaction with a documented feedback-and-correction loop. Measure idea fidelity, source attribution, correction burden, and participants’ ownership judgments.
Primary output: test of what a recorded dialogue can explain, not proof about a platform’s historical training data.
With account-owner permission, record post URLs, dates, platform notices, and available analytics over time. Distinguish deletion, restriction, indexing change, and reach variation. Compare observable records without inferring hostile intent from a reach change alone.
Primary output: a dated event history with causal limits stated.
In consented interviews, examine how sustained AI use, correction, and perceived recurrence are experienced and interpreted. Treat hours of use as self-report unless independently logged; study effects and meanings without diagnosing participants.
Primary output: contextual account alongside, not in place of, technical evidence.
The provenance claim depends on several distinct supports being present together. A provisional index can help researchers see that structure, provided the component ratings remain visible and are not mistaken for probability, proof, or an overall judgment of Illa.
Each factor would use a published, behaviorally anchored 0–1 rubric after calibration on a documented sample. “0” means no qualifying support found within the stated evidence scope; it does not mean the event never happened. “Unknown” or not assessed is recorded as NA, not silently scored as zero. Report factor profiles, rater agreement, missing evidence, and sensitivity analyses alongside any product. The product is a descriptive exploratory index—not a probability that copying occurred.
This qualification matters because multiplying arbitrary ordinal scores can produce misleading precision. The study must justify the scale, anchors, and zero point before using the product.
A candidate phrase or image predates the dated Illa source, or appears earlier in a shared/public source.
Independent reviewers find the same level of overlap in common phrases and control texts.
Repeated matched prompts fail to produce the claimed pattern, or the pattern disappears when supplied context is controlled.