# PRE-REGISTRATION — The AI Creative Rights Contradiction Index (2026) > **Status: PRE-REGISTERED. No index computed yet.** This document commits the four contradiction-class definitions (as exact predicates on the data), the scoring rule, the falsification threshold, the honest-negative path, and the sample-selection rule BEFORE the index is computed. Committing it first is the DOI-Worthiness Gate's V2 requirement and the anti-HARKing discipline: a class definition, filter, or threshold added after the headline is seen reads as HARKing even when honest. > > **Author:** Vincent Wesley Couey (ORCID 0009-0005-6869-308X). **Pre-registered:** 2026-07-10. **Gate:** `GeneralBrains/methodology/gates/DOI-WORTHINESS-GATE.md`. **Scope:** `lattice-core/rinzara-site/RINZARA-DOI-STUDY-SCOPE.md`. **Working title of the eventual object:** "The AI Creative Rights Contradiction Index: quantifying where creative-AI tool grants and platform rules structurally contradict each other (2026)." --- ## 0. WHY THIS FRAMING, AND WHAT IT IS *NOT* The public discourse around AI creative rights is loud and anecdotal: legal blogs narrate single cases (Thaler, the Suno and Runway suits), vendors advertise "you own your output," and platforms publish their own AI rules, but **nobody has quantified how often these layers contradict each other from the standpoint of a good-faith creator.** A study that just compiled prices or feature lists would be a directory, not research, and would fail the gate (this is exactly why the §45 default "pricing index" was rejected at V1). **What this study measures instead (the novel lens, V1):** the *density of structural rights contradictions* across the creative-AI stack — the cells where a creator who complied with one layer (the tool's grant, or the platform's rule) is still exposed by another layer. This is a disagreement index for rights: it measures the conflict between layers, not a ranking of tools. Nobody publishes this. **It is NOT:** a claim that any tool or platform is acting in bad faith; a legal conclusion about any specific creator's situation; or a ranking of "best" tools. It reports *what the documented terms were, on a dated snapshot, and where they structurally conflict*, and nothing more. It is educational information, not legal advice (the YMYL honesty floor is load-bearing here). --- ## 1. THE METRIC (committed) The substrate is the on-disk `rinzara-data/` layer (every fact already carries `source` + `lastVerified` = 2026-06-08): **16 tools** (image 5, video 5, music 3, voice 3) and **9 publishing platforms**. The four contradiction classes are scored as **exact predicates on committed fields**: ### The four contradiction classes (frozen predicates) 1. **OWN-vs-COPYRIGHT gap** (tool-intrinsic). Fires when a tool grants ownership but the output is not copyrightable: `ownsOutput ∈ {"you-own","yes-granted"}` **AND** `copyrightable == "no"`. Meaning: you are told you own a thing you cannot legally stop a competitor from copying verbatim. 2. **SELL-vs-DEFEND gap / indemnity trap** (tool-intrinsic). Fires when a tool licenses paid commercial use but offers no IP indemnification: `commercialUse.paid == "yes"` **AND** `indemnification == "no"`. Meaning: you may sell it, but if it is claimed to infringe you defend the claim alone. 3. **OWN-vs-MONETIZE gap** (tool x platform, matched by modality). Fires for a (tool, platform) pair sharing a modality when the tool permits paid commercial use but the destination platform blocks or endangers monetization: `tool.commercialUse.paid == "yes"` **AND** (`platform.monetizable == "no"` **OR** `platform.strikeRisk == "high"` **OR** `platform.disclosure == "required-or-strike"`). Meaning: the tool says sell it, the platform will demonetize or strike it. 4. **FREE-TIER trap** (tool-intrinsic). Fires when a tool is marketed for commercial use but its free tier specifically cannot be sold: `commercialUse.paid == "yes"` **AND** `commercialUse.free == "no"`. Meaning: the "commercial-use" headline does not apply to the tier most people start on. > **Field-value normalization (committed):** any `ownsOutput` value containing "own" (case-insensitive) or equal to `"yes-granted"` counts as an ownership grant; anything containing "license" counts as license-only. `copyrightable`, `indemnification`, and `commercialUse.*` are read as their literal `yes`/`no` values; any value that is not exactly `"yes"`/`"no"` (e.g. "conditional") is recorded as `AMBIGUOUS` and **excluded from the numerator and denominator of that class** (never silently coerced), with the count of exclusions reported. ### The indices (committed) - **Tool-level contradiction density** = fraction of the 16 tools exhibiting at least one tool-intrinsic contradiction (classes 1, 2, 4). Reported overall and per-modality, with each class's own rate reported separately (never hide the classes behind the composite). - **Path-level contradiction density (the headline)** = fraction of all modality-matched (tool, platform) **paths** that contain at least one contradiction (classes 1, 2, 3 apply to a path; class 4 is tool-only and excluded from the path metric). A "path" = a good-faith creator using a tool's paid tier and publishing to a platform that accepts that modality. Reported overall and per-modality. - **Per-class density** = each class's own firing rate, reported for every class, so the composite can never over-read (a single near-universal class must be visible as its own number). - The **full per-cell contradiction matrix** (every tool, every path, which classes fired, with the source receipts) is deposited, not just the percentages. --- ## 2. HYPOTHESIS, FALSIFICATION THRESHOLD, AND HONEST-NEGATIVE PATH (committed) - **H1 (directional):** the creative-AI rights stack is *densely* contradictory, i.e. a majority of good-faith creator paths contain at least one unresolved rights contradiction. **Threshold to support:** path-level contradiction density >= 50%. - **H2 (structure, where measurable):** the contradiction is *not* uniform across classes; specifically the tool-intrinsic gaps (classes 1-2) are more prevalent than the tool x platform gap (class 3), i.e. the exposure is baked into the tools more than into the platforms. **Threshold to support:** class-2 (indemnity) rate > class-3 (monetize) rate across the matched paths. - **PRE-COMMITTED HONEST-NEGATIVE PATH (V5):** if path-level density is **< 50%** (contradictions are the exception, not the rule), that is a **valid, publishable finding** and the object ships as such: *"The AI creative rights market is more internally coherent than the alarmist discourse suggests, as of 2026."* This disconfirms the loud framing and is genuinely interesting. **We commit to minting on the null.** If we would only publish on a high-contradiction result, the design is confirmation-shaped and must be redrawn. - **The composite is descriptive, not evaluative.** Because one class (indemnity, class 2) is expected to be near-universal, the headline could be trivially high. The pre-committed guard: the object **leads with the per-class breakdown**, states plainly which classes drive the composite, and treats the *variance* (per-modality, class-3, and the temporal drift) as the substantive finding, not the near-certain classes. A high composite driven entirely by one near-universal class is reported as exactly that, never dressed as a surprise. - No re-cutting of the class predicates, the threshold, the normalization, or the sample after seeing the numbers. If the design is found flawed mid-run, the run is voided and re-pre-registered, not silently patched. --- ## 3. THE SAMPLE (selection rule committed, anti-B2) - **Tools:** all 16 tools present in `rinzara-data/tools.json` as of the frozen snapshot. This is the full on-disk set (the major image/video/voice/music generators the site covers), **not** a subset selected after seeing contradiction counts. The exact 16 are enumerated in the deposited `tools.json` snapshot; the selection rule is "every generator in the maintained data layer," disclosed so a challenger can inspect it. - **Platforms:** all 9 platforms in `rinzara-data/platforms.json`, likewise the full on-disk set, matched to tools by shared `modality`/`modalities`. - **No tool or platform is added or dropped to move the index.** If a reviewer thinks the sample is unrepresentative, the fix is to expand the maintained data layer and re-run, not to re-pick for this run. - **Snapshot:** the index is computed on a dated freeze of the four JSON files (`tools`, `platforms`, `legal-facts`, `policy-changes`), hash-recorded, so the exact inputs are reproducible. --- ## 4. SCORING RUBRIC (deterministic; committed) - The scorer is a single deterministic script (`score.py`) that reads the frozen JSON, evaluates the four predicates per tool and per matched path exactly as defined in section 1, and emits: the per-cell matrix, the per-class rates, the tool-level and path-level densities, the per-modality breakdowns, and the `AMBIGUOUS`-exclusion counts. - **No human judgement in the loop for the counts** — the predicates are pure field comparisons, so re-running the script on the deposited data reproduces the index exactly (this is what makes V3 hold). Any place the data is ambiguous is surfaced as `AMBIGUOUS` and excluded, never guessed. - The script + its output + the frozen data are all deposited so a hostile stranger re-runs and lands the identical numbers. --- ## 5. ADVERSARIAL AUDIT, PRE-COMMITTED (V4; the B1-B6 pass runs BEFORE mint) - **B1 (over-generation):** confirm every counted contradiction is a *real* good-faith-creator exposure, not a technicality. For a sample of fired cells in each class, write the concrete creator scenario it represents and verify the exposure is genuine (e.g., class 2 = "you can be sued over this output and the vendor will not defend you" is a real exposure, not a formality). Reconcile any cell that is a technicality out of the count. - **B2 (biased sample):** the "every tool/platform in the maintained layer" rule + the enumerated deposited snapshot defend against a contradiction-maximizing seed set. State the selection rule in the object and publish the full list. - **B4 (over-read summary stat):** always lead with the per-class breakdown and the full matrix; never let the single composite stand alone, especially given class 2 is expected near-universal. - **B5 (metric floor):** verify the scoring imposes no artificial contradiction floor; report each class raw so a floor would be visible. - **B6 (unrealistic construct):** verify each contradiction is a genuine legal/financial exposure a real creator would actually hit, not a contrived edge case; the class-3 platform predicate in particular must reflect a real "you published and got demonetized/struck" path. - **B-meta:** a guessed critique fails the audit as badly as a guessed result; the audit is written against the actual fired cells, with receipts. ## 6. REPRODUCIBILITY + MINT PLAN (V3, then the DOI) Deposit, dated: the frozen `tools.json` / `platforms.json` / `legal-facts.json` / `policy-changes.json` (+ their hashes), `score.py`, and the full per-cell contradiction matrix + all distributions — so a hostile stranger re-runs the scorer and lands the same index. THEN author the research-object page at `rinzara.com/research/ai-creative-rights-contradiction-2026/` (OLV-gated wrapper) and mint a Zenodo DOI under the ORCID, CC-BY, cross-linked into `data.deepsynthesis.org` + the Rinzara entity spine. **Only after V1-V6 + the floor all pass, AND after operator review of the computed result** (the DOI standard: permanent representations of the operator's work are not minted unilaterally). ## 7. GATE MAP (how the eventual object clears each vector) - **V1 novel lens:** contradiction density across the tool x platform x modality space, a measurement nobody publishes. ✔ (pending compute) - **V2 pre-registration:** THIS document, dated, pre-compute. ✔ - **V3 reproducible:** frozen data + deterministic scorer + full matrix deposited. ✔ (plan) - **V4 adversarial audit:** section 5, pre-committed. ✔ (plan) - **V5 honest-negative-safe:** section 2 commits to minting on the coherent-market null. ✔ - **V6 proud-to-sign:** a rigorous, sourced, reproducible quantification of a real socio-legal structure in the AI economy. ✔ (with the YMYL honesty floor) - **Floor:** every cell sourced + dated; framed as "documented terms as of 2026-06-08," never legal-conclusion language; zero em-dashes in the deposited object; no fabricated values. ## 8. LIVING AMPLIFIER (committed) The tool/platform rules change monthly and `policy-changes.json` already tracks them. Re-run the identical scorer on a fresh monthly snapshot and track how contradiction density **drifts** (is the market getting more or less coherent?). The drift is the temporal novelty and wires directly to the still-open monthly policy-change-alert magnet. --- ## ADDENDUM A (2026-07-10, dated, pre-computation) — Class 3 predicate correction **What happened:** on freezing the data snapshot and auditing the platform field *vocabulary* (the set of enum values, NOT the computed index), the original Class 3 predicate was found to reference enum values that do not exist in `platforms.json`. The real vocabularies are: `monetizable ∈ {yes, conditional}` (no "no"); `strikeRisk ∈ {low-if-compliant, low-if-disclosed, low-if-tagged, medium}` (no "high"); `disclosure ∈ {required, required-via-distributor, required-if-realistic, required-in-marketplace}` (no "required-or-strike"). The original predicate would therefore fire zero times purely because it named non-existent values. That is a specification bug, not a finding. **How it is corrected (on principle, before any index is computed):** Class 3 (OWN-vs-MONETIZE) is re-specified to the actual vocabulary on a strict, defensible principle: a **contradiction** exists only where a platform genuinely *endangers monetization* of compliant AI content, not where it merely imposes a *compliance step*. Therefore: - **Counts as a Class 3 contradiction:** `platform.monetizable != "yes"` (i.e. "conditional" or any future "no") — the platform does not cleanly allow monetizing the content. - **Does NOT count** (deliberately, to avoid over-generation, B1/B6): a mandatory-disclosure requirement (`disclosure == required*`) or a non-"high" strike risk. Disclosing AI use is a compliance burden the tool's grant does not conflict with; it is not a contradiction. Counting disclosure as a contradiction would inflate the index and fail the audit. **Integrity note (stated plainly, for the object and the audit):** this correction was made after seeing the field *vocabulary* but before computing any contradiction counts or the index. It therefore has *weaker* pre-registration strength than classes 1, 2, and 4, and B6 must scrutinize it hardest. The scorer will ALSO emit, clearly labelled as **exploratory and not part of the pre-registered index**, a broader "platform friction" variant (counting `conditional` monetizable + `medium` strike + mandatory disclosure) so a reader can see the full sensitivity, but the headline index uses only the strict corrected predicate above. The likely consequence, flagged in advance, is that Class 3 is rare or zero and the index is driven by the tool-intrinsic classes, which would *strongly support H2* (tool gaps > platform gaps). If so, that is the honest finding and is reported as such, not massaged. --- ## STATE Pre-registered 2026-07-10 (Addendum A same day, pre-computation). **INDEX NOT YET COMPUTED.** Next step: freeze the four JSON snapshots (+ hashes), write the deterministic `score.py`, compute the matrix + indices, run the B1-B6 audit, THEN pause for operator review, THEN author the object + mint. Do not skip to the headline or the content page; the number is computed from the frozen data by the deposited script, never asserted.