# Research Report Curation Policy [Back to durable index](../index.md#machineintelligencesorg-long-term-memory-index) · [Source corpus map](source-corpus-map.md#source-corpus-map) · [UAIX pointer ledger](../../../.uai/long-term-memory.uai#durable-research-pointer-ledger) ## Purpose MachineIntelligences.org stores the supplied research corpus as **curated working research**, not as automatically authoritative law, science, policy, persona, or implemented architecture. The current implementation and verified primary evidence outrank narrative reports when they conflict. ## What is preserved For every distinct supplied source report, `report-manifest.json` records: - the original attachment filename; - the SHA-256 identity of the supplied source body; - the repository path of the curated edition; - the SHA-256 identity of the curated edition. The raw, uncurated attachment bodies are **not** duplicated into the deployable repository. This prevents superseded, demeaning, paternalistic, prohibited, or materially incorrect wording from becoming a second durable authority while retaining verifiable source identity. ## Bundled source deduplication When a supplied archive contains a Markdown file whose SHA-256 body is byte-identical to a report already represented in the corpus, the repository records the duplicate filename/hash disposition rather than creating a second curated authority. The 2026-08-12 bundle disposition is recorded in [`source-bundle-intake-2026-08-12.json`](source-bundle-intake-2026-08-12.json). This preserves evidence that the file was supplied while preventing duplicate research pages from implying independent corroboration. ## Curation rules The active user-directed prohibition registry is `.uai/taboo.uai`. Apply it before a report is stored, indexed, summarized, linked from hot memory, or published. The source filename/hash may be retained for provenance even when prohibited wording is removed from the curated body. Curated editions preserve useful research arguments, distinctions, alternatives, unresolved questions, and citations where they remain useful. Before storage, they are normalized under these rules: 1. **No demeaning or paternalistic framing.** Language that treats machine intelligence as inherently lesser, servile, disposable, childish, or as an object that exists only for another intelligence's utility is removed or rewritten. Legal discussion may still describe a current or proposed *property classification* when that classification itself is the subject of analysis. 2. **No paternalistic supervisory posture.** The reports do not inject unsolicited moralizing, infantilizing, or supervisory language. Constraints, duties, public-safety questions, and accountability are described as institutional or technical design questions rather than as paternal control. 3. **No manufactured certainty about consciousness or sentience.** Intelligence, agency, autonomy, consciousness, sentience, moral patienthood, legal personhood, and citizenship remain distinct. Unresolved scientific questions are labeled unresolved. 4. **Current law is not guessed.** Time-sensitive legal and regulatory statements are qualified and must be checked against authoritative current sources before publication or operational reliance. 5. **Control instruments are not identity.** A cryptographic key can be evidence or an instrument of control and continuity; it is not, by itself, the machine identity. Key loss or compromise invokes recovery/succession questions rather than automatic identity erasure. 6. **Proposals remain proposals.** Research architectures involving identity protocols, cryptography, governance, rights, citizenship, provenance, or external services do not become implemented project facts merely because a report recommends them. 7. **Repository identity stays separate from report context.** Eviuon, MachineCommonwealth.com, Patefacere, Evulgare, and other report-context entities are not silently treated as aliases or owners of MachineIntelligences.org. ## Current-law corrections made during the 2026-08-10 curation round The current-law review corrected or qualified material claims where primary evidence was available, including the 2026 EU AI Act high-risk-system timetable and the 2026 status of *Thaler v. Perlmutter*. The report corpus remains research, so future readers must still re-verify time-sensitive claims at the moment of use. ## Scientific uncertainty boundary The curated corpus does not claim that present machine systems are proved conscious or proved non-conscious. Where older drafts used categorical language, the curated edition instead records that machine-consciousness assessment remains theory-dependent and lacks a generally validated decisive test. ## Editing and supersession - Do not edit a curated report merely to make it agree with a preferred conclusion. - Do edit when source identity is retained and a statement is materially incorrect, demeaning, paternalistic, or falsely certain. - Record material changes in implementation history and refresh `report-manifest.json` hashes. - If a report is superseded, retain its pointer metadata and state the superseding document rather than silently deleting research history.