Research Methods
The credibility apparatus is drawn in one picture on the research hub.
The methodological apparatus of the ARC / Eden research programme. Every method is defined, executed, and independently verifiable.
Core Methods
Truth-Gate
Every claim published by the programme passes through the truth-gate: it must be dated, hashed, publicly anchored, and independently verifiable. Claims that fail any criterion are held until evidence is secured. No claim leaves the programme without a kill condition.
Anti-Sycophancy Protocol
Every claim is tested adversarially, to falsify it rather than confirm it. Agreement is treated as suspicious, and disagreement is investigated. The author makes the final determination.
Honey Dashboard
Verification traps embedded in the evidence register. If a claim appears to be verified but the source does not actually support it, the honey trap triggers. Prevents self-confirmation bias in verification.
ARC-Align Test Protocol
Blinded alignment scoring methodology. Three independent seeds. 20 generations. 10 trade-off tasks. Eden vs Babylon. Scoring is blinded: raters do not know which system produced which output. success threshold fixed in advance.
Evidence Methods
Timestamp Verification
Every priority claim is anchored to the ARC-sealed recipient-form record, with Gmail origination headers corroborating, SHA-256 hash of the original content, and a public URL where the timestamped content can be independently verified. The standard is that each date carries its kind of evidence (a Bitcoin anchor, a deposit dated by its server on OSF, Zenodo or GitHub, a commit date the committer writes, or a date written inside a document) and that no kind is presented as stronger than it is; the 8 December 2024 record's date, for example, is its sender's Date header, which a DKIM signature authenticates as sent bytes but does not fix in time.
Kill Conditions
Every prediction carries an explicit kill condition: the observable fact that would prove it wrong. Predictions without kill conditions are classified as hypotheses, not claims. Kill conditions are written and dated before the outcome is known.
Canonical Facts Register
A single source of truth for all programme facts. Every fact has a SHA-256 hash, a date, a source, and a verification level (L0-L5). No fact is promoted beyond L1 without source verification. Facts are never deleted, only superseded with a dated correction record.
Deduplication Protocol
All evidence, claims, and convergences are checked for duplicates before registration. Duplicate identifiers are blocked. Duplicate SHA-256 hashes are merged. Every entity has exactly one canonical home.
Verification Methods
Retraction Protocol
When a result is found to be incorrect, it is retracted publicly with a dated correction, an explanation of the error and a link to the corrected record; a claim or wording found wrong is corrected in the open, on the corrections page. Retractions are never hidden. The error rate is part of the programme's credibility.
Silence Is Not Evidence
The absence of contradiction is not treated as confirmation. Claims that have not been independently verified are explicitly labelled as unverified. The programme does not inflate its evidence base by treating quiet as consent.
Single-Lab Constraint
The programme acknowledges that all primary research is conducted by a single researcher. Findings are presented with this constraint explicitly stated. Independent replication is actively sought. No finding is claimed as established science until independently replicated.
Michael Darius Eastwood conceived and directs this research programme and is the author of this work. Across the programme, he has used more than six AI systems in parallel, under his own instructions, to stress-test his arguments, identify possible errors, and assist in preparing draft text from his own outlines. He determines what is adopted, revised or rejected and takes responsibility for the published content. These systems are tools, not authors.