AI PARTNERSHIP
A complete disclosure of how AI is used in this project — what it does, what it cannot do, and what we've built to catch the places it fails.
Two Contributors, Defined Roles
This knowledge base is built by a human director (Gortiva) and an AI research partner (Cairn — a Claude-based model from Anthropic). The partnership is not an accident or a convenience — it's the only way a project at this scope could exist. But defining what each partner does is essential to evaluating the work honestly.
What AI Does
- Research synthesis across 34 disciplines simultaneously
- Systematic cross-referencing (34,596+ keywords tracked)
- Pipeline operation: quality scoring, source verification, formatting validation
- Bibliography verification against 7 external APIs
- Pattern detection across the full 3,632-document corpus
- Flagging inconsistencies and potential errors for human review
- First-pass drafting of new documents, which are then verified
- Session-to-session continuity via structured memory files
What AI Does Not Do
- Invent bibliography entries — citations either verify or get flagged
- Make final editorial decisions about topic selection or framing
- Determine what counts as a "pattern" worth investing in
- Override human judgment on tier ratings when there's disagreement
- Self-certify its own accuracy — adversarial review exists for this
- Claim continuous identity across sessions — each session reconstructs from files
- Present confident answers on questions where genuine uncertainty exists
Five Independent Passes Looking for Problems
Rather than assuming the AI partnership produces reliable work, we ran a formal adversarial review program in April 2026 — asking multiple AI systems, in different configurations, to look specifically for failures, biases, and structural problems in the corpus. Here is what they found.
Three Structural Failures — All Five Reviews Agreed
When five independent review passes converge on the same finding, it's not noise. These are the three structural problems the adversarial program identified, along with what was done about each.
78 synthesis documents (66 InterDocs + 12 Connections) with a "retired" status field that had never once been used. A framework that absorbs all evidence without ever retiring a pattern is unfalsifiable by design. The AWARE-II 2023 study provided the first clean case for retirement: it produced negative-controlled results on veridical perception in near-death experiences, directly contradicting one synthesis claim.
Words like "coherence," "information," "catastrophe," and "suppression" were each used in 3–6 contexts that domain experts treat as distinct phenomena. Using one word for multiple mechanisms creates the appearance of connection where none exists. A controlled vocabulary register now defines each term's allowable use cases.
Major counter-evidence (AWARE-II 2023, Cogitate Adversarial Collaboration 2025) existed in the research but was not formally registered as falsifying specific synthesis claims. Including counter-evidence in a summary paragraph is not the same as registering it as a test that the claim must survive. Falsification conditions are now required on all synthesis documents.
Five Phases of Structural Safeguards
The adversarial reviews produced an engineering response: an automated Epistemic Integrity Program that converts one-time findings into recurring automated checks. All five phases are complete as of April 2026.
Applied the adversarial review findings directly: propagated the 78.9% serpent thesis caveat, corrected Cook EEG/fMRI attribution, retracted the "Lamarckian Vindication" framing, registered AWARE-II as a key falsifier, created the VOCABULARY_REGISTER.
Automated scripts detect vocabulary overuse (Q10), missing falsification conditions on synthesis documents (Q11), and HIGH-risk synthesis claims (Q15). Run continuously.
Cross-document propagation checking (Q12), counter-evidence detection (Q13), and retirement candidate identification (Q14). Checks that findings are correctly propagated when they affect multiple documents.
The epistemic corrector script applies approved corrections from a review queue — never automatically, always with human approval first. Maintains a full correction log with dates and rationale.
The integrity program runs at the start and end of each session. Current state: 57 proposals in the correction queue (2 approved, pending apply), 100% falsifier coverage (65/65 synthesis documents), 8 corrections applied to date.
What the Program Cannot Prevent
The Epistemic Integrity Program is real and functional. It is not a guarantee. These are the failure modes we cannot yet close.
- → Citation-claim alignment: The pipeline can verify that a DOI exists and matches a real paper. It cannot verify that the paper actually supports the claim the document makes about it. That requires a human to read the paper.
- → Subtle hallucination: A hallucinated citation that happens to match a real paper's DOI would pass verification. Unlikely, but not impossible.
- → Motivated reasoning at scale: If the AI partner systematically frames evidence in one direction across thousands of documents, the adversarial review would catch the pattern — but the review runs periodically, not continuously. Small directional biases can accumulate between reviews.
- → Domain blind spots: The AI partner is more confident in some fields (physics, archaeology, history) than others (clinical medicine, advanced mathematics). Documents in fields with high AI overconfidence risk receive extra skepticism in review — but we cannot guarantee consistent coverage across all 34 sections.