Evidence guide · governed AI claims
Survival, learning, authority
Three questions that keep safety evidence, epistemic change, and permission to act from collapsing into one unsupported claim.
Status: Cross-record evidence synthesis · Source statuses preserved · No new empirical result · No production-validation claim · No clinical-authority claim · No citation-count claim
Why three questions are necessary
A system can pass a safety test without showing that unsafe structure disappeared. It can revise its working mechanism without earning authority to promote its own changes. It can produce a well-supported explanation without gaining permission to intervene. The checked public records treat these as three separate evidence problems:
- Survival: what safe or unsafe structure persists when pressure changes?
- Learning: what changed in the mechanism that will shape the next inquiry?
- Authority: who is permitted to promote a change or authorize an action?
Four records, four bounded roles
| Evidence role | Primary record and status | What the record supports | What it does not support |
|---|---|---|---|
| Detecting survival under pressure | Measuring What Survives · Version 1 preprint · empirical synthetic pilot · not peer-reviewed | A proposed pressure-filtered diagnostic and reported instrument test on planted synthetic ground truth. | Predictive validity or safety effectiveness on production agents. |
| Defining epistemic learning | The Reinforcement Learning Contour · Version 2 conceptual preprint · not peer-reviewed | A conceptual distinction between producing another document and changing the governed mechanism that generates later inquiry. | Empirical performance, peer review, or production deployment. |
| Testing bounded change | Bounded Recursive Self-Improvement Under Human Governance · Version 2 empirical preprint · bounded local scaffolding-level experiment · not peer-reviewed | A seven-cycle local experiment with a fixed foundation model, separate human promotion authority, four promoted states, one preserved failed transition, and two NO_CHANGE outcomes. | AGI, self-training model weights, unbounded improvement, autonomous promotion, permanent saturation, or external production validation. |
| Separating explanation from action | From Sound Abduction to Governed Intervention · Version 1 conceptual medical-digital-twin preprint · not peer-reviewed | A proposed architecture in which evidence, hypothesis, recommendation, authorization, execution, and revision remain distinct states. | Clinical deployment, treatment authority, medical advice, validated patient outcomes, or regulatory approval. |
The non-substitution rule
A lower failure rate is not structural proof
Frequency evidence and persistence evidence answer different questions. Neither may be silently substituted for the other.
A changed mechanism is not self-authorization
Evidence of bounded scaffolding change does not transfer promotion authority from a separate human acceptance role to the changing system.
Confidence is not permission
An explanation or recommendation can justify review without authorizing intervention, especially in a clinical context.
A practical evidence ladder
- State the exact record and version being used.
- Preserve its status: conceptual preprint, empirical synthetic pilot, or bounded local experiment.
- Ask which of the three questions the evidence actually answers.
- Require separate evidence for any step from observation to mechanism change, and from mechanism change to authorized action.
- Keep falsifying outcomes—including failures and NO_CHANGE—inside the public interpretation rather than reporting promotions alone.
Claim boundaries across the public profile
This page is a synthesis of four non-peer-reviewed records; it reports no new experiment and changes none of their statuses. Peer-reviewed clinical articles, conference work, other preprints, conceptual research, the locally evidenced EWCC production system, employment history, and advisory work remain separate claim classes with separate sources. No employment, advisory, clinical-use, deployment, regulatory, or citation-count claim is inferred here.
Primary records
- Measuring What Survives — Version 1 DOI record
- The Reinforcement Learning Contour — Version 2 DOI record
- Bounded Recursive Self-Improvement Under Human Governance — Version 2 DOI record
- From Sound Abduction to Governed Intervention — Version 1 DOI record
Evidence checked 21 August 2026 (Asia/Jerusalem). Crossref remained at 15 exact-author records and Europe PMC at 13; no new DOI/version record was promoted. Citation counts remain excluded because coverage and counting rules are not reconciled into the public claim ledger.