Skip to content

Reading Routes

John Kuhles edited this page Oct 31, 2026 · 3 revisions

Reading Routes

Fifteen-minute orientation

  1. Deep Ethics as a Working Concept
  2. Architecture Map
  3. C1 Before C2
  4. Correspondence, Corrigibility and Reality Veto

One-hour architecture route

  1. Ten Plus One
  2. Twenty-Seven Plus Twelve
  3. Believism and Value–Usefulness
  4. Topologetics and Compression
  5. Governance and Correction

Research and evaluation route

  1. Methods, Evidence and Experiments
  2. Provenance, Claim Status and Contribution
  3. Beta Paper V03b
  4. Cold Tests
  5. Decision Ledger

Journalism and media route

  1. Believism and Value–Usefulness
  2. PR Psychology
  3. Cognitive Warfare
  4. Truth Seekers vs Cautionmurmelism

A.I. design and evaluation route

  1. C1 Before C2
  2. Topologetics and Compression
  3. Meta Processing
  4. Neural Ethics
  5. Autocomplete

Biography and conditions route

  1. Biography and project origins
  2. Ten Plus One
  3. Iris
  4. Dutch Directness

Encyclopedic proposal route

  1. Deep Ethics as a Working Concept
  2. Provenance, Claim Status and Contribution
  3. Sources and Recovery Routes
  4. Grokipedia proposal

Intelligence properties and process quality

  1. Intelligence Properties and 1888
  2. C1 Before C2
  3. Process Quality and Reciprocal Scrutiny
  4. Meta Processing

Competing assessments and later correction

  1. Longitudinal Correction and VCRAS
  2. Asymmetric Grace and Bidirectional Scrutiny
  3. Correction Metabolism
  4. Controversy Inquiry and Source Recovery

A practical source-recovery route

Choose one case and source passage through Controversy Inquiry and Source Recovery. Recover the underlying material, record what it changes and use Contributing and Corrections to make the update inspectable. Keep useful contributions, unresolved findings and successful repairs available alongside detected errors.

Local AI and a testable operating environment

  1. Intelligence Properties and 1888 and Ten Plus One — the property field and mutually correcting orientation.
  2. Local AI and DeepEthicalOS — triad, attention, retained context and proposed host components.
  3. DeepEthicalOS candidate tests and Herdr — separate runtime continuity from correction transfer.
  4. Longitudinal Correction and VCRAS — follow dependent repairs and the next unfamiliar encounter.

Start with one source and a bounded comparison. Retain the original, actual access, authorization and remaining disagreement; include a useful no-change case.

Clone this wiki locally