Evidential & Epistemic Framework

This specification defines the evidential classifications and forecasting requirements for The AI / SI Encyclopedia (thebestai.co).

---

#1. The Six Epistemic Classifications

Every substantive statement regarding AI capabilities, mathematical proofs, historical milestones, and future outlooks must be anchored to one of six evidential categories:

[Established] ──► [Well Supported] ──► [Preliminary] ──► [Contested] ──► [Hypothetical] ──► [Forecast]
 (Proof/Replication)  (Strong Empirical)    (Single Study)      (Disputed)       (Theoretical)     (Future Target)

#1.1 Established

• Definition: Findings that are mathematically proven, universally recognized across scientific literature, or replicated across multiple independent laboratories over an extended period.
• Criteria: Appears in standard peer-reviewed textbooks (e.g., Bishop, Russell & Norvig); supported by definitive mathematical proofs (e.g., Universal Approximation Theorem); verified historical facts with primary records (e.g., 1956 Dartmouth Workshop).
• Presentation Rule: State directly as facts without hedging.
• Example: "The backpropagation algorithm calculates the gradient of the loss function with respect to the network weights via the chain rule of calculus."

#1.2 Well Supported

• Definition: Empirically robust claims with broad consensus and extensive experimental replication, though subject to evolving technical nuances or boundary conditions.
• Criteria: Multiple independent academic evaluations, standardized public benchmarks across different institutions, verified industrial implementations.
• Presentation Rule: State authoritatively while noting known operational bounds or dependencies.
• Example: "Transformer architectures trained on large corpora demonstrate predictable performance scaling according to empirical power-law relationships over several orders of magnitude of compute."

#1.3 Preliminary

• Definition: Recent research breakthroughs, preprints, single-laboratory demonstrations, or novel algorithmic proposals that have not yet undergone widespread independent replication.
• Criteria: arXiv preprints, newly published conference papers, laboratory announcements with published code/weights.
• Presentation Rule: Explicitly identify the reporting institution and state that findings await independent confirmation.
• Example: "A 2024 laboratory study reported that test-time compute scaling improved mathematical reasoning on benchmark suites; however, independent multi-model evaluations are ongoing."

#1.4 Contested

• Definition: Scientific debates, disputed evaluation results, or active controversies where credible researchers hold conflicting views supported by competing evidence.
• Criteria: Competing papers disputing benchmark contamination, debates over whether current LLMs exhibit true understanding versus statistical pattern matching, disagreements on emergent capabilities.
• Presentation Rule: Present both positions neutrally with their respective arguments, experimental data, and counter-arguments without false equivalence.
• Example: "Whether performance jumps on reasoning benchmarks represent genuine emergent cognitive abilities or are artifacts of metric non-linearity remains actively contested between research groups."

#1.5 Hypothetical

• Definition: Formal theoretical concepts, thought experiments, or mathematical models of systems that do not currently exist in physical or computational reality.
• Criteria: Recursive self-improvement loops, cognitive architectures for Artificial General Intelligence, brain-uploading scenarios, Coherent Extrapolated Volition.
• Presentation Rule: Explicitly label as theoretical scenarios or philosophical models. Never describe as implemented or established capabilities.
• Example: "Under the hypothetical intelligence explosion model proposed by I.J. Good and expanded by Nick Bostrom, an advanced agent capable of recursive self-improvement could initiate rapid capability compounding."

#1.6 Forecast

• Definition: Explicit, testable predictions about future technological capabilities, economic impacts, safety milestones, or societal outcomes.
• Criteria: Must include operational metric, timeline, explicit assumptions, and confidence level.
• Presentation Rule: Document as a registered forecast with structured tracking.

---

#2. Standard Forecast Record Specification

Whenever an encyclopedia entry incorporates a forecast regarding AI, AGI, or ASI milestones, it must include a structured forecast block: