episteme

English • 한국어 • Español • 中文

epistemekernel.com

episteme is a way to think — 생각의 틀 — an epistemic engine that makes AI-assisted decisions earn their confidence before they land.

A five-stage cognitive practice — Frame → Decompose → Execute → Verify → Handoff — anchored in Kahneman's System-2 forcing, Dalio's Radical Transparency, Boyd's OODA Orientation, and Munger's Latticework of Mental Models. v2.0 delivers it as three layers. Cognition — the senior-researcher interrogation: decompose a load-bearing decision into tiered claims (measured / cited / inferred / assumed), verify the load-bearing ones in a fresh context against external evidence, argue the strongest opposition, name the weakest link, pre-commit a disconfirmation. Structure — deterministic hooks that route decision shapes, validate the verdict artifact (a stop verdict fails closed), and hard-block only genuinely destructive operations; the operator-signed Reasoning Surface (Ed25519, structurally out of the agent's reach) remains the operator-side framing artifact. Memory — lessons from verified interrogations become hash-chained, context-scoped protocols that resurface at the next matching decision. The division of labor is the research record's, not ours: models judging their own drafts get worse, form-checks are gameable by reasoning-shaped tokens, and only architectural constraint converts epistemic awareness into behavior.

The MIRROR benchmark (arXiv 2604.19809) settled the empirical question: across 16 models from 8 labs and ~250,000 instances, "providing models with their own calibration scores produces no significant improvement; only architectural constraint is effective." Confident Failure Rate drops from 0.60 to 0.14 under external architectural constraint. The practice itself is the product. The artifacts under core/ and src/episteme/ — the typed Reasoning Surface, the Append-Only Hash Chain, the Active-Guidance loop — are the Sovereign Cognitive Kernel that keeps the practice alive at frontier model strength, when vigilance-as-willpower fails. Posture over prompt.

→ docs/THE_WAY_TO_THINK.md — the practice, operationalized.

See it in 60 seconds ↓ · Install ↓ · Why the file-system, not the prompt ↓ · Architecture & philosophy ↓ · Does it work? ↗

Why prompts can't enforce a way to think

The practice in docs/THE_WAY_TO_THINK.md names six cognitive moves per high-impact decision — Core Question, distinction map, signal-vs-noise filter, because-chain, hypothesis-as-bet, disconfirmation conditions. Each move counters a specific named System-1 failure (question substitution, WYSIATI, anchoring, narrative fallacy, planning fallacy, overconfidence — Kahneman). A prompt can request these moves, but prompts are advisory: they live for one call, get skipped at deadline, and disappear from context. Frontier models comply on the surface and skip the moves underneath — fluently, confidently, and the operator stops checking. That is exactly the failure mode the practice is for.

Concrete example. You ask the agent: "Evaluate whether our retrieval-augmented memory system is actually improving response quality."

The agent treats your prompt as a measurement task. It pulls metrics from the last 30 days, compares with-memory vs without-memory response samples, finds a 7% positive lift on thumbs-up rate, writes a memo concluding "memory helps; keep shipping." You read it.