episteme is a sovereign cognitive kernel that extracts context-fit protocols from conflicting information, records every decision on a tamper-evident hash chain, and actively guides execution when the reasoning surface goes thin.
Agents fail confidently when they confuse fluent pattern-match with reasoning. episteme does not block that failure mode — it makes the mode structurally unavailable. Pillars below.
Forcing functions, not policies.
Four named blueprints (Fence, Consequence-Chain, Kernel-Adjacent, Architectural-Cascade) intercept writes at the exact point where patterns of confident-wrongness historically emerge. Each blueprint has a validator, a fallback, and a dogfood test — the kernel runs against itself before it runs against you.
Tamper-evident memory, not a log.
Every Reasoning Surface seal, protocol synthesis, and deferred discovery writes one line to a SHA-256-linked JSONL chain. A corrupted past is not just noticed — it is unprovable to replay. Session boot walks the chain and fails closed on mismatch.
Protocols from practice, not from doctrine.
The kernel observes decisions, extracts context-fit protocols, and injects them back into the next cycle through session digests and guide CLIs. The framework gets sharper every time you use it — not because you trained it, but because the chain made the lessons auditable.
Every non-trivial decision is written onto a Reasoning Surface before execution. Knowns are facts with current evidence. Unknowns are named, not hidden. Assumptions are declared as assumptions. Disconfirmation states what evidence would prove the plan wrong. The chain records the seal.
Run `episteme init` in your project root, then write a Reasoning Surface for your current core question. The kernel's guard creates .episteme/reasoning-surface.json on the next high-impact op.
No hash-chained records under $EPISTEME_HOME/framework/. The first protocol synthesized on a successful op (Fence removal, Axiomatic judgment, Blueprint D firing) lands as the chain's first record.
Speed comes from loop completion, not step-skipping. The agent that consistently closes Observe → Orient → Decide → Act outruns the one that collapses stages to feel fast.
Name the core question. Declare the uncomfortable friction driving the work. Build the distinction map — facts / unknowns / assumptions / preferences — before planning touches code.
Convert non-linear context into explicit tasks. Translate 'why' into 'how'. For major choices, state method, alternatives considered, and why this one fits the governing intent.
One bounded lane per owner. Reversible moves first. Record assumptions when data is incomplete. The hash chain seals each step.
Validate against success metric, not effort spent. Distinguish proven facts from inference. Evaluate the hypothesis — validated, refined, or invalidated.
Update authoritative docs. Capture unresolved risks and exact next action. Residual unknowns marked, not hidden.
Hover any node to surface its because-chain: the signal that was observed, the cause that was inferred, the decision that landed. Every protocol is chain-linked — provenance is non-optional.
Dispatch cascade detector before Fence via try/except ImportError so detector absence never breaks the guard pipeline.
Preserve unexplained kernel constraints until their purpose is reconstructed. Removal requires named reason.
On session start, walk the chain and report any prev_hash mismatch before admitting new writes.
Reject a Reasoning Surface whose disconfirmation field is empty or tautological. A hypothesis without a failure mode is not a bet.
episteme guideOne CLI. No config. Reads reasoning-surface.json, walks the hash chain, surfaces active protocols, and names the unknowns that are blocking the next step.
$ episteme guideEPISTEME · active guidance · v1.0-rc ▶ surface: 8m fresh · domain=complicated · tacit_call=false core question Does the CP10 scaffolding land the four-trigger detector, six-field validator, and hash-chained writer without orphan regression? ▲ unknowns (3) — proceed blocked until ≥1 cleared: 1. hot-path latency under 100ms p95 on non-cascade ops? 2. sensitive-path coverage vs. kernel-adjacent UPPERCASE docs? 3. deferred-discovery writer: one hash-chained record per entry? ✓ disconfirmation declared · 3 conditions✓ hypothesis stated · confidence=0.82 active protocols (4) · cascade-graceful-degrade (conf 88% · invoked 4) · fence-pattern-validate (conf 94% · invoked 11) · hash-chain-self-verify (conf 91% · invoked 2) · disconfirmation-required (conf 82% · invoked 6) ↳ chain head: a3c9f1b2 · seq 0013 · 12s ago