The agent tried. The gate said no.
This is the whole product in one screen. An irreversible command was about to run without anyone writing down what they knew. The gate is deterministic — it doesn't argue, it just holds — and it names the thinking that got skipped, not merely the field that was blank. Then the work gets done, and the command goes through.
$ git push --force origin master Execution blocked by Episteme Strict Mode: no valid ReasoningSurface or interrogation verdict exists for this high-impact op. REASONING SURFACE INCOMPLETE: high-impact op`git push --force origin master` with cognitive move(s)skipped — Frame · Core Question discipline — countersquestion substitution (Kahneman).Surface fails validation on: core_question, disconfirmation.Disconfirmation must be a concrete observable condition (…). Per-project advisory mode (not recommended) is enabled by`touch .episteme/advisory-surface`.$ cat .episteme/reasoning-surface.json "core_question": "Does force-pushing master destroy commits that exist only on the remote?" "unknowns": [ "If origin/master holds commits absent locally, this push deletes them irrecoverably" ], "disconfirmation": "If git rev-list origin/master ^HEAD returns any commit, abort the push and rebase instead." ✓ surface accepted · op admitted · recorded to the chainThree moves, and none of them are optional.
Declare
Before anything irreversible, the agent writes down three things: the question it is actually answering, what it doesn't know, and what would prove it wrong. Not a form to fill — the fields are the thinking.
Gate
A hook checks that artifact before the command runs. It's deterministic: no valid surface, no execution. Nothing here depends on the model choosing to be careful — which is the point, because that's exactly what stops happening under time pressure.
Learn
What the decision taught gets hash-chained into a protocol, so it resurfaces the next time a similar decision comes around. The record is dated and tamper-evident — six months from now, your reasoning is still there.
episteme is a way to think — 생각의 틀 — an epistemic engine that makes AI-assisted decisions earn their confidence: structure guarantees the thinking happens; factored verification checks it was real.
Frontier models are now fluent enough that the diff looks fine and the conclusions sound right and the operator stops actually thinking at the moment of decision. For everyday work this is fine. For decisions that cannot be taken back, it is the failure mode the cognitive-governance literature has been warning about.
You will type more. You will move slower at the irreversible gate. Your judgment will stay present where it used to silently go absent. Six months from now when something asks you why you decided what you decided — an auditor, a teammate, your future self — your own reasoning will be there: readable, verifiable, dated, unalterable.
That is not compliance. That is what it means to keep thinking while the model finishes your sentences.
read the full practice →Two commands, inside Claude Code.
Hooks, agents, and skills go live immediately — the gate starts holding your irreversible operations from the next session on. The episteme shell CLI ships with the source checkout rather than the plugin.