Frontier models are fluent enough now that wrong answers arrive sounding finished. episteme makes a load-bearing conclusion earn its confidence before it lands — decomposed into claims, verified against evidence, argued from the other side, stamped with a verdict.
conclusion “memory system improves response quality”
01 · decision
“memory system improves response quality — keep shipping.”
02 · decompose · tiered claims
+7% thumbs-up lift over 30 days
supportedwith-memory responses run ~30% longer
the lift comes from memory, not length
unverifiablethumbs-up tracks answer quality
refuted03 · verify · fresh context
the verifier receives the claim only — never the draft's prose. it answers from evidence, not from the argument.
04 · oppose
length alone predicts thumbs-up — the lift may be the length effect wearing a memory badge.
weakest link c3 · cheapest decisive test: re-run with length controlled
disconfirmation lift disappears under length control
05 · verdict gate
the gate starts closed and only a valid verdict opens it — a stop verdict fails closed.
06 · lesson → chain
control for response length before attributing quality lift
verified lessons become context-scoped protocols — hash-chained, resurfacing at the next matching decision.
● load-bearing claims are verified in a fresh context · supported / refuted / unverifiable · most interrogations teach nothing durable — the lesson is nullable.
The interrogation itself: decompose the conclusion into claims, send the load-bearing ones to a fresh context that never sees the draft, argue the other side, commit to what would prove it wrong.
Model-judged, because meaning lives where rules can't reach.
File-system hooks route decision shapes, check that the verdict artifact exists and holds (a stop verdict admits nothing), and hard-block only the genuinely destructive. Everything is recorded.
Deterministic, because deadlines are exactly when discipline gets skipped.
When a verified interrogation teaches something durable, the lesson is sealed into a hash-chained protocol scoped to its context — and resurfaces, unasked, at the next matching decision.
Chained, so the agent gets sharper on your codebase — not on the average of the internet.
It runs where you already work: a Claude Code plugin and a CLI today — and because the kernel is plain files (markdown, JSONL, hooks), the practice travels with you across tools, not with any one vendor.
A falsifiability condition only counts if it can fire. For 49 days the framework ran live and synthesized zero protocols — the one emit path was tied to the rarest operation class. Condition E1 fired. First the kernel was made to measure the miss in its own reports; then the loop got a real source — every verified interrogation whose lesson survives becomes a context-scoped protocol. The first one landed the same night. It's transcribed below, exactly as sealed.
“Providing models with their own calibration scores produces no significant improvement; only architectural constraint is effective.”MIRROR · 16 models · 8 labs · ~250,000 instances — arXiv 2604.19809 →
In context `episteme release flow under operator-granted overnight autonomy`, treat release-please merges as in-scope when the directive asks for a finished product by morning — and name the judgment call explicitly in the handoff.
№1 — synthesized 2026-06-10 from a verified interrogation · verdict: proceed · sealed against sha256:GENESIS · the live chain renders in the dashboard
Frame names the one question. Decompose turns the why into a how. Execute moves in reversible steps. Verify judges against the metric, not the effort. Handoff persists what was learned. Each stage leaves an artifact on disk — a surface, a verdict, a sealed envelope.
Five stages. The gate reads the artifacts they leave behind.
# interrogation №1 — transcribed from the chainEPISTEME · epistemic interrogation conclusion under review "merge release PR #94 — cut v1.7.0-rc1" ⊢ claims decomposed · 3 — all load-bearing✓ verified against evidence — CI green ×3, release flow cited✓ opposition argued — a public rc the operator hasn't read▲ weakest link — ‘merge it’ may not have covered the release PR✓ disconfirmation — operator reverts the tag on waking verdict · proceed↳ lesson lh_71f88adef21147df → protocol №1 · chained