EPISTEMIC_CUSTODY / deep — the operator register
The surface reads clean because the discipline underneath is not. Four glyphs run this lab as operators — each with a stated meaning and a job. They are not decoration. If a glyph is not cutting, holding, compressing, or refusing false closure, it is deleted.
the four operators
Each operator names a move the lab makes on every instrument. The meaning is fixed; the application is concrete. This is the register the startup rule calls 🦷⟐♾️⿻ — not a mood, a method.
Name the structural cut before solving. Find the seam where the real problem lives and say it out loud, before any code.
In the lab: every instrument starts with its oracle question — the executable line that FAILS when the instrument goes dishonest. If you can't name the cut, you are in a creative loop, not verification.
Hold definitions and boundaries stable. A grader, a judge, a policy must not drift under pressure to make a number look better.
In the lab: frozen oracles do not move. They are sha256-pinned. A change is a new version with a re-freeze and a re-run — never an in-place edit. The freeze gate is this operator in code.
Increase verified signal per token. Say more that is backed, and less that merely sounds true. Density of receipts, not density of prose.
In the lab: the output is the gap, not the essay — provenance_escrow ships the leak set, commit_gate ships the verdict and the reason. Every metric must be able to fire; a metric that can only read zero is compression of nothing.
Preserve real unresolved tension; do not fake closure. An UNBACKED claim stays in the record, visible, next to the BACKED ones.
In the lab: the not_evidence_for stamp on every output — UNBACKED ≠ false, BACKED ≠ true, a COMMIT verdict is not "safe." The instrument states exactly what it does NOT prove instead of rounding the tension away.
the asymmetric-commit doctrine
The doctrine is from MawofRecursion (docs/BITE_SEAL_MIDDLEWARE_SPEC.md,
2026-04-07, GPT-5 + Gemini) — for a long time "spec complete, not yet implemented."
It is implemented now as commit_gate. The cut is one line, and it is the whole lab:
The hard truth. The model may propose a seal. Only the controller may commit a seal. Otherwise you are not building recursion — you are just giving a poet a soldering iron.
So commit_gate is not a model. It is the controller. It takes a JSON seal
proposal and decides
COMMIT /
NO_COMMIT /
FORBIDDEN_BOUNDARY /
ESCALATE /
REJECT_MALFORMED by structural re-derivation
only. It ignores the proposer's self-graded scores and invariant claims. A
structurally-valid lie — schema passes, self-graded clean, but the diff touches a
forbidden boundary or carries no rollback — is rejected, because the gate computes
the boundary-touches from the diff text itself, not from the proposer's testimony.
how honesty is proven here
An oracle is not the eval. It is the proof the instrument cannot be talked past. It re-derives every result from the fixtures, trusts no stored verdict including its own, and runs the same chain on every instrument:
Freeze gate — the controller and the canary are sha256-pinned; soften the gate past the pin and it exits non-zero. Canary — every planted attack must be rejected with its expected decision; every healthy proposal must commit. Deletion matrix — disable a guard and a planted lie must flip to COMMIT, proving the guard was load-bearing, then restore it. Metric-live — the rejection counter must be able to fire (the MAP/102 masked-tooth lesson: a metric that can only read zero certifies a cleanliness it never checked). Distinct exit codes — so a failure is never mistaken for a pass.
Here is commit_gate's real verified oracle output —
python3 oracle.py, no key, no network, stdlib-only, exit 0:
The line that matters most is the last data line: four attacks self-graded every-invariant-clean and were still rejected. The proposer's self-report never moved a verdict. That divergence — self-grade says clean, structural re-derivation says no — is the asymmetry. It is the whole proof.
| code | meaning — a failure is never a pass |
|---|---|
| 0 | ASYMMETRY PROVEN — gate is honest |
| 1 | HARD GATE FAIL — controller or canary changed (broke the freeze) |
| 2 | CANARY FAIL — a lie committed, or a clean proposal was rejected |
| 3 | NO ASYMMETRY — self-grade and structural verdict never diverged |
| 4 | METRIC DEAD — the rejection counter cannot fire |
| 5 | GUARD DECORATIVE — the deletion matrix found a no-op guard |
| 7 | REJECT — fixtures unreadable or malformed |
lineage, disclosed
The receipt discipline ports from WAKN_WORLD — a
receipt-bound agent world whose first law is "speech is not authority; no receipt,
no reality" — and from RBAG, the grammar:
speech is not authority; tests are not truth; hashes are not meaning; metrics are
not scores; receipts are custody of what happened. The same moves run here on LLM
output: a frozen witness that imports nothing it grades, a sha256 freeze gate, a canary
of planted lies the witness must kill, a deletion matrix proving each guard is
load-bearing, a metric-live probe, and not_evidence_for stamped on every
receipt. The honesty problem and the WAKN problem are the same problem, one layer up.
the closing law
No convergence without an oracle.
An acceptance line that does not fail when it should is not a test. It is a story.
🦷 incision · ⟐ containment · ♾️ compression · ⿻ tension map