Confidentiality labels that constrain value flow, with declassification only at explicit audited points: a standalone information-flow typing project.
Question. Which value flows does a confidentiality label permit, and what may
a lower observer learn from an output? A value type Secret ℓ A classifies a
value at a label ℓ; the first model uses a two-level Public ⊑ Secret label
order over a pure, total calculus, and the target property is
termination-insensitive noninterference between secret inputs and public
observations.
Boundary. A label is not cryptography and not an epistemic standpoint. A
Secret ℓ A does not make a value unreadable, authenticate a principal, enforce
runtime access control, or hide timing, storage or network side channels. The
baseline declassifies nothing: an ordinary Secret-to-Public flow is outside
the discipline, and any release requires a separate policy and theorem. As of
2026-10-04 no consumer and no frontier item is selected for this project.
That decision is about selection, not about existence, and the difference is
worth stating because it was measured on 2026-11-01:
metadatastician/lintel imports
SecretTypes.{Prelude,Vars,Lattice.Record,Lang,Noninterference.Generic}, defines
reader-sets : Lattice, and opens the generic TINI at it, so the estate does
have a mechanised consumer of this calculus. It is pinned at d1d843b; Lintel’s
own gate was run against this repository’s HEAD on 2026-11-01 and passes, so the
lattice re-representation described in the
proof-debt record is non-breaking for it. Recording a consumer that exists is
not a change to the 2026-10-04 decision, which stands: no consumer is selected,
and nothing here is designed around Lintel’s requirements.
The canonical statement is docs/secret-types.adoc. It records the owner decisions (D153, D154, and the 2026-10-04 confirmation of this standalone home), the first model, and the explicit non-claims.
Core theorem PROVEN (2026-10-12): termination-insensitive noninterference
for the two-point lattice, src/SecretTypes/Noninterference.agda
(tini), with the explicit and implicit flows refused by the type system
and shown to leak by computation (tests/reject/, Witness.agda). Since
2026-10-19 that theorem is an instance: typing and TINI are stated over a
Lattice record and any down-closed observer
(src/SecretTypes/Noninterference/Generic.agda), and the laws the proof
uses are certified (scripts/minimal-laws.sh) and each shown necessary
(Countermodels.agda). The fragment is a while language with fuel
semantics over any variables with decidable equality (two, lo and hi,
in the two-point development); see PROOF-STATUS.adoc. Termination is
the channel the theorem leaves open, by name.
The explicit public-outcome model in
src/SecretTypes/Noninterference/OutcomeVisibility.agda makes finishing versus
non-finishing observable. It proves that termination-sensitive
noninterference is false for the current typing rules, using the existing
well-typed secret loop with equal fuel on two low-equivalent stores. This
does not change the finishing-runs-only TINI theorem or decide which stronger
typing discipline a future termination-sensitive language should adopt.
Conjecture (not claimed): for a well-typed program, the public observer’s
Echo fibre over its outputs does not separate secret inputs — the Echo
reading of noninterference, and the first candidate connection
Secret → Echo for nextgen-typing/docs/TYPE-CONNECTIONS.adoc.
Grade. The proof receipt
(verification/receipts/agda_praxis.deed,
PROOF-POLICY.adoc §8.5.1 and §8.5.2) describes commit 292d566 on main. It
holds the receipts of CI run 43455458225 on commit 7dfeab2, and an echo
records that the squash merge into 292d566 left every leg’s inputs equal.
From them the pinned reader derives Proof Witness Grade PWG-2ii for each
proved claim. It is not PWG-1 because the Rocq and Lean legs of that run ran
on a bare runner with no image digest, and PROOF-POLICY §8.2 counts a witness
toward PWG-1 only with one; the legs job now runs in a digest-pinned image,
so the next recorded run restores PWG-1. The grade is what the reader
derives, never a value typed here. The proved claims type-check in Agda under the
canonical §2.1 invocation, --cubical-compatible included, with stronger flags
besides (scripts/proof-flags.sh). Each is also re-proved by hand in Rocq
(Coq 8.18.0, prelude only) and in Lean (4.24.0, core only, no Mathlib), with no
axioms in either, and the statements are compared in
verification/legs/README.adoc. That
comparison was written by Claude Code, and the owner reviewed it on 2026-11-13.
Not +foundation-diverse: no witness outside dependent type theory.
Per-claim evidence is in the Proof Ledger,
the machine record is
verification/receipts/agda_praxis.deed
(checked in CI), and proof debt is empty in all four
dispositions. This paragraph states no grade the receipt does not support.
-
This repository owns the confidentiality-label definitions, the typing and observation semantics, and its own proof evidence.
-
TYPE-CONNECTIONS is the shared family map. This family’s row landed on 2026-10-05 and nextgen-typing#118 is closed, which satisfies D153’s registration requirement. A row is not a dependency: no import exists in either direction.
-
Neighbours and their limits (
epistemic-types,systemet/anytype,echo-types,tropical-types) are covered in the canonical spec. None is a dependency. They are also not one another: a confidentiality label, an epistemic standpoint, a resource grade and an Echo residue are four distinct structures, and the spec says why they stay distinct. -
The origin record stays with
epistemic-types— its boundary/history note — and it is a record, not a dependency: that repository proves no confidentiality or noninterference theorem and hosts noSecretAPI. Its issue #32 was closed by the owner on 2026-10-05T06:37Z as delivered here, with no transfer and none required; its successor guard #39 closed on 2026-10-25. Nothing is to be reopened. The transition review records the verified cross-references, keeps its 2026-10-05 text as history, and carries a 2026-11-01 reconciliation addendum.
Code, configuration and scripts are Mozilla Public License 2.0
(MPL-2.0); prose documentation is CC-BY-SA-4.0. Both texts live in
LICENSES/, and per-file SPDX-License-Identifier headers are authoritative.