Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions CHANGELOG.adoc
Original file line number Diff line number Diff line change
Expand Up @@ -35,9 +35,9 @@ Ledger: `+L4-bond-merge+`, `+L4-lifting+`, `+A2-chain+` and `+retention-order+`
Lean legs, with receipts from CI run 44006871652 on `+d5234b1+`. The census moves
from 28 obligations (`+PWG-3+` 18, ungraded 10) to 28 (`+PWG-3+` 15, `+PWG-2ii+` 7,
ungraded 6). The derivation does not reach `+PWG-1+`, because the receipts are not at
the ledger's source commit. Whether the core-Lean leg may count, given the
2026-11-09 ruling that Lean is a generator and not a leg, is put to the owner.
If it may not, the seven rows derive `+PWG-2i+`.
the ledger's source commit. The core-Lean leg counts by owner ruling of 2026-11-17,
which ADR 0005 records in its amendment: the 2026-11-09 ruling ("Lean is a
generator, not a leg") covers the Mathlib development only.

==== Fixed

Expand Down
6 changes: 3 additions & 3 deletions README.adoc
Original file line number Diff line number Diff line change
Expand Up @@ -174,9 +174,9 @@ under `reference`, never as a leg, so it can neither raise nor lower a grade.
Both are rechecked in Rocq (`verification/proofs/rocq/`) and in core Lean with no
Mathlib (`verification/proofs/lean4/ratified-order/`). The ledger grades L4 as
four obligations and A2 as two (owner ruling 2026-11-16). All six, and the
retention order (a decidable preorder), derive `PWG-2ii`. Whether the core-Lean
leg may count is put to the owner (PROOF-STATUS, *Recheck legs of the order
laws and L4*). Completeness of the chain, which spec §7 also uses, is proved in
retention order (a decidable preorder), derive `PWG-2ii`. The core-Lean leg counts
by owner ruling of 2026-11-17, which ADR 0005 records in its amendment: Lean with
Mathlib stays a generator, never a leg. Completeness of the chain, which spec §7 also uses, is proved in
no prover. The atom
factorisation (O1's carrier half, proved on the ratified `GradeR` in
`EffectAtomsRatified.lean`; O1-coord stays open) and the tropical factorisations
Expand Down
82 changes: 82 additions & 0 deletions docs/decisions/0005-v0.2-consolidation-o5-and-prover-roles.adoc
Original file line number Diff line number Diff line change
Expand Up @@ -118,3 +118,85 @@ decided by this ADR.
* *Leave O5 as the workbench proposal* — rejected: an obligation cited by the
checker's firewalls must live in the checker's spec, not in a proposal
downstream cannot cite as normative.

== Amendment, 2026-11-17: D4 and D5, the §2 recheck legs

*Status of this amendment:* owner rulings of 2026-11-09 and 2026-11-17, recorded
here. The ADR as a whole stays *Proposed* until the owner accepts it.

*Why.* D4 made "Lean 4 + Mathlib" this repository's provenance prover. PROOF-POLICY
§2 (standards `3-practice/PROOF-POLICY.adoc`) defines no such role. It requires
every Agda theory proof to be rechecked in Rocq and in Lean, with no Mathlib: "a
proof that needs a large external library to compile is a dependency, not a proof
the estate owns". Two owner rulings close that gap.

* *2026-11-09, Lean as a generator.* The Mathlib development (the
`grade-boundary` and `grade-factorisation` packages) keeps Mathlib and keeps its
purpose: finding and stress-testing theorems, which are then restated in core
Agda. It is never a §2 leg. The ledger records each of its legs as
`route: re-proved, state: BRIDGE-PENDING`, which does not count (§2.2, §8.2).
* *2026-11-17, rechecked in Rocq and Lean.* Theory goes in Agda and is rechecked
in Rocq and Lean, with no Mathlib. Asked whether the 2026-11-09 ruling also
covers a Lean leg that imports nothing, the owner ruled that it does not. Such a
leg is a §2 recheck leg, and it counts when its correspondence is `RECONCILED`.

*D4, as amended (prover roles).*

[cols="1,3"]
|===
| Prover | Role

| Agda
| The normative kernel, unchanged: Theorem B and every theory obligation, under the
§2.1 invocation `agda --safe --without-K --cubical-compatible`.

| Rocq
| A §2 recheck leg, route `re-proved`: `verification/proofs/rocq/`, with no
`Require` beyond the prelude. Gate: `scripts/check-recheck-legs.sh rocq`.

| Lean, core
| A §2 recheck leg, route `re-proved`: `verification/proofs/lean4/ratified-order/`,
with no import (no Mathlib). Gate: `scripts/check-recheck-legs.sh lean`. Lean is
an independent kernel of the same dependent-type family as Agda and Rocq, so
§8.3's `+foundation-diverse+` tag is not earned by it.

| Lean + Mathlib
| A generator, never a leg (2026-11-09): `verification/proofs/lean4/grade-boundary/`
and `grade-factorisation/`. Its results become gradeable only once restated in
core Agda. Its rows stay `BRIDGE-PENDING`.

| Idris2
| The executable reference, unchanged. PROOF-POLICY §8.1 keeps it off the PWG
ladder, so the ledger records it under `reference`, never as a leg.

| Rust
| The fast core, unchanged. It is validated against the reference over the
conformance corpus and proves nothing.
|===

*D5, as amended.* D5 removed three mint templates, among them
`verification/proofs/coq/TypeSafety.v`, because they were never instantiated and
the Coq recipe skipped silently. That reasoning stands. The Rocq development added
on 2026-11-17 is different in kind. It is an instantiated recheck leg. Its gate
asserts the Coq version, fails on a missing prover, fails on any axiom and on an
empty audit, and runs planted controls in CI. The repository therefore has four
proof developments (Agda, Rocq, Lean, Idris2) plus the Rust fast core. "Three
provers" in D5 and its consequences is superseded by this table.

*Consequences of the amendment.*

* `docs/status/PROOF-STATUS.adoc` no longer marks the core-Lean legs as awaiting
the owner. The paragraph there on the 2026-11-09 ruling points here instead of
to an amendment "tracked separately".
* The seven rows whose Rocq and core-Lean legs count (`L4-fate`, `L4-fidelity`,
`L4-bond-merge`, `L4-lifting`, `A2-directions`, `A2-chain`, `retention-order`)
keep their derived `PWG-2ii`. No grade moves because of this amendment: it records
the ruling the ledger already followed.
* A new theory result reaches `PWG-2ii` only by the same route: an Agda source,
then a Rocq leg and a core-Lean leg, each with a recorded §2.2 correspondence
and a receipt. A Mathlib Lean proof of it is welcome as a generator, and it
counts for nothing on the ladder.
* Not changed by this amendment: `just proof-status` and the other recipes D5
touched still describe the developments they ran on 2026-10-27. Bringing them in
line with this table is upkeep (doctrine 16), and it is tracked as such rather
than done here.
16 changes: 8 additions & 8 deletions docs/status/PROOF-STATUS.adoc
Original file line number Diff line number Diff line change
Expand Up @@ -353,9 +353,9 @@ their titles state, and that theory "goes in Agda, rechecked in Rocq and Lean",
with no Mathlib. The 2026-11-09 ruling below ("Lean is a generator, not a leg")
was given about the Mathlib development, for the Mathlib reason. The Lean leg here
is a separate file with no import at all. It is recorded `RECONCILED` and it
counts. *That reading is put to the owner for confirmation.* If the owner keeps
every Lean row off the ladder, the Lean legs of the seven rows become
`BRIDGE-PENDING`, and those rows derive `PWG-2i` instead of `PWG-2ii`.
counts. Asked on 2026-11-17, the owner confirmed that reading. The 2026-11-09
ruling covers the Mathlib development only, and a core-Lean leg is a §2 recheck
leg. ADR 0005 records both rulings in its amendment of 2026-11-17.

[cols="1,3"]
|===
Expand Down Expand Up @@ -426,16 +426,16 @@ Mathlib cone therefore never enters the trusted base, and no Lean row here is or
will be a §2 recheck leg. The ledger records every Lean leg of the Mathlib
development as `route: re-proved, state: BRIDGE-PENDING`, which does not count
(§2.2, §8.2). (The core-Lean legs added on 2026-11-17 for the seven order and L4
rows import nothing, Mathlib included, and are recorded `RECONCILED`. Whether this
ruling also covers them is put to the owner: see <<recheck-legs>>.) The
rows import nothing, Mathlib included, and are recorded `RECONCILED`. The owner
ruled on 2026-11-17 that this ruling does not cover them: see <<recheck-legs>>
and ADR 0005, amendment of 2026-11-17.) The
two Lean-only obligations — the O1 atom factorisation and the tropical (B′)
sighting — stay `ungraded` until an Agda statement of them exists at all; tier-2
membership has its Agda witnesses already and stays `ungraded` only until the
exhaustive step joins them. The census at the head of this section is the
authoritative split.
The amendment to ADR 0005 D4 implementing this is tracked separately; until it
lands, the ADR and the policy disagree on paper and this paragraph is the
operative reading.
ADR 0005 records this ruling, and the 2026-11-17 one, in its amendment of
2026-11-17 (D4 and D5, the §2 recheck legs).

One measurable cost of the dependency, worth keeping in view: `Trope.nine_families`
is proved by `rfl` yet reports `[propext, Classical.choice, Quot.sound]`, reaching
Expand Down
Loading