From a03ae6645da9945e422efb9a287438fab3cd5004 Mon Sep 17 00:00:00 2001 From: Claude Date: Sun, 15 Nov 2026 09:58:01 +0000 Subject: [PATCH 1/2] docs: correct stale project metadata and the changelog The Sonar project key still named zigprove. The citation abstract claimed more than the README. The changelog had no entry for the 2026-11-07 repair, the u8 pilot or the finite-model replay. Co-Authored-By: Claude Claude-Session: https://claude.ai/code/session_0136P7Zmp1i8PGVHkq3u2iNz --- CHANGELOG.adoc | 13 +++++++++++++ CITATION.cff | 13 +------------ sonar-project.properties | 4 ++-- 3 files changed, 16 insertions(+), 14 deletions(-) diff --git a/CHANGELOG.adoc b/CHANGELOG.adoc index d9f8ce70..8231fb4d 100644 --- a/CHANGELOG.adoc +++ b/CHANGELOG.adoc @@ -53,9 +53,22 @@ to end, plus the plan for the phases after it. source "GitHub Actions"; `pages.yml` lost its `push` and `workflow_dispatch` triggers. The helper's old `publish` route is no longer used by any workflow and is left in place for a follow-up removal. +* Soundness repair (2026-11-07): the gate clears valid writable outputs before + wrapper-level errors, and its tests start with poisoned bytes. The 14-byte + field extent within a 16-byte struct is corrected, and the invalid foreign + binding is removed. Experimental: the evidence report records observed + results only. ==== Added +* Source-bound u8 pilot (experimental): `source-vc` and `source-coq` lower a + closed Zig u8 fragment. The generated Coq theorem covers all 256 inputs. + Only the tiny closed fragment is covered; the general Zig subset is not. +* Finite-model replay (experimental): the generated Coq proof runs through + ECHIDNA `prove --output json`, an independent `coqchk` pass and a + no-axioms audit. The receipt status is always `INCOMPLETE`. Compiler, + refinement and full Dedukti/Coq/Lean4 chain arguments remain open. + * **Chain minting: a minted repo can be minted again.** "Use this template" on a minted repo now mints correctly (owner ruling 2026-11-04, after the Codex review of #251 showed the grandchild inherited its parent's identity). diff --git a/CITATION.cff b/CITATION.cff index 12a9ee72..69db5dfb 100644 --- a/CITATION.cff +++ b/CITATION.cff @@ -5,18 +5,7 @@ authors: given-names: "Jonathan D.A." website: "https://github.com/hyperpolymath" title: "Atoma" -abstract: >- - Atoma is a verification toolchain for Zig at the Idris2 ABI seam. It takes the - contract a repository already declares in Idris2 and the contract its Zig - export fn already documents, canonicalises both into one hash-checked Core - Spec IR, checks the implementation against a deliberately small verifiable - subset of Zig, and derives SMT verification conditions from the result. The - shape is Ada/SPARK or Rust/Creusot: one name for the subset, the contracts and - the toolchain. Witnesses are intended to be exported as canonical Dedukti and - rechecked by coqchk and lean4checker before any claim is promoted from - INCOMPLETE (Warrant) to COMPLETE (Receipt). Status: Phase 0/1 - the seam, - layout, subset and VC gates run and the soundness theorem typechecks for a - store-free core fragment; the kernel chain is specified, not yet run. +abstract: "Atoma is experimental, source-bound verification for an explicitly closed Zig subset. It provides canonical contracts, a syntax screen, typechecked Idris2 specifications, Agda proofs of a store-free model and a finite u8 source pilot. It is not yet a general sound verifier for Zig. Evidence remains INCOMPLETE (Warrant): finite-model replay runs through ECHIDNA and coqchk, while compiler and refinement arguments and the full kernel chain remain unfinished." version: "0.1.0" date-released: "2026-10-22" url: "https://github.com/hyperpolymath/atoma" diff --git a/sonar-project.properties b/sonar-project.properties index d4d82818..8b28f9b9 100644 --- a/sonar-project.properties +++ b/sonar-project.properties @@ -1,9 +1,9 @@ # SPDX-License-Identifier: MPL-2.0 # SonarQube Cloud (SonarCloud) configuration. -# Project: https://sonarcloud.io/project/overview?id=hyperpolymath_zigprove +# Project: https://sonarcloud.io/project/overview?id=hyperpolymath_atoma # Requires the SONAR_TOKEN repository secret + a SonarCloud project (owner setup). sonar.organization=hyperpolymath -sonar.projectKey=hyperpolymath_zigprove +sonar.projectKey=hyperpolymath_atoma # Analysable surface = shell scripts + any JS/TS the template carries. Idris2, # Zig, Elixir and AffineScript have no SonarCloud analyser; vendored, generated, From d4bb1674a69f0e6125f7a5c0e79a75184ec8cd01 Mon Sep 17 00:00:00 2001 From: Claude Date: Sun, 15 Nov 2026 09:58:02 +0000 Subject: [PATCH 2/2] docs(plan): add the family plan and the execution model FAMILY-ULTRAPLAN.adoc holds the work that Atoma shares with quintessence, Apodeixis.jl, bactrian, canonry and chancel: measured state, aspect evaluation, shared packages F1 to F12, waves W0 to W5, model roles, CI design and the ECHIDNA coprocessor findings. ULTRAPLAN.adoc gains one section that maps stages R0 to R5 to that plan. Existing stages, gates and task IDs do not change. Co-Authored-By: Claude Claude-Session: https://claude.ai/code/session_0136P7Zmp1i8PGVHkq3u2iNz --- docs/plan/FAMILY-ULTRAPLAN.adoc | 326 ++++++++++++++++++++++++++++++++ docs/plan/ULTRAPLAN.adoc | 33 ++++ 2 files changed, 359 insertions(+) create mode 100644 docs/plan/FAMILY-ULTRAPLAN.adoc diff --git a/docs/plan/FAMILY-ULTRAPLAN.adoc b/docs/plan/FAMILY-ULTRAPLAN.adoc new file mode 100644 index 00000000..6370081c --- /dev/null +++ b/docs/plan/FAMILY-ULTRAPLAN.adoc @@ -0,0 +1,326 @@ +// SPDX-License-Identifier: CC-BY-SA-4.0 +// SPDX-FileCopyrightText: 2026 Jonathan D.A. Jewell += Verifier family: shared plan to v1.0.0 +:revdate: 2026-11-15 +:toc: +:toclevels: 2 + +This plan covers six sibling verifier repositories. It holds the work that is +the same in more than one of them. Each repository keeps its own plan for the +work that only it can do, at `docs/plan/ULTRAPLAN.adoc`. + +[cols="1,1,3",options="header"] +|=== +| Repository | Target | Role in the family +| `atoma` | Zig | Pathfinder. It has the strongest evidence discipline today. +| `quintessence` | Elixir/OTP | Most code. Weakest evidence integrity today. +| `Apodeixis.jl` | Julia | Source-policy audit only. Clean tests. +| `bactrian` | OCaml, BEAM actors, FFI | Design documents and a readiness preflight. +| `canonry` | Agda theory for Chapel arrays | One checked theory module. +| `chancel` | Chapel front end (Rust) | Placeholder crate. +|=== + +This file lives in `atoma` because `docs/decisions/ATOMA-SIBLING-TRANSFER.adoc` +already makes `atoma` the first slice. The owner can move it. The repositories +stay separate. Nothing here merges them. + +== Status vocabulary + +A statement in this plan is *measured* only if a command ran on 2026-11-15 and +the result was seen. All other statements are *read* (from source) or *planned*. +Estimates are judgments. They are not measured velocity. + +== Measured state on 2026-11-15 + +Each row ran in one cloud session. Tool versions: Julia 1.12.1, Zig 0.15.2 and +0.17.1, Agda 2.6.3 with standard library 1.7.3, Z3 4.8.12, cvc5 1.1.2, Rust +stable. + +[cols="1,3,2",options="header"] +|=== +| Repository | Command and result | Not run +| `Apodeixis.jl` | `Pkg.test()`: 366 passed, 0 failed (347 before this session, plus 19 new). | Julia 1.9.4 leg. Idris2 and Zig scaffold. +| `atoma` | `zig build test`: 24 of 24 passed. `verification/atoma/check-agda.sh`: 4 modules, 0 postulates, 0 holes. | `scripts/atoma-gate.sh` (needs Z3 4.13.3, Idris2, ECHIDNA build). +| `quintessence` | `zig build test`: 10 of 10 passed. `tests/quintessence/run_suite.sh`: 14 passed, 2 failed. | Elixir front end. Suite with the solver versions that the README names. +| `bactrian` | `tests/bactrian/doctor_test.sh`: PASS. `scripts/bactrian/doctor.sh`: exit 2 (blocked, by design). | No product code exists to run. +| `canonry` | `scripts/check-proofs.sh agda`: 2 of 2 gated modules PASS under `agda --safe --without-K --cubical-compatible`. | Rocq and Lean legs do not exist. +| `chancel` | `cargo test --locked`: 2 passed. `cargo clippy -D warnings` and `cargo fmt --check`: clean. | No verifier code exists to run. +|=== + +== Aspect evaluation + +One row for each aspect. `ok` means measured and sound for the scope that the +repository claims. `gap` means a defect or a missing part. `none` means the +part does not exist yet. + +[cols="2,2,2,2,2,2,2",options="header"] +|=== +| Aspect | atoma | quintessence | Apodeixis.jl | bactrian | canonry | chancel +| Claims match code | ok | gap: fixed in README today; receipt format still overclaims | ok | ok after README repair | ok after README repair | ok after wording repair +| Soundness of results | ok for the u8 pilot; model proof only | gap: structural Alethe check, heuristic Layer C, fixed BMC model | gap closed today: deny-list held 7 bypasses | none | ok: 0 postulates | none +| Source to proof link | partial: one closed u8 function | gap: example-specific special cases | none | none | none: no Chapel link | none +| Independent replay | partial: ECHIDNA and coqchk, receipt INCOMPLETE | none: no ECHIDNA call | none | none | none: no Rocq or Lean leg | none +| Falsifiers | ok: 8 groups | partial: 16 cases, 2 fail off the pinned solvers | ok: tests fail without the fix | ok for the preflight | none | none +| Project CI gate | ok: `atoma.yml` | gap: no workflow runs the engine | ok: `julia.yml` | added today: `bactrian.yml` | added today: `canonry.yml` | ok: `rust-ci.yml` +| Reproducible toolchain | partial: pins, runner not frozen | gap: Zig 0.17 in code, 0.16.0 in `mise.lock`; solvers not pinned; NIF hash does not reproduce | partial: no `Manifest.toml`, `julia = "latest"` | gap: 8 authorities UNPINNED | partial: apt pins | ok: `Cargo.lock`, no dependencies +| Project benchmarks | none: template only | none: template only | none: template only | none | none | none +| Proof grade (theory) | ungraded: no receipt | none | none | none | `proven`, ungraded: no receipt | none +| Deed identity | set | set, no uuid | UNASSIGNED | UNASSIGNED | UNASSIGNED | UNASSIGNED +| Plan to v1 | yes | two plans that disagree | detailed, no v1 gate | detailed, no v1 gate | roadmap only | none +|=== + +== What is common and what is bespoke + +=== Common: do it one time, then apply it six times + +[cols="1,4,2",options="header"] +|=== +| ID | Work | Kind of sameness +| F1 | Template truth-up at the source (`rsr-template-repo`), then one sync of all six. Three template revisions are in use now. See the defect list below. | Identical files +| F2 | Deed identity fields: IS, IS-NOT, pipeline position, clade. Owner only. | Same form, different values +| F3 | Evidence kit: one receipt schema, one status vocabulary, one replay adapter, one falsifier harness. Start from `scripts/atoma-replay.sh` and `scripts/atoma-gate-falsifiers.sh`. | Same thing, change of name +| F4 | CI gate shape: `.yml`, `scripts/-gate.sh`, `scripts/-ci-toolchain.sh`, evidence artifact, required check. | Same thing, change of name +| F5 | Project benchmarks in the estate bench format. They replace `benches/template_bench.sh`. | Same harness, different workload +| F6 | One toolchain pin table for Zig, Z3, cvc5, Agda, Coq, Idris2 and ECHIDNA. | Identical values +| F7 | ECHIDNA client contract: `echidna prove --output json`, schema `echidna.prove.result/1`, one pinned ECHIDNA revision. | Identical adapter +| F8 | Zig seam. `atoma` checks the Zig code that `quintessence` (NIF) and `bactrian` (FFI) admit. | One provider, two consumers +| F9 | BEAM actor protocol layer. `quintessence` Layer B and `bactrian` Layer B use the same K-CUT rules, occupancy bound and upstream Agda authorities. | Same semantics, two front ends +| F10 | Coprocessor contract with ECHIDNA. See the section below. | One contract +| F11 | Proof Witness Grade recipe for theory: receipt for `PWG-3`, then a Rocq leg and a Lean leg. Applies to `canonry` and to the `atoma` Agda modules. | Same recipe +| F12 | Release engineering: changelog, citation, version tags, the standards V1-GATE audit. | Same checklist +|=== + +=== Bespoke: each repository must do it alone + +* The front end: parser or compiler adapter for the target language. +* The closed subset and its exclusions. +* The semantics: memory and aliasing (Zig), dispatch and numeric tiers (Julia), + actors and supervision (Elixir), modules and effects (OCaml), strided domains + and array loops (Chapel). +* The refinement argument from admitted source to the proved semantics. +* The pilots and the diagnostics for users of that language. + +`ATOMA-SIBLING-TRANSFER.adoc` forbids a shared semantic IR until two adapters +show the same meaning. This plan keeps that rule. + +=== Template defects found in more than one repository (input to F1) + +Fix each one in the template. Do not patch the six copies. + +* `docs/status/TEST-NEEDS.adoc` claims "CRG Grade C, achieved 2026-04-04" in + repositories born in October and November 2026. +* `docs/status/READINESS.adoc` ships unfilled `_<...>_` placeholders. +* `docs/status/PROOF-NEEDS.adoc` names a file on the owner's desktop. +* `docs/EXPLAINME.adoc` is titled "RSR Template Repo" in every child. +* `SUPPORT.adoc` uses Markdown syntax. +* `scripts/check-bench-regression.sh` names `benchmarks/`; the directory is `benches/`. +* `benches/template_bench.sh` measures template scripts in every child. +* `.gitignore` ignores `verification/proofs/agda/*.agdai` and misses nested modules. +* The deed carries `:last-updated "2026-04-11"`, before the birth date of the child. +* The deed sets `:port 0` and the mint answers set `:port "8080"`. +* `docs/decisions/` has two files numbered 0001. +* `pages.yml` runs `cabal build` in repositories with no cabal file. +* `rust-ci.yml` is green by no-op where no `Cargo.toml` exists. +* `docker-build.yml` triggers on `build/container/**` where that directory is absent. +* `sonar-project.properties` and `sonarqube.yml` keep an old project key. +* `Justfile` recipes `readiness`, `build-release` and `test-smoke` are TODO stubs that fail. +* The sample proofs under `verification/proofs/` say "Replace with your project's proofs" and stay gated. + +== Definition of v1.0.0 + +`atoma/docs/plan/ULTRAPLAN.adoc` defines seven conjunctive gates. Every +verifier in the family uses the same seven. Only the acceptance text changes. + +[cols="1,4",options="header"] +|=== +| Gate | Family rule +| Language boundary | A closed, typed subset with stated exclusions. An unsupported construct fails closed. +| Source connection | Obligations come from the body. A change to a body, an import or a target invalidates old evidence. +| Semantics | The semantic rules that the subset needs are proved, and the trusted computing base is published. +| Replay | A source-derived witness passes every kernel that the specification requires. Unknown, timeout, missing and malformed evidence fail closed. +| Practical coverage | Three pilots. A second consumer needs no special case. +| Assurance | Mutants are killed by gates that ran. Evidence is reproducible and bound to source. +| Release and estate | Claims, deed, README and installation are true. The standards V1-GATE audit passes. +|=== + +`canonry` is a theory library, not a verifier. Its v1 gates are in its own +plan. A repository can release `0.x` versions before it passes all gates. + +== Waves: how the six move forward together + +A wave closes when each repository has its exit evidence. A repository that +cannot close a wave narrows its subset. It does not skip the wave. + +[cols="1,3,3",options="header"] +|=== +| Wave | Shared work | Exit evidence in each repository +| W0 Truth and wiring | F1, F2, F4 (first half) | README is true. One project gate runs in CI. Each known hole is closed or is an open obligation in `PROOF-STATUS.adoc`. +| W1 Evidence spine | F3, F6, F7, F5 | The gate uses the evidence kit without a local fork. One benchmark has a baseline from a real run. +| W2 First source-bound obligation | none (bespoke front ends on the shared kit) | One closed function goes from source to obligation to ECHIDNA replay. One body-only mutant is rejected. Receipt status is INCOMPLETE. +| W3 Subset and semantics | F8, F9 | The checked argument covers the admitted subset. An alias, overflow or nontermination counterexample is rejected. +| W4 Kernel chain and grades | F7 (exchange), F10, F11 | End-to-end replay. Adversarial certificates fail. Theory theorems have receipts. +| W5 Pilots and release | F12 | All seven gates pass on the release commit. +|=== + +W0 is partly done (see "Done in the first session"). W2 is the first wave +that tests the transfer claim: measure time to a valid proof, time to reject a +body-only mutant, adapter size, and the count of kit components reused +unchanged. Re-estimate W3 to W5 from those numbers. + +== Models, effort and cost + +=== Roles + +[cols="1,3,3",options="header"] +|=== +| Model | Use it for | Do not use it for +| Fable (advisor) | Plan ratification. Wave gate review. Review of each claim promotion. Review of a cross-repository contract. | Execution. Routine review. +| Opus (lead) | Proof work. Semantics. Refinement arguments. Contract and schema design. Repair of a soundness hole. Review of Sonnet work on a trust path. | CI, template drift, status files (doctrine rule 16). +| Sonnet (executor) | Implementation against a written specification, 3 to 20 files. CI workflows. Adapters. Tests. README and status documents. | Design of a trust rule. +| Haiku (bulk) | Scout pass before each step. Mechanical edits. Template sync. Counts and stale-string sweeps. | Any edit that needs judgment about a claim. +|=== + +=== Orchestration rules + +. Opus leads one wave in one repository at a time. It writes the brief for each Sonnet or Haiku task. +. One writer for each repository tree at a time. Each brief names one allowed tree. +. A Haiku scout runs before each Sonnet or Opus step. It reports blockers in 20 lines or fewer. +. The advisor is called at four points only: plan ratification, the close of each wave, each promotion of a status word, and each change to a shared contract. The input is evidence files, not a narrative. +. A subagent never states that a check passed unless it ran the check. The lead reruns one check from each report. +. Work that touches a status word, a receipt, a trust level or a proof goes to Opus, with advisor review. +. Common work (F1 to F12) is done in one repository first. Sonnet or Haiku then replays it in the others from a replay brief. +. Shared code is extracted only after two repositories use the same thing. + +=== Effort scale + +`S` is one agent session. `M` is 2 to 4 sessions. `L` is 5 to 10 sessions. +`XL` is more than 10 sessions and must be split before it starts. Measured in +the first session: a recon of one repository used about 180,000 to 260,000 +tokens on Sonnet or Haiku. A polish task used about 160,000 to 180,000 tokens. + +[cols="1,4,1,1,1",options="header"] +|=== +| ID | Work | Lead | Workers | Size +| F1 | Template truth-up and one sync | Sonnet | Haiku | M +| F2 | Deed identity | Owner | Haiku to regenerate the arrival pack | S +| F3 | Evidence kit: schema, adapter, falsifier harness | Opus, advisor review | Sonnet | L +| F4 | CI gate shape in each repository | Sonnet | Haiku | M +| F5 | Benchmarks | Sonnet | Haiku | M +| F6 | Toolchain pin table | Sonnet | Haiku | S +| F7 | ECHIDNA client contract and exchange | Opus | Sonnet | L +| F8 | Zig seam through `atoma` | Opus | Sonnet | L +| F9 | Actor protocol layer | Opus, advisor review | Sonnet | XL +| F10 | Coprocessor contract | Opus, advisor review | Sonnet | M +| F11 | Proof Witness Grade recipe and legs | Opus | Sonnet for receipts | L +| F12 | Release engineering | Sonnet | Haiku | M +|=== + +The bespoke work is much larger than the shared work. The existing plans give +30 to 60 person-weeks for `atoma` and about 160 person-weeks for +`quintessence`. Those figures are the authors' judgments. This plan does not +shorten them. It removes repeated work and it orders the risk. + +== CI/CD design (F4, F5) + +Every repository has one project workflow with the same shape. `atoma.yml` is +the reference. + +. `actions/checkout`, no persisted credentials, read-only permissions. +. `scripts/-ci-toolchain.sh`: installs pinned tools and checks each digest. +. `scripts/-gate.sh`: one fail-closed entry point. A missing tool is a failure, not a skip. +. The gate runs unit tests, project proofs, falsifiers, then the benchmark. +. The evidence directory is uploaded even on failure. +. The job is a required check on `main`. +. After a workflow change, run `scripts/lock-sync-fix.sh`, then `scripts/check-lock-sync.sh`. Do not edit `actions.lock` by hand. + +[cols="1,2,3",options="header"] +|=== +| Repository | Gate today | Next step +| `atoma` | `atoma.yml`: full gate | Add the replay time and the mutant-rejection time as benchmarks. +| `Apodeixis.jl` | `julia.yml`: package tests on two Julia versions | Add audit throughput on a fixed corpus as a benchmark. +| `quintessence` | none | Add `quintessence.yml` with the Zig unit tests first. Add the suite only after the solvers are pinned and the NIF receipt reproduces. +| `bactrian` | `bactrian.yml`: preflight test | Flip the expected preflight status when WP-00 pins the authorities. +| `canonry` | `canonry.yml`: Agda proof gate | Add the receipt step (F11). +| `chancel` | `rust-ci.yml` | Add the subset-gate fixtures when the gate exists. +|=== + +A benchmark is valid only if its baseline comes from a recorded run on the +pinned toolchain. A benchmark gate compares against that baseline with the +estate script `scripts/check-bench-regression.sh`. + +== ECHIDNA coprocessor (F10) + +Question from the owner: are these repositories wired to the ECHIDNA +coprocessor layer and its Julia side, `src/julia/coprocessor.jl`, and can that +be done without friction? + +Answer: no repository in the family refers to the coprocessor. The wiring is +not possible without friction today. These are the facts, read from the +ECHIDNA source on 2026-11-15: + +* The Rust coprocessor layer has 16 kinds. The in-process backends are real code. +* No external entry point exists. The CLI, REST, gRPC and GraphQL interfaces have no coprocessor route. +* `coprocessor.jl` implements 6 `Math` operations. No recipe, container or Julia file starts it. Rust code never constructs the bridge. No round-trip test exists. +* The coprocessor default URL and the ML sidecar default URL both use port 8090. +* The wire format has no version field and no published schema. +* The trust tier of a coprocessor is defined. No code outside the module reads it. +* The ECHIDNA plan (decision D-G) defers the coprocessor trait "until a second consumer needs it". + +The verifier family is that second consumer. The path, in order: + +. The owner rules on D-G. +. ECHIDNA publishes a versioned contract, `echidna.coprocessor/1`, in the style of `PROVE-RESULT-CONTRACT.adoc`. +. ECHIDNA adds one entry point, for example `echidna coprocessor dispatch --output json`, and a capability probe. +. Trust rule: a coprocessor answer is a hint. It never raises a status word. A prover or an exhaustive check must confirm it. +. The Julia server gets its own port, a launcher and a round-trip test. + +First uses that pay for the work: + +* `canonry` and `chancel`: exact `gcd` and `lcm` for strided-domain alignment. Use them to search for counterexamples before a proof attempt. +* `Apodeixis.jl`: exact integer and rational cross-checks for the exact numeric tier. Interval arithmetic needs new operations. +* `atoma` and `quintessence`: evaluation of bit-vector counterexamples. + +Until the contract exists, each verifier calls ECHIDNA only through +`echidna prove --output json`. + +== Shared libraries + +[cols="2,1,3",options="header"] +|=== +| Candidate | Verdict | Reason +| Evidence kit (schema, replay adapter, falsifier harness) | Extract in W1 | It holds no language semantics. `ATOMA-SIBLING-TRANSFER.adoc` rates its reuse high. +| Canonical JSON and hashing | Use `hyperpolymath/ijson-jcs` | `atoma` (`canon.zig`) and `quintessence` (`jcs.zig`) each carry a private copy. +| SMT driver and Alethe check | Do not share the `quintessence` copy | Its Alethe check is structural. Use the ECHIDNA dispatch and certificate path (F7). +| Mutation engine | Use the ECHIDNA mutation planner | The `quintessence` engine is a local imitation with the ECHIDNA name. +| Finite arithmetic and linear resources in Agda (`Bv.agda`, `Lin.agda`) | Extract when a second repository needs them | The second-adopter rule. +| Shared semantic IR | Not yet | Forbidden until two adapters show the same meaning. +|=== + +The home of the evidence kit is an owner decision. The options are a new +small repository, a directory in ECHIDNA next to the prove contract, or a +directory that the template sync carries. + +== Enhancements that the lead recommends + +. Make `quintessence` consume `atoma` for Layer C. It replaces a pattern heuristic with a checked result, and it gives `atoma` its second consumer. +. Pair `canonry` and `chancel` under one wave schedule. `chancel` has no purpose until `canonry` M1 defines the obligation format. +. Put `bactrian` Layer B behind `quintessence` Layer B. Both need the same protocol rules. Build them one time. +. Record the reuse measures from `ATOMA-SIBLING-TRANSFER.adoc` in each W2 exit report. Decide the order of W3 from them. +. Give every repository the same plan path, `docs/plan/ULTRAPLAN.adoc`, and the same seven gates. An agent that learns one repository can then work in the others. +. Keep one page for each repository that lists the measured state with the date. Refresh it at each wave close. +. Unify the status words. `quintessence` uses `TrustLevel`. The estate grades theory with `PWG`. Receipts use INCOMPLETE and COMPLETE. The evidence kit must map them one time. + +== Done in the first session (2026-11-15) + +* `Apodeixis.jl`: the policy audit accepted `@ccall`, `unsafe_wrap`, `pointer_from_objref`, `include`, `llvmcall`, an alias of a forbidden call and `getfield(Base, :unsafe_load)`. It now rejects them. 19 new tests fail without the fix. +* `quintessence`: a `coqc` and a `lean` binary on PATH gave TrustLevel 5 with no check. The maximum is now 4. Suite case 16 fails without the fix. The README states the measured state and the limits. +* `canonry`: `canonry.yml` runs the Agda proof gate. The README describes the theory. +* `bactrian`: `bactrian.yml` runs the preflight test. The README describes the project. +* `chancel`, `atoma`: stale statements corrected. +* Each repository has a plan at `docs/plan/ULTRAPLAN.adoc`. + +== Stop points + +* If a W2 spike fails in a language, narrow the advertised subset or change the architecture. Do not add features. +* If the evidence kit needs a language-specific branch, stop F3 and review the schema with the advisor. +* If a status word would be promoted without a receipt, stop. A grade is derived from receipts. It is never asserted. diff --git a/docs/plan/ULTRAPLAN.adoc b/docs/plan/ULTRAPLAN.adoc index f895e4bd..f86b700a 100644 --- a/docs/plan/ULTRAPLAN.adoc +++ b/docs/plan/ULTRAPLAN.adoc @@ -157,3 +157,36 @@ the scaffold aspect script does not demonstrate coverage of these paths. Exercise sequences across aspects, not isolated checklists. Record reusable lessons in `docs/decisions/ATOMA-SIBLING-TRANSFER.adoc`; extract a shared engine only after a second language demonstrates the same semantic boundary. + +== Execution model and family links + +Added 2026-11-15. The stages, gates and task IDs above do not change. + +The work that Atoma shares with its sibling verifiers (`quintessence`, +`Apodeixis.jl`, `bactrian`, `canonry`, `chancel`) is in +link:FAMILY-ULTRAPLAN.adoc[the family plan]. That plan holds the shared work +packages F1 to F12, the waves W0 to W5, the model roles and the effort scale. +Atoma is the pathfinder: each shared package is done here first. + +Measured on 2026-11-15: `zig build test` in `tools/atoma` gave 24 of 24 passed +(Zig 0.15.2). `verification/atoma/check-agda.sh` gave 4 modules, 0 postulates +and 0 holes (Agda 2.6.3). `scripts/atoma-gate.sh` was not run in that session; +the hosted workflow is the evidence for it. + +[cols="1,2,1,1,3",options="header"] +|=== +| Stage | Family wave and packages | Lead | Size | Notes +| R0 | W0 | Sonnet | S | Done, apart from the inherited template drift (F1). +| R1 | W1 and W2: F3, F6, F7 | Opus | L | Extract the evidence kit from the replay scripts here. Then finish the alias counterexample and the typed boundary. +| R2 | W2 and W3 | Opus | XL (split first) | Tasks S1 to S6 and T6. Sonnet implements from Opus briefs. +| R3 | W3: F8 | Opus, advisor review | XL (split first) | Tasks T1 to T4 and T7. Proof work. `quintessence` and `bactrian` become consumers of the Zig seam. +| R4 | W4: F7 exchange, F11 | Opus | L | Tasks K1 to K9. The Agda modules get receipts, then the Rocq and Lean legs. +| R5 | W5: F12 | Opus for pilots; Sonnet for release | L | Tasks H1 to H6. +|=== + +Size `S` is one agent session, `M` is 2 to 4, `L` is 5 to 10, and `XL` is +more than 10 and must be split before it starts. Haiku runs the scout pass +before each step and does the mechanical edits. The advisor reviews the close +of each stage and each change to a status word. The person-week ranges in the +sequence table above stay the planning basis; the sizes here only order the +agent work.