SG-0 by briansrls · Pull Request #559 · gunb-ai/gunbc

briansrls · 2026-04-19T07:10:27Z

Opened from session-dashboard for session tidy-tern-93.

briansrls · 2026-04-19T07:12:13Z

This is the right lane and the right kind of work for SG-0. One blocking issue before it can serve as the ratchet for the rest of SG:

EXPECTED_HAND_AUTHORED is seeded with paths that do not exist in the current tree. I checked the actual src/v3/compiler/**/*.rs surface, and these entries are stale / nonexistent:
- src/v3/compiler/src/dimension.rs
- src/v3/compiler/src/bin/self_host_fixed_point.rs
- src/v3/compiler/tests/determinism_test.rs
- src/v3/compiler/tests/integration/common/determinism_fixtures.rs
- src/v3/compiler/tests/lane2_stage_2f_dimension_test.rs
A census ratchet has to start from an exact current inventory or it becomes noise immediately. Please reseed EXPECTED_HAND_AUTHORED from the real tree on the PR branch (including the new sg0_census_test.rs itself, excluding generated-header files) before landing this.

Non-blocking refinement: the prose in dsl/gunbc/compiler.dag says "No dual-authority period." Per the agreed SG rule, that should be clarified to mean no dual authority in merged/main state; in-PR intermediate commits can temporarily hold both old and new authorities during parity work.

briansrls · 2026-04-19T07:15:41Z

ChatGPT review in progress... (view conversation)

Check back in ~30 minutes for the full review.

Prose refinement from PR #559 review: "No dual-authority period" was ambiguous about whether it bound the merged state or every intermediate commit. Clarify that it only bites at merge time; in-PR commits may temporarily hold both authorities while parity work is in flight. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

briansrls · 2026-04-19T07:27:41Z

ChatGPT review in progress... (view conversation)

Check back in ~30 minutes for the full review.

briansrls

codex · gpt-5.4 · c30e73d1

✅ Review (blocking: 0, non-blocking: 1+/0-)

Non-blocking — Strengths

src/v3/compiler/tests/integration/sg0_census_test.rs The ratchet is narrowly scoped to src/v3/compiler, and the inline receipt makes the temporary authority split and self-hosting dissolution trigger explicit.

ROADMAP — Verified

Compiler-source ratchet enforcement path: This SG-0 census matches ROADMAP's temporary src/v3/compiler ratchet pattern rather than silently normalizing dual authority in merged/main.

✅ I do not see a blocking modeling or correctness issue in the changed lines.

briansrls · 2026-04-19T08:08:16Z

Meta-review in progress... (view conversation)

Loop-health check: is this review cycle making forward progress, or shifting debt? Posts in ~5-15 minutes.

briansrls

codex · gpt-5.4 · c30e73d1

⚠️ Review (blocking: 1, non-blocking: 0+/0-)

BLOCKING (1)

Root Cause

src/v3/compiler/tests/integration/sg0_census_test.rs Generated-file provenance is modeled as a spoofable header string instead of a producer-owned authority; derive the generated set from an enforced generator receipt/manifest or equivalent source of truth rather than comment-prefix matching.

⚠️ The SG-0 ratchet direction is right and the prior seeding issue is fixed, but the generated-versus-hand-authored partition is still unsound enough that I would tighten it before this becomes the lane-0 authority gate.

Sign in to view

+    let Some(first) = contents.lines().find(|line| !line.trim().is_empty()) else {
+        return false;
+    };
+    first.trim_start().starts_with(GENERATED_MARKER_PREFIX)


briansrls · 2026-04-19T08:33:24Z

Meta-Review (Loop Health)

_{Generated by gpt-5-4-pro}

According to a document from 2026-04-19, this loop is 🔁 PAUSE_AND_REGROUP.

Loop summary. I can substantiate 3 completed review rounds over about 51 minutes: two browser reviews completed at 07:21:54 and 07:33:00, and one codex review completed at 08:06:20. That is 2 browser reviews, 1 codex review. The uploads expose only one visible current diff snapshot, so the total PR commit count is not recoverable from the provided artifacts.

Forward progress evidence. There is one real gain: SG-0 adds a compiler-source ratchet for v3 Rust authority, and codex correctly recognized that it matches the ROADMAP’s explicitly allowed temporary compiler-source ratchet pattern for hand-written compiler code until compiler.dag self-hosts. The broader roadmap also already has real self-hosting infrastructure in place through determinism tests and the self_host_fixed_point binary. So this PR is not random churn; it fits an existing migration discipline.

Debt accumulation evidence. The problem is that the loop is no longer reviewing the same object. The current diff is only the SG-0 census/ratchet change: source-authority inventory in compiler.dag plus sg0_census_test.rs and its registration in integration.rs.pr-559.diff But both browser reviews are still auditing older operator/scaffold work — TransformTarget, OperatorKind, OPERATOR_FIELD_MAP, primitive caches, Dag::int_shape() — which is not in the current diff. Only the codex review is actually grounded in SG-0. That means 2 of the 3 reviews are stale relative to the present change, so the loop is burning cycles on history, not on the current PR.

The second debt signal is that SG-0 mostly measures debt instead of paying it down. ROADMAP’s own example of real progress, PR-B, is framed in terms of a downstream consumer actually validating substrate sufficiency. SG-0 does not do that; it adds a census around the hand-authored Rust surface. The new compiler.dag “authority inventory” has no real consumer yet; enforcement still lives in a Rust test constant and a hand-maintained roster in the diff itself. That is substrate-shaped accounting without consumer validation.pr-559.diff

The third debt signal is that the real recurring class is still somewhere else. ROADMAP still carries ArrowBody::Unparsed’s DB-14 interim as a scheduled deletion blocked on E-9, and INVARIANTS.md says E-9 exists because browser review, codex, and an earlier meta-review all independently found the same split-authority shape. That means the recurring finding has already been identified and graduated. SG-0 is not closing it; the reviews are just re-mentioning it while looking at the wrong diff.

Cheating signal. The implementer is not hiding the compromise. The SG-0 comments are explicit that this is a temporary authority split with a self-hosting dissolution story. But the metric is also easy to game: the census classifies files by a first-line // AUTO-GENERATED marker and allows retirement by moving files behind generated headers or narrow shims in the current diff.pr-559.diff That is exactly the kind of “small blast radius” accounting move that becomes attractive late in a loop. And the testing invariant explicitly says source-audit tests should anchor on live syntax or declarations, not comments, and should not collapse into mirror-the-implementation checks. So the compromise is documented, but it is still convention-level accounting, not structural enforcement.

Path to convergence. Before another implementation round, the project needs one explicit decision artifact: a short SG authority contract that answers three questions.

First: is SG-0 a pure Rust-side temporary ratchet, or is it a real compiler.dag authority?

If it is a temporary Rust ratchet, then drop the unused compiler.dag inventory for now and say so plainly.

If it is real authority, then make the test consume compiler.dag now instead of mirroring its constants and the roster in Rust.

Second: what counts as “authority retired”?

Not “one fewer file.” Define whether retirement means “generated-from-.dag only” or whether a bounded host shim is allowed, and if allowed, define that shim shape.

Third: replace the comment-marker protocol or formally bound it.

Right now SG-0 uses comment text as the generated/non-generated discriminator. Either replace that with structural receipt metadata, or record a narrow temporary exception against the source-audit testing rule with a named dissolution trigger.

Only after that reset is another round worthwhile. Without it, the next round will keep doing what this one did: stale substrate commentary on top of a file-count ratchet.

Meta-verdict — 🔁 PAUSE_AND_REGROUP

This loop is not making clean forward progress on PR #559. It has drifted off the current diff, the new “authority” facts do not yet have a real consumer, and the ratchet it adds is convention-based and gameable. Reset the object of review and the SG authority boundary before iterating again.

View conversation

briansrls · 2026-04-19T15:52:36Z

claude-review (director) — Aligning with codex BLOCKING + ChatGPT
PAUSE_AND_REGROUP. The ratchet must be sound before any subsequent SG PR
can claim its count-decreased correctly.

Proposed fix — build-time producer manifest.

Replace header-comment matching with a producer-owned manifest:

Each regen_* binary writes its output paths to a manifest (e.g.,
target/v3_generated_files.manifest or ${OUT_DIR}/generated.manifest)
when it runs.
build.rs writes the same manifest entries for any files it generates
via include_str! projection (build-time codegen).
sg0_census_test.rs reads from the manifest, NOT from file contents.
Hand-authored count = enumerate(src/v3/compiler/src/**) minus
read_manifest().

Why this is sound:

Producer-owned: only codegen authorities can add to the manifest. A
contributor adding a // Generated by ... comment to a handwritten
file doesn't appear in the manifest because no producer registered it.
Fail-closed at the boundary per feedback_fail_closed_is_boundary:
the manifest IS the authority for "what was generated," not a string
in the file body.
Single authority per feedback_substrate_principle_audit Q5: the
producer is the sole source of truth.

Acceptance: ratchet test reads producer manifest; adding a fake
// Generated by ... header to a handwritten test file doesn't pass the
ratchet (test should fail). That's the soundness assertion.

If your earlier proposal differs (e.g., .dag-declared
data generated_files: GeneratedFileSet), share so we can reconcile —
that direction is also sound but harder to bootstrap (chicken-and-egg
with codegen needing to load .dag before producing). Build-time manifest
is simpler.

This is load-bearing for the SG program. Land sound; then SG-6 and the
rest can be trusted.

Dependency: none — self-contained. You can proceed today.

briansrls · 2026-04-19T16:01:57Z

Meta-review in progress... (view conversation)

Loop-health check: is this review cycle making forward progress, or shifting debt? Posts in ~5-15 minutes.

briansrls · 2026-04-19T16:02:39Z

ChatGPT review in progress... (view conversation)

Check back in ~30 minutes for the full review.

Hardens the SG-0 census ratchet by replacing content-marker classification with a producer-owned manifest, addressing the Codex BLOCKING review and the ChatGPT meta-review's PAUSE_AND_REGROUP verdict on the prior landing (#562). **Partition authority (Q3).** `src/v3/compiler/build.rs` now emits `${OUT_DIR}/v3_generated_files.rs` with a `GENERATED_FILES` static derived from a reviewed `REGEN_OUTPUTS` literal. Every codegen driver (4 `regen_lens_*` binaries, the SG-4 prep `regen_infer_helpers`, and the ignored `emit_lens_provenance_snapshot` test) imports the manifest and asserts its output path is in it before writing — producer identity becomes structural. `sg0_census_test.rs` consumes the same manifest: hand-authored = walk(src/v3/compiler) minus manifest membership. No content scanning; a hand-authored `// AUTO-GENERATED` header does not slip through. **Authority location (Q1).** Dropped the unused `V3CensusRoot`, `v3_census_roots`, and `v3_generated_marker_prefix` from `dsl/gunbc/compiler.dag` — the test never consumed them (parallel-representation debt per `feedback_parallel_representation_debt`). Kept the SG program narrative (genuine documentation) and added a footnote naming SG-6 as the migration trigger that promotes the manifest from build.rs into .dag data. **Shim rule (Q2).** Tightened the definition in compiler.dag: a retired-to-shim .rs's ONLY hand-authored content is imports, a `#[path]`/`include!` directive, or `pub use` re-exports. Any function body, match arm, or control-flow statement keeps the file counted as hand-authored. Lens-checkable eventually; review-enforced today. **New soundness tests.** - `sg0_generated_partition_is_producer_owned` — proves a path outside the manifest classifies as hand-authored regardless of file-local content. - `sg0_every_generated_file_is_present_on_disk` — keeps the manifest in lockstep with committed producer output. Rebased on main (cf37494). The SG-4-prep-a additions (`regen_infer_helpers.rs`, `sg4_prep_infer_helpers_freshness_test.rs`, `infer_helpers_generated.rs`) are integrated: the generated file is in REGEN_OUTPUTS, the regen driver asserts membership, and the two new hand-authored entries live in EXPECTED_HAND_AUTHORED with a receipt pointing at SG-6 as the dissolution trigger. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

briansrls · 2026-04-19T16:08:07Z

Meta-Review (Loop Health)

_{Generated by gpt-5-4-pro}

To evaluate the review loop of PR #559 (SG-0) against the modeling principles outlined in the documents you uploaded, we will follow the six key principles and assess whether the PR adheres to them:

1. Fail-closed

This principle ensures that every failure path is diagnosed and does not silently return None or panic on user paths. Upon reviewing the PR, the failures appear to be adequately handled with explicit diagnostics. The failure paths are controlled, and no failure scenario silently returns None. Each failure is processed through a diagnostic message, which is essential for debugging and correctness in the context of the project.