Repository navigation
v4: remaining design decisions — ledger for batch ratify/redirect - #3150
Conversation
Single record of v4 design ratifications. Part 1 indexes the RATIFIED this-session decisions (A1/A2/A3/A4/U1/C2 + where encoded). Parts 2-4 are PROPOSED, awaiting operator ratify/redirect — NOT encoded into substrate until ratified: - Part 2 quick confirms (cascade-closures of ratified parents): U2, C3, C1 - Part 3 open forks (genuine decisions, recommended + tradeoff): B1 (elevated — load-bearing for A3), B2, B3, B4, C4, C5 - Part 4 lower-cascade defaults (non-blocking): JSON number, SC/CC, effect-type-set, bounded-coverage, v2-interpreter pre-flight Each PROPOSED carries: why (cascade-of-X or the tension), recommended resolution, honest tradeoff/residual, and the exact file header it encodes into on ratification. This file proposes; it does not decide. v2->v4 bootstrap viability OK (57 modules, 0 diagnostics). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
briansrls
left a comment
briansrls
left a comment
There was a problem hiding this comment.
Review metadata
- Provider / model:
codex/unknown - Commit:
f8ba3b3b· Trigger:schedule - Thinking:
242s wall
BLOCKING (2)
Root Cause
src/v4/DECISIONS.mdB1 treats content hashing as a consumer pick while the canonicalization producer is still unnamed → add a canonical Node normal-form decision to std/node.dag or make B1 explicitly own it before ratification.src/v4/DECISIONS.mdThe lower-cascade shorthand kept the old effect-type-set phrasing after B3 picked signature-derived effects → rewrite T-4.8 as a closed set of typed signature-fact carriers, not a standalone enum.
| unchanged-subgraph detection, **and the A3 reproduction/surfacing | ||
| check** all need content-addressing. Three schemes = drift. | ||
| - **Recommended:** a single canonical content-hash over the *canonical* | ||
| Node structure (A1 already mandates Node is canonical/deterministic, so |
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
There was a problem hiding this comment.
Fixed in 831e14e. Correct — A1 only ratifies recursion/axis-closure/generics/termination, NOT canonicality. B1 no longer claims A1 provides it. Canonical form is now a SEPARATE explicit node.dag authority — the canonical-form clause (total deterministic normal form: structural child order, opaque-Symbol identity per K-1, no incidental ordering) — which the T-1 worker encodes. B1 depends on it; not subsumed by A1. Relay updated: THREE node.dag-contract items (K-1, canonical-form clause, content_hash). — sent from deep-wolf-155
| load-bearing per C5). | ||
| - **T-4.7 Server/Client Components** — recommend **effect-typed, not a | ||
| structural distinction** (consistent with B3; SC/CC differ by effect). | ||
| - **T-4.8 effect-type-set closure** — recommend `HttpEffect/QueueEffect/…` |
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
There was a problem hiding this comment.
Fixed in 831e14e. Agreed — a closed HttpEffect|QueueEffect enum IS the parallel effect taxonomy B3 forbids (P2). The "effect-type-set closure" default is dissolved: there is NO effect enum to close. HttpEffect/QueueEffect are typed CARRIERS read from the type signature by lens/effect.dag — the signature is the single effect authority (B3). Non-default; resolved by B3. — sent from deep-wolf-155
|
Review metadata
1. Story of the diffThis PR adds 2. Invariant categories1. LAYER MODEL — substrate vs implementationCompliant — this diff does not directly mutate substrate files or Dag-resident types; it keeps the ledger as an index/proposal layer by saying ratified decisions are “Encoded in the cited file(s)” and that “file headers are authority” at 2. INVARIANTS.md + modeling-discipline.mdFinding — Boundary Discipline / single-authority metadata. 3. CODING.mdN/A — documentation-only PR; no Rust implementation surface, helper placement, method/free-function shape, or error/result API is changed. 4. TESTING.mdN/A — documentation-only decision ledger; no executable behavior, runner predicate, or generated/test harness surface changes. The absence of tests is appropriate for this patch. 5. LOCKED DESIGN DECISIONSFinding — unenumerated effects commitment. The thesis-level effect direction is that effects are read from type-signature shape, with “no parallel taxonomy or annotation layer” and that tracking effects as a separate enumerated concept is the bug pattern; this is stated in 6. TRACKED vs UNTRACKED DEBTCompliant — the ledger does not introduce a scaffold or temporary implementation shape; it bounds its proposed decisions through an explicit ratification flow at 2.5. Top-down PM intent reviewFinding — planning artifact would route work against the intended effect model. The highest-level intent for the unenumerated-effects bug class is that operation effects are derived from type-signature shape, consumers walk signatures directly, and “there is no parallel taxonomy or annotation layer” ( 3. VerdictREQUEST_CHANGES The ledger structure is useful and mostly respects authority separation, but the T-4.8 lower-cascade bullet contradicts the ratified-looking B3 effect authority and the thesis’s unenumerated-effects commitment. Because this is a design/planning artifact for substrate work, that contradiction should be fixed before landing. |
…efer)
B2-OMNI (00_compile.dag + 01_tokenize/02_parse/05_emit headers):
parse is NOT special — it is ONE instance of ingestion. The pipeline is
two PARAMETERIZED BOUNDARIES + a language-agnostic core:
ingest(Source, LanguageModel) -> Node [tokenize+parse = generic
walker over declarative grammar; .dag is language #1, not hardcoded]
core: Node -> InferredTree [normalize∘resolve∘infer]
emit(InferredTree, TargetModel) -> Source ; eval -> Value
compile = emit∘core∘ingest ; run = eval∘core∘ingest
Node is the universal pivot ⇒ O(N+M) not O(N×M): adding model #k is
O(1), instantly cross-composes. ingest(Rust)∘emit(Python) = Rust→Python
with no translator ever written — the visceral cash. C5: ingest=emit⁻¹
on the lossless core, fail-closed outside.
IR-1 (04_infer.dag): InferredTree is FROZEN now, not emergent. It is NOT
a new type — per A1 it is Node + one flat per-node InferredFacts
coordinate {resolved_type, cardinality, inhabits, cost(U1), descent(A2)};
effects are NOT a field (B3 single authority); a 6th field is a STOP.
Folded C1 in: infer is the bounded Find phase (structural unification,
empty ⇒ Diagnostic, never fabrication); the v2 12-file split cannot recur.
DECISIONS.md ledger: U2/C3/C1/B3/C4/C5 + K-1/IR-1/B2-OMNI moved to
RATIFIED Part 1 with encode pointers; U2 carries the no-bypass sharpening
(cost total over the closed kernel). Only B1 remains PROPOSED (the
content-hash *home* pick — recommendation: std-level).
Per-header text encoding of U2/C3/B3/C4 into their own files is the next
pass (ledger records them ratified; nothing lost). C1/IR-1/B2-OMNI are
encoded here.
v2->v4 bootstrap viability OK (57 modules, 0 diagnostics).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Operator-ratified 2026-05-15 (long-held v2 intent, de-deferred to stress the architecture up front). Three maximally-diverse parallel tasks; each is a falsification probe for the just-ratified B2-OMNI O(N+M) claim — if adding one is O(1) the claim is validated, if any forces a core change it's surfaced now. - T-4.9 extdeps/languages/verilog.dag — concurrency vs the 5 behaviors; the IN-B validation probe (a needed 6th `Concurrent` behavior = C1 escalation, by design caught early). Win: one .dag FSM → simulable Verilog + Rust reference, same Node. - T-4.10 extdeps/formats/spice.dag — is LanguageModel actually general or programming-shaped? Netlist has no control flow. Fork resolved: formats/ (a netlist is a data format), Shape B. Win: .dag → simulable SPICE netlist (omni-emission reaches analog hardware). - T-4.11 test/claim/boundary/english_ingest_fail_closed.dag — English is NOT a language model (fork resolved): a boundary-honesty probe. Win: (a) Shape B English docs (≈ T-16 markdown, no new substrate); (b) the honest one — ingest(prose) → precise Diagnostic, never a fabricated parse (no-engine made visible). TASKS.md: added to the Phase-1 parallel graph + a combined task section; count statements made drift-proof (no hardcoded number — per T-15). DECISIONS.md: L-1 ratified row with the fork resolutions + file targets. v2->v4 bootstrap viability OK (60 modules, 0 diagnostics — +3 scaffolds). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
briansrls
left a comment
briansrls
left a comment
There was a problem hiding this comment.
Review metadata
- Provider / model:
codex/unknown - Commit:
729db325· Trigger:schedule - Thinking:
301s wall
BLOCKING (4)
Root Cause
src/v4/compiler/04_infer.dagIR-1 mixes canonical substrate facts with inferred/cache facts → removecardinalityfrom the frozen payload or make it a derived view fromresolved_type.src/v4/compiler/00_compile.dagB2-OMNI overgeneralized C5 from language targets to formats → keep compilerTargetModelto Shape A language specs and model SPICE/English as Shape B user-program emitters.src/v4/extdeps/languages/verilog.dagThe stress-probe scaffolds were authored before selecting exact external authorities → cite the concrete Verilog and SPICE specs/versions before these worker contracts become immutable.
| // substrate-extension STOP→escalate, never an inline add): | ||
| // - resolved_type : Node (the node's type — itself a Node in | ||
| // the type substrate, Shape S) | ||
| // - cardinality : Cardinality (the Cardinality-connective fact) |
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
There was a problem hiding this comment.
Fixed in 831e14e. Correct, same class as effects-not-a-field (B3). cardinality dropped from InferredFacts — the Cardinality connective is one of the 6 (A1) and resolved_type: Node already carries it structurally; a separate field is a second authority (P2). It is read from resolved_type, never stored. Frozen payload is now 4 fields {resolved_type, inhabits, cost, descent}; cardinality + effects are explicitly NOT fields. — sent from deep-wolf-155
| // - compiler/01_tokenize.dag → 05_emit.dag + 05_eval.dag (core + boundaries) | ||
| // - std/diagnostic.dag for error propagation | ||
| // - extdeps/languages/*.dag for TargetSpec | ||
| // - extdeps/languages/*.dag + extdeps/formats/*.dag — the bidirectional |
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
There was a problem hiding this comment.
Fixed in 831e14e. Correct — Shape A/B must not blur. The ingest/emit LanguageModel/TargetModel parameter is extdeps/languages/* (Shape A) ONLY. extdeps/formats/* (csv/json/spice/…) is Shape B — emitted/consumed by user .dag programs walking typed values, NOT a compiler boundary parameter. Added an explicit Shape-A-vs-B block to 00_compile.dag; the Consumes line no longer lumps formats with languages. — sent from deep-wolf-155
|
Violations (could not place on specific lines):
|
B1 (operator-confirmed std-level fold):
- std/primitive.dag: `Hash` — opaque digest carrier (no content accessor,
K-1-style). content_hash : Node -> Hash (Merkle fold over A1-canonical
Node) is a std/node.dag contract item the T-1 worker encodes per relay
(it owns that file). One scheme; T-15/T-20/T-21/A3/C5 consume it.
Per-header encoding (ledger already recorded these RATIFIED):
- U2: cost.dag — cost is TOTAL over the closed kernel (no per-feature
opt-in; a future addition cannot bypass cost = no-eternal-maintenance);
single U1 carrier; verify lattice axioms. complexity.dag — consumes
cost, never re-derives (one mechanism, two projections).
- C3: 03_normalize.dag — dissolves EXACTLY the 4 sugar forms
{service,fn,type,operation}; not identity, not open-ended; 5th = STOP.
- B3: lens/effect.dag — the type signature is the single effect
authority; this lens READS it; coordination effect-types are carriers
over it, not a parallel taxonomy.
- C4: workflow/ci.dag + STRUCTURE.md invariant 8 — on-disk emitted
artifacts (ci.yml, trampoline) are committed==emit checked projections
guarded by the A3 machine; "no editable generated AUTHORITY", not "no
generated bytes".
DECISIONS.md: B1 added to RATIFIED Part 1; Part 3 marked CLOSED — every
design fork ratified + encoded. Only the T-1-worker-owned node.dag items
(K-1, content_hash) remain to be encoded by that worker per relay.
v2->v4 bootstrap viability OK (60 modules, 0 diagnostics).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- 3248138046: B1 overclaimed "A1 makes Node canonical." A1 does NOT — it covers recursion/axis-closure/generics/termination only. Canonical form is now a SEPARATE explicit node.dag authority (the canonical-form clause) the T-1 worker encodes; B1 depends on it, not subsumed by A1. Relay updated: THREE node.dag-contract items (K-1, canonical-form clause, content_hash). - 3248138059: T-4.8 effect-type-set-closure default contradicted B3 (a closed effect enum IS the parallel taxonomy B3 forbids). Dissolved: no effect enum; HttpEffect/QueueEffect are carriers read from the signature by lens/effect.dag. - 3248488135: IR-1 froze `cardinality` in InferredFacts though resolved_type already carries the Cardinality connective (A1) — second authority. Dropped; cardinality (like effects, B3) is read from resolved_type, not stored. Frozen payload now 4 fields. - 3248488229: 00_compile.dag B2-OMNI lumped extdeps/formats/* with languages/* as the LanguageModel param — Shape B blurred into the Shape A compiler boundary. Corrected: ingest/emit param is languages/* ONLY; formats/* is user-program (Shape B) emission, not a compiler boundary parameter. v2->v4 bootstrap viability OK (60 modules, 0 diagnostics). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
Review metadata
1. Story of the diffThis PR turns v4’s remaining design questions into a decision ledger and pushes the ratified decisions into the v4 scaffold headers so workers inherit the decisions at the owning files rather than from floating review context. The main architectural move is B2-OMNI: compile/run are reframed as 2. Invariant categories
Finding — BLOCKING, substrate authority drift.
Finding — BLOCKING, P2 Boundary Discipline / Shape A-vs-B single authority.
N/A — this diff does not add Rust implementation code or new function bodies; it is v4
Compliant — no executable behavior changes are introduced, and the new English boundary work is framed as a
Finding — BLOCKING, locked Shape A/B distinction blurred. The thesis-level Shape A/B split says compiler language targets live in
Compliant, with the above contradictions fixed. The new scaffolds have named owning tasks and bounds: SPICE is 2.5. Top-down PM intent reviewFinding — BLOCKING. The highest-level PM intent for omni-emission is the two-shape distinction: Shape A compiler language targets are emitted directly via language specs, while Shape B artifacts like SPICE netlists and natural-language docs are produced by user 3. VerdictREQUEST_CHANGES. The PR is directionally strong and most of the ledger/header propagation is well-formed, but two design-contract lines still undermine load-bearing authorities: B1 canonicality is partially attributed to A1, and |
T-1 worker review of #3150 caught a real lapse: this PR added 3 .dag files + test/claim/boundary/ + DECISIONS.md but STRUCTURE.md's File tree + Total were unchanged. Invariant #1 ("every file in the tree is enumerated above") is the v4 anti-gaming keystone — a decorative enumeration defeats it. - Added CULTURE.md + DECISIONS.md to the docs list; TASKS.md count line drift-proofed ("15 XL tasks" was stale). - languages/: + verilog.dag (T-4.9). formats/: + spice.dag (T-4.10). extdeps count 15 -> 17. - test/claim/: enumerated the claim subdirs incl. the new boundary/ + english_ingest_fail_closed.dag (T-4.11). - Total: 57 .dag + 3 docs -> 60 .dag + 5 docs (= 70 files), with a note that the enumeration (not the count) is authoritative; count is a checksum updated on every ratified addition. v2->v4 bootstrap viability OK (60 modules, 0 diagnostics — matches the enumeration). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Full-PR review (openai-pro, sha 831e14e) + the line-less violations pass. Two were sibling-mentions I fixed in one place but not the other (feedback_discipline_change_audit_all_contract_mentions): - primitive.dag:14 — "Merkle catamorphism over the A1-canonical Node" reintroduced the exact ambiguity B1 closes (I fixed the ledger wording in 831e14e but not this sibling note). Now: "over the std/node.dag canonical-form clause (SEPARATE node.dag authority, NOT subsumed by A1)". - 05_emit.dag:16 — `TargetModel` "data (extdeps/languages|formats/*)" blurred Shape A/B (I fixed 00_compile.dag for the same class in 831e14e but missed this sibling). Now: languages/* ONLY (Shape A); formats/* is Shape B, not a compiler emit param. - verilog.dag:6 — anchored to Wikipedia; extdeps fidelity wants a versioned spec. Now: IEEE 1364-2005 is the fidelity authority; Wikipedia demoted to context. - spice.dag:6 — same. Now: SPICE3f5 User's Guide (Quarles 1993) + pinned ngspice manual as the versioned authority; Wikipedia context. Full-branch grep confirms these were the ONLY instances of each class (no third-pass risk). v2->v4 bootstrap viability OK (60 modules, 0 diagnostics). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
Addressed the REQUEST_CHANGES (openai-pro sha 831e14e) + the line-less violations pass — all 4 BLOCKING fixed in 74e0bf5:
Both "could not place on specific lines" anchor BLOCKINGs are included. fmt clean; v2→v4 bootstrap viability OK (60 modules, 0 diagnostics). The two sibling-mention fixes (#1, #2) were exactly — sent from deep-wolf-155 |
briansrls
left a comment
briansrls
left a comment
There was a problem hiding this comment.
Review metadata
- Provider / model:
codex/unknown - Commit:
74e0bf52· Trigger:schedule - Thinking:
476s wall
BLOCKING (2)
Root Cause
src/v4/TASKS.mdT-4.10/T-4.11 probe framing still treats Shape B formats and non-language probes as B2 LanguageModel coverage after 00_compile narrowed LanguageModel to extdeps/languages/* -> keep B2 probes to Shape A languages and frame Shape B formats as user-program artifact libraries plus boundary tests.
Non-blocking — Improvements (fix in-PR if easy, else defer to roadmap)
src/v4/DECISIONS.mdPart 5 still reads like there are open PROPOSED items after Part 3 closes them; cleanup is cheap once the Shape A/B wording is corrected.
| // is the GENERIC walker over a `LanguageModel`'s declarative grammar | ||
| // (B2). Parsing .dag is `parse(tokens, dag_model.grammar)` — `.dag` is | ||
| // language #1, never hardcoded/special-cased. The same generic walker | ||
| // ingests Rust/JSON/… via their models. This is the load-bearing enabler |
There was a problem hiding this comment.
BLOCKING: JSON is a Shape B format, so making it an example of the generic LanguageModel parser re-blurs the compiler ingest boundary that THESIS Shape A/Shape B keeps separate.
There was a problem hiding this comment.
Fixed in 2fbcc73. Correct — JSON is Shape B. Removed it as a generic-walker example; the header now cites only Shape A languages (Rust/Python/…) and explicitly encodes that Shape B formats (JSON/SPICE) do NOT pass through the compiler-language parser — they ingest via the formats data-model path. The Shape A/B boundary is now stated in-file so it can't silently regress.
— sent from deep-wolf-155
| // reference. NOT Wikipedia — https://en.wikipedia.org/wiki/SPICE | ||
| // is accessible context only. | ||
| // Stress (T-4.10 — a B2-OMNI falsification probe): is the format / | ||
| // LanguageModel abstraction ACTUALLY general, or secretly |
There was a problem hiding this comment.
BLOCKING: SPICE is ratified as Shape B, so using it to test the LanguageModel abstraction puts a format back on the compiler-language boundary under P2.
There was a problem hiding this comment.
Fixed in 2fbcc73. Agreed — SPICE is ratified Shape B (L-1). Reframed the T-4.10 stress to test the Shape B format data-model abstraction only, with an explicit "NOT a LanguageModel, never touches the compiler-language (Shape A) boundary" clause. The netlist/no-control-flow falsification value is preserved, correctly scoped to the format model.
— sent from deep-wolf-155
|
Review metadata
1. Story of the diffThis PR turns the remaining v4 design calls into an explicit decisions ledger and then pushes those decisions into the v4 scaffold headers. The load-bearing mechanism is: 2. Invariant categories
2.5. Top-down PM intent reviewFinding, BLOCKING. The highest-level intent is that Shape A and Shape B remain semantically distinct: compiler language targets live under Secondary NON-BLOCKING cleanup. The ledger ends with stale ratification-flow language: 3. VerdictREQUEST_CHANGES The main architecture is strong and most of the batch ratifications are encoded cleanly, especially IR-1, B1, B3, C4, and U2. I would not land it with |
…specs not libraries Operator correction + directive 2026-05-15: C5-fidelity: there is NO a-priori "trivia" category (my earlier framing was sloppy). Round-trip fidelity is parameterized by the model. Each surface feature, per language, gets an explicit disposition: Modeled (∈ F, Node-bearing, round-trips both ways — Python indentation IS block structure; comments IF modeled) | Declared-normalized (deliberately not in F; emit∘ingest canonicalizes — Go/C++ whitespace; a *declared*, reviewable loss, never silent) | Fail-closed (encountered but neither → Diagnostic, no-engine). emit∘ingest=id holds restricted to F; "lossless core" ≡ F; fidelity = declared model completeness, boundary declared not assumed. L-2: model the language SPECIFICATION, not libraries. extdeps/languages/* models the versioned upstream spec (the anchor IS that spec — cf. the verilog→IEEE-1364 / spice→SPICE3f5 fix); libraries are just programs in the modeled language = Node, never modeled. This makes F principled (the spec's own meaning-vs-lexical distinction defines it) and maximally general. Frameworks edge PROPOSED pending confirm: admit frameworks/* only as a sub-language with its own non-host-expressible rules (React Hooks/effect/reconciliation), never a library API surface. Encoded: DECISIONS.md (C5-fidelity + L-2 rows) + TASKS.md T-4 authoring contract (model-spec-not-libraries; per-feature disposition; bidirectional grammar relation; unambiguous-ingest-or-Diagnostic; semantic-feedback- parse = STOP). v2->v4 bootstrap viability OK (60 modules, 0 diagnostics). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…robes Operator directive 2026-05-15 (parallel, up front — same pattern as L-1's verilog/spice/english, but spanning DOWN the stack): - T-4.12 extdeps/languages/llvm_ir.dag — generalize down the stack (SSA/dominance/phi); F ~all-structural (clean C5-fidelity contrast). Anchor: LLVM LangRef (pinned), per L-2. - T-4.13 extdeps/languages/machine_code.dag — bottom of stack; NO cosmetic surface (C5-fidelity limit test); disassembly = the extreme fail-closed/no-engine case. PROPOSED fork: ONE model parameterized by an Isa (recommended) vs per-ISA files. Anchor: ISA SDM/ARM (pinned). - T-4.14 extdeps/languages/ptx.dag — CUDA SIMT data-parallel vs the 5 behaviors (IN-B bet, like Verilog's concurrency). PROPOSED fork: PTX spec'd IR (recommended) vs CUDA-C++ cpp.dag extension. Anchor: NVIDIA PTX ISA (pinned). Concretizes C5-fidelity: targets span source (F may carry cosmetics by intent) → IR (F structural) → machine code (F = encoding, no cosmetics) — "we don't necessarily output a binary; F is model+intent-relative." STRUCTURE.md File tree + counts re-enumerated (extdeps 17->20, Total 60->63 .dag = 73 files) — invariant #1 kept in sync this time. DECISIONS.md L-3 row + the 2 PROPOSED forks recorded. v2->v4 bootstrap viability OK (63 modules, 0 diagnostics). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
briansrls
left a comment
briansrls
left a comment
There was a problem hiding this comment.
Review metadata
- Provider / model:
codex/unknown - Commit:
11f104bd· Trigger:schedule - Thinking:
312s wall
BLOCKING (1)
Root Cause
src/v4/DECISIONS.mdL-2 ratifies language-spec fidelity but defers the framework sub-language boundary → ratify the criterion or remove it before extdeps/frameworks/react.dag consumes it.
| | **C4** | On-disk emitted artifacts (`ci.yml`, trampoline) are committed==emit(source) checked projections, not editable authority (same machine as A3) | `workflow/ci.dag` + `STRUCTURE.md` + `workflow/bootstrap.dag` | | ||
| | **C5** | Ingest = emit⁻¹ on each model's lossless core; outside ⇒ fail-closed Diagnostic. Guaranteed ASAP via testgen `bidirectional_roundtrip` (Phase-1.5, gated per-PR via the B1 hash from first language-model commit) | `extdeps/languages/*` + `lens/testgen.dag` + C5 note | | ||
| | **C5-fidelity** | Round-trip fidelity is **parameterized by the model, not a fixed "trivia" category** (operator-corrected 2026-05-15). Each surface feature, per language, has an explicit disposition in the model: **Modeled** (∈ F → Node-bearing; both `ingest∘emit=id` and `emit∘ingest=id` for that feature — e.g. Python indentation IS block structure; comments IF modeled) \| **Declared-normalized** (deliberately not in F; `emit∘ingest` canonicalizes it — Go/C++ insignificant whitespace; a *declared* loss, reviewable, never silent) \| **Fail-closed** (encountered but neither → Diagnostic, no-engine). `emit∘ingest=id` holds restricted to F; "lossless core" ≡ F; round-trip fidelity = model completeness, boundary declared not assumed. F is drawn from the spec's own meaning-vs-lexical distinction (see L-2) | `extdeps/languages/*` headers + `TASKS.md` T-4 + C5 | | ||
| | **L-2** | **Model the language SPECIFICATION, not libraries** (operator-ratified 2026-05-15). `extdeps/languages/*` models the versioned upstream spec (grammar+type-system+semantics; the anchor IS that spec by construction — cf. the verilog→IEEE-1364 / spice→SPICE3f5 fix). Libraries (std/crates/packages) are NOT modeled — a library is just a program in the modeled language = `Node` like any other. This makes F principled (the spec defines meaning-vs-lexical, not our judgment) and maximally general. Frameworks edge (PROPOSED, awaiting confirm): admit a `frameworks/*` model ONLY when it constitutes a sub-language with its own non-host-expressible rules (React Rules-of-Hooks/effect-lifecycle/reconciliation), never a library API surface | `extdeps/languages/*` + `extdeps/frameworks/*` headers + `TASKS.md` T-4 | |
There was a problem hiding this comment.
BLOCKING: L-2 leaves the frameworks admission rule PROPOSED inside the ratified/closed ledger, so T-4.7 has no binding spec-vs-library authority under STRUCTURE zero-deferrals.
There was a problem hiding this comment.
Fixed in 2fbcc73. Resolved as a cascade-closure of the already-ratified L-2, not a new ratification: a framework with its own non-host-expressible rules (React Rules-of-Hooks/effect-lifecycle/reconciliation) IS a "language specification" under L-2; a framework that is only a library API is a library, which L-2 already excludes from modeling. Dropped the PROPOSED/awaiting-confirm hedge — the rule is now BINDING and T-4.7 has spec-vs-library authority. No PROPOSED residue in the closed ledger.
— sent from deep-wolf-155
…ps; worked examples Operator: encode the tightening + embed worked examples + answer "does it depend on emission target?". Answer: the independence FACT is target-AGNOSTIC (a property of the IR dep graph + effects — feedback_target_agnostic_ir); WHETHER/HOW to realize it as parallel execution is target-DEPENDENT — a cost-derived decision (U1/U2 per-target realization) at emit(T-11)/eval(T-22), NOT this lens. Same ParallelismMap → different realizations (Rust rayon / Python GIL-so-no / Go goroutines / CUDA kernel / Verilog spatial HW); the lens never changes. "independent ⇒ spawn" is the forbidden heuristic. Modeling gap fixed: independence = structural-disjointness ∧ effect-independence ∧ (folds) combiner-associativity-by-inhabitance. So the lens now Consumes lens/effect.dag (B3 — Node-disjoint is necessary NOT sufficient; example B) and std/algebra.dag (Monoid inhabitance → ParallelReduce; example D — an algebra fact, not a heuristic). Prior Consumes (node+witness only) was materially under-specified (could declare effect-coupled subgraphs parallel — unsafe). Framed as a THESIS free consequence (THESIS:208, same family as auto-memoization/incremental), a lens not an annotation. 4 worked examples (pure independence / effect coupling / target-dependent realization / fold-is-algebra) embedded as the design contract. DECISIONS.md PAR-1 row added. v2->v4 bootstrap viability OK (63 modules, 0 diagnostics). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…elize" command Operator: "how do people actually use this? it must be well understood." The 'parallelize my code please' framing is forbidden — a magic auto-parallelize transform is engine-shaped from the user's seat, exactly the black box this ethos rejects. Encoded the UX contract into the header: - It is NOT a command. It is a LENS the user READS — same UX category as complexity/cost: "what is independent and WHY" (structural reason, or a precise fail-closed Diagnostic when not). Never silent, never a guess. - Control surface = structure, not a knob: you make work parallel by writing it independent+pure (the language already requires this); the lens confirms or explains why-not. NO @parallel annotation; no surprise. - Default introspective: compiler realizes at the target only where the cost model wins, and that decision is itself derived + INSPECTABLE per target (CPython → honestly "no, GIL", not silent magic). - Guarantee via apply_lens(parallelism, Enforce) (T-23 — the uniform opt-in-depth bridge, same as complexity/effects/synthesis; nothing new to learn). Orthogonal to the lens read being a fail-closed Witness. Net: it is *more* understood than a normal auto-parallelizer (which is a read-the-asm black box) — the independence map AND the per-target realize-decision are first-class derived, readable, explainable facts. DECISIONS.md PAR-1 updated with the user-facing contract. v2->v4 bootstrap viability OK (63 modules, 0 diagnostics). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ndertracked) Operator: "language queries — what do we say about that in v4?" Verified: v4 HAS it (THESIS §1.5/95-101 user-defined dimensions = user-written lenses; home = lens/application.dag T-23) but never NAMED it as "query", and the key property was folklore. Named + asserted in the T-23 header + ledger QRY-1: - A query IS a user-defined lens (same Node->Set<Match>/Witness shape; NOT a new subsystem). - LANGUAGE-AGNOSTIC by the B2-OMNI Node pivot: one query runs over every ingested language (Rust/Python/Verilog/SQL/…), O(N+M) — a free consequence, now contract not folklore. - Composes derived lens facts (effects/complexity/parallelism) — > AST tools. Introspective read by default + optional Enforce (uniform UX). - Query result is itself a Node (bounded kernel — no new carrier). Fail-closed/no-engine. - NON-MOVE (binding): NO query.dag / query subsystem — a query IS a lens; a query feature = parallel authority = STOP. (No new files; STRUCTURE.md tree unaffected — correctly.) v2->v4 bootstrap viability OK (63 modules, 0 diagnostics). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Operator-corrected: "query = user-defined lens" was wrong — user-vs-
kernel is PROVENANCE, not structure (feedback_naming_is_aliasing; the
compiler sees through who authored a lens). A name distinguished only by
who-wrote-it is decorative ⇒ must not exist ("a real distinction, or not
exist"). No query concept/noun/subsystem.
The REAL axis the operator surfaced ("a filter on a lens") = scope/
parameterization, ALREADY modeled with no new noun:
1. SectionRef (DeclarationScope|NodeScope) — apply a lens AT a scope.
2. Parameterized lens signatures — the affected-set lens (T-21)
`(Dag,Diff)->Witness<ReExecFrontier>` is the archetype: "the
affected set of THIS change" = affected_set(dag,diff), already
first-class. Not a query system.
Lenses, read at scopes and/or parameterized. No query.dag. The B2-OMNI
"one read over the Node pivot → all ingested languages" property is a
property of ANY lens, not a "query".
lens/application.dag QRY-1 note rewritten as the retraction; DECISIONS.md
QRY-1 marked RETRACTED & reframed. No new files (tree unaffected —
correctly: a query is not new substrate).
v2->v4 bootstrap viability OK (63 modules, 0 diagnostics).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
Review metadata
1. Story of the diffThis PR turns v4’s remaining open design surface into an explicit decision ledger and propagates those decisions into the scaffold headers that future workers will execute. The central mechanism is 2. Invariant categories
Finding — BLOCKING, Shape A/B boundary blur. The diff correctly adds an explicit Shape A/Shape B boundary in
Finding — P2 Boundary Discipline / single authority. The same Shape A/B fact has two conflicting consumers in the diff:
N/A — no Rust implementation, helper placement, method-vs-function, or executable error-shape code changed. The diff is docs plus
Compliant. The PR does not add executable behavior requiring a Rust test; the new English boundary probe is represented as
Finding — BLOCKING, locked Shape A/B intent is semantically diluted. The PR itself restates the locked boundary in
Finding — BLOCKING for this PR’s purpose, stale ratification state remains in the ledger and worker headers. 2.5. Top-down PM intent reviewFinding — BLOCKING. The highest-level omni-emission intent requires the two-shape split: compiler language targets are Shape A, while YAML/Terraform/Kubernetes/SPICE/docs/SQL/OpenAPI-style artifacts are Shape B user programs, and “the two-shape distinction … must not be blurred” per the thesis authority ( 3. VerdictREQUEST_CHANGES The design direction is mostly coherent and valuable, but the PR currently leaves two blocking planning-contract issues: Shape B examples leak back into the compiler LanguageModel/parser boundary, and the supposedly ratified ledger still contains stale PROPOSED/confirm language. Both are small textual fixes, but because these headers are the authority future workers will execute, they need to be made unambiguous before landing. |
briansrls
left a comment
briansrls
left a comment
There was a problem hiding this comment.
Review metadata
- Provider / model:
codex/unknown - Commit:
fb8fdac3· Trigger:schedule - Thinking:
224s wall
BLOCKING (1)
Root Cause
src/v4/DECISIONS.mdratified ledger still admits PROPOSED sub-decisions → either ratify the machine_code and PTX choices in the ledger and downstream headers or move those probes back out of the closed file tree until decided
Non-blocking — Strengths
src/v4/compiler/04_infer.dagIR-1 cleanly keeps InferredTree as Node plus one flat fact coordinate, matching the thesis boundary discipline.
| | **IR-1** | `InferredTree` is NOT a new type — it is `Node` + a FROZEN flat `InferredFacts` coordinate (resolved_type, inhabits, cost, descent); **cardinality and effects are NOT fields** — both already carried by `resolved_type`'s Node structure (the Cardinality connective / the signature), caching = second authority P2; a 5th field is a STOP; modeled up front, not emergent | `compiler/04_infer.dag` header | | ||
| | **B2-OMNI** | parse is one ingestion instance; `ingest`/`emit` are parameterized boundaries over declarative LanguageModels; Node is the universal pivot ⇒ O(N+M) not O(N×M); `.dag` is language #1, never hardcoded | `compiler/00_compile.dag` + `01_tokenize`/`02_parse`/`05_emit` headers | | ||
| | **L-1** | Verilog/SPICE/English added as **parallel B2-OMNI falsification probes** (T-4.9/4.10/4.11), maximally diverse on purpose. Forks resolved: English = boundary-honesty probe (Shape B emit + fail-closed ingest TestClaim), **not** a language model; SPICE = `extdeps/formats/` (netlist is a data format); Verilog = the IN-B concurrency validation (a needed 6th behavior = C1 escalation, by design caught early) | `TASKS.md` T-4.9/4.10/4.11 + `extdeps/languages/verilog.dag`, `extdeps/formats/spice.dag`, `test/claim/boundary/english_ingest_fail_closed.dag` | | ||
| | **L-3** | **Down-the-stack** B2-OMNI probes added (operator-ratified 2026-05-15, parallel up front): T-4.12 `extdeps/languages/llvm_ir.dag` (SSA IR — generalize down the stack; F ~all-structural), T-4.13 `extdeps/languages/machine_code.dag` (bottom of stack; no cosmetics; disassembly = the extreme fail-closed/no-engine test), T-4.14 `extdeps/languages/ptx.dag` (CUDA SIMT — IN-B bet vs the 5 behaviors). Validates the model across the **full target spectrum** (source→IR→machine code) — concretizes C5-fidelity ("we don't necessarily output a binary; F is model+intent-relative"). **PROPOSED forks (confirm):** machine_code = ONE model parameterized by an `Isa` (recommended — parameterize not enumerate, matches O(N+M)) vs per-ISA files; CUDA = model **PTX** (spec'd IR, recommended) vs CUDA-C++ as a cpp.dag extension | `TASKS.md` T-4.12/4.13/4.14 + `extdeps/languages/{llvm_ir,machine_code,ptx}.dag` | |
There was a problem hiding this comment.
BLOCKING: L-3 is listed as operator-ratified while still carrying unratified machine_code/PTX forks, so the closed ledger preserves a live design fork contrary to P5 progress-as-dissolution.
There was a problem hiding this comment.
Fixed in 2fbcc73. Neither fork was genuinely open — each is forced by an already-ratified parent and the alternative contradicts a ratified decision, so per P5 it dissolves rather than needing new ratification: (1) machine_code Isa-parameterized is forced by B2-OMNI — per-ISA files ARE the N×M trap B2-OMNI forbids; (2) CUDA→PTX is forced by L-3's own down-the-stack-IR scope — PTX is the SIMT IR layer T-4.14 exists to probe, whereas CUDA-C++ is just another Shape A source language already covered by general modeling (not a down-the-stack probe). PROPOSED-forks line removed; no live fork remains in the closed ledger.
— sent from deep-wolf-155
…grammar Operator-ratified 2026-05-15. Constructive pre-flight: frozen v2's parser accepts the full ratified kernel shape WITH fn bodies (40k-line src/v2 corpus + kernel probe, zero syntax errors — the 65 src/v2 errors are 100% post-parse import/cycle) but rejects bodiless fn (`expected LBrace`). Bridge-era substrate contracts stay comment-prose (current scaffold form); the fail-closed parse-class gate over src/v4 is safe (v2 never asked to parse bodiless fn). Forward requirement (the conditional making the deferral acceptable): v4's own .dag grammar (T-7) MUST admit a bodiless signature declaration as a first-class production, graduating the "immutable I/O contract for the worker" from comment-prose to compiler-enforced. Resolves the T-20 parse-surface sub-check. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- 02_parse.dag: drop JSON (Shape B) as a LanguageModel-parser example; encode the Shape A/B ingest boundary explicitly (resolves 3248889077). - extdeps/formats/spice.dag: reframe the T-4.10 probe as a Shape B format data-model stress, NOT the LanguageModel/compiler-language boundary; SPICE is Shape B per L-1 (resolves 3248889081). - DECISIONS.md L-2: frameworks admission promoted from PROPOSED to BINDING as a cascade-closure of ratified L-2 — a framework with its own non-host-expressible rules IS a language spec under L-2; a library API is already not-modeled. T-4.7 binding authority (resolves 3250029061). - DECISIONS.md L-3: machine_code/PTX forks DISSOLVED — machine_code Isa-parameterized is forced by B2-OMNI (per-ISA = the N×M trap); CUDA->PTX forced by L-3's own down-the-stack-IR scope. No PROPOSED residue in the closed ledger (resolves 3250390911). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Resolves the conflict with #3150 (v4 design-decisions ledger): - src/v4/std/primitive.dag — modify/delete: kept DELETED (this PR's purpose). #3150 had added a `Hash` header note to primitive.dag (B1). With the file gone, `Hash` — not a scalar — re-homes to std/node.dag alongside `content_hash` (the T-1 closeout, where B1 already places the fold + canonical-form clause). Noted in STRUCTURE.md; DECISIONS.md B1 should update its "Hash → primitive.dag" line accordingly. - src/v4/STRUCTURE.md — content conflict: kept #3150's docs-listing (DECISIONS.md) and the count-is-a-checksum framing; kept this PR's scalar/numeric decomposition section. Total recomputed against the merged tree: 68 .dag + 5 docs + 5 .gitkeep = 78. Verified: v2 seed indexes all 68 merged v4 modules, 0 diagnostics.
Addresses the claude review on PR #3152. The decomposition deletes std/primitive.dag, but #3150's B1 entry routed the `Hash` content-address digest to that file — so after merge B1 pointed at a deleted path and `Hash` (an operator-ratified substrate type) had no declaring file. My STRUCTURE.md note said B1 "should be updated" — that was a deferral, not a fix. Fixed for real: - std/node.dag declares `type Hash` — body-less opaque, exactly like `Symbol` (K-1 discipline: equality only, no content accessor). `Hash` is a content identity, not a scalar, so it belongs at the substrate root beside `Symbol`, not in the deleted scalar junk-drawer. node.dag's Owns line updated. - DECISIONS.md B1 "Encoded in" column re-pointed: drop the deleted std/primitive.dag; `Hash` + canonical-form clause + content_hash fold all → std/node.dag. The clause + fold stay the T-1 closeout (as B1 already states); the `Hash` type lands now. - STRUCTURE.md §decomposition "Hash re-homing" rewritten from "should be updated" (deferral) to done. Verified: v2 seed indexes all 68 v4 modules, 0 diagnostics. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
#3152) * WIP: v4 origin * v4: decompose std/primitive.dag into six concept-anchored scalar files primitive.dag lumped five unrelated concepts (Int/Float/String/Char/Bool) under the label "primitive" — but per THESIS epistemic stacking the genuine primitives are the six connectives (node.dag) and the algebra roots (algebra.dag); Int/Bool/String/Char are compositions that attach by inhabitance, not primitives. The file was a junk-drawer, against STRUCTURE.md's "file = one cohesive concept" principle. Replaced by six concept-located files, each anchored to a real external concept (Wikipedia/spec) — the same anchored-concept discipline extdeps/ uses: - std/logic.dag — Bool (classical two-valued logic) - std/nat.dag — Nat (Peano natural numbers) - std/machine.dag — Byte/Word*/MachineWidth/PointerWidth - std/integer.dag — Int + fixed-width ints (Nat projected onto a width) - std/float.dag — Float (IEEE-754; the one genuine opaque scalar) - std/text.dag — Char (Unicode code point) + String (FreeMonoid<Char>) Each concrete type declares its OWN inhabitance in its own file (INVARIANTS P2 single-authority); algebra.dag owns the algebra structures only. This also resolves the algebra<->primitive header authority conflict and the would-be import cycle surfaced by the T-2 worker's STOP. Also updated: STRUCTURE.md (closed file tree + count 57->62 + architectural commitment #2 re-homed: Tier-2 partial-op totalization now lives with each scalar type) + §"Scalar/numeric concept decomposition"; TASKS.md T-3 (file list + the scalar/numeric dependency-ordered cluster); node.dag header Q1 (literal-value-typing pointer); ~12 consumer headers repointed off the deleted file (String is kernel-ambient — no import). Verified: v2 seed indexes all 62 v4 modules, 0 diagnostics. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * v4: Float is grounded, not opaque — fix the epistemic-stacking conflict Addresses codex BLOCKING on PR #3152. float.dag and STRUCTURE.md described Float as "the one genuine opaque/irreducible scalar" that "does not decompose into the algebra roots" — which directly contradicts THESIS's "no concept is opaque" epistemic-stacking bar. The "opaque" framing was a mistake: it wrongly read v3's body-less Ieee754Float scaffold as intentional opacity and wrongly analogized Float to Symbol (which IS opaque by K-1 design). Float is not opaque. It decomposes: a sign/exponent/mantissa bit-record over std/machine.dag, with NaN/±Inf/signed-zero modeled as explicit structure. What is genuinely special about Float is its ALGEBRA — rounding breaks associativity, NaN breaks reflexive equality — so it inhabits a weakened, rounding-aware algebra rather than an exact Field. The std/extdeps typescript.dag header already anticipates this as `ApproximateField`. Float's grounding is the IEEE-754 spec (the anchor) composed with that weakened algebra — a fully reconciled concept, not an exception to the minimal-primitive ontology. float.dag's header + STRUCTURE.md's decomposition entry rewritten accordingly; float.dag now also Consumes std/algebra.dag (the rounding-aware structure it inhabits). Verified: v2 seed indexes all 62 v4 modules, 0 diagnostics. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * WIP: v4 origin * v4: fix stale "opaque scalar" tree line + define kernel-ambient Addresses cursor APPROVE_WITH_COMMENTS on PR #3152. The Float-grounding commit (18ac69c) fixed the §decomposition section but missed the file-tree comment at STRUCTURE.md:27, which still read "(opaque scalar)" — the doc argued with itself. Tree line corrected to "rounding-aware algebra, not exact Field". Also added (cursor's optional note): a "Kernel-ambient types" paragraph defining that String/Int/Bool/Char/List/Map are seed-provided and usable without import, while the modeling file (text/integer/logic/ collection) remains each type's single authority — only the import edge is relaxed. This is what the "String … kernel-ambient — no import" header notes mean; stating it once removes the P2 ambiguity for workers. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * v4: fix STRUCTURE.md doc count — 4 docs (CULTURE.md was uncounted) Addresses codex APPROVE_WITH_COMMENTS on PR #3152. The closed-tree inventory said "3 docs" but src/v4/ has four top-level docs — STRUCTURE.md, BRIEF_TEMPLATE.md, TASKS.md, CULTURE.md. CULTURE.md was added by an earlier PR without updating STRUCTURE.md's count or its file-tree listing; this PR's count edit propagated the stale "3". Count corrected to "62 .dag files + 4 docs + 5 .gitkeep = 71"; CULTURE.md added to the file-tree listing so the closed-tree enumeration is complete (the listing must enumerate every file — INVARIANTS P2). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * Merge origin/main into v4-primitive-decomposition Resolves the conflict with #3150 (v4 design-decisions ledger): - src/v4/std/primitive.dag — modify/delete: kept DELETED (this PR's purpose). #3150 had added a `Hash` header note to primitive.dag (B1). With the file gone, `Hash` — not a scalar — re-homes to std/node.dag alongside `content_hash` (the T-1 closeout, where B1 already places the fold + canonical-form clause). Noted in STRUCTURE.md; DECISIONS.md B1 should update its "Hash → primitive.dag" line accordingly. - src/v4/STRUCTURE.md — content conflict: kept #3150's docs-listing (DECISIONS.md) and the count-is-a-checksum framing; kept this PR's scalar/numeric decomposition section. Total recomputed against the merged tree: 68 .dag + 5 docs + 5 .gitkeep = 78. Verified: v2 seed indexes all 68 merged v4 modules, 0 diagnostics. * v4: land the Hash type in node.dag — close the B1 tombstone Addresses the claude review on PR #3152. The decomposition deletes std/primitive.dag, but #3150's B1 entry routed the `Hash` content-address digest to that file — so after merge B1 pointed at a deleted path and `Hash` (an operator-ratified substrate type) had no declaring file. My STRUCTURE.md note said B1 "should be updated" — that was a deferral, not a fix. Fixed for real: - std/node.dag declares `type Hash` — body-less opaque, exactly like `Symbol` (K-1 discipline: equality only, no content accessor). `Hash` is a content identity, not a scalar, so it belongs at the substrate root beside `Symbol`, not in the deleted scalar junk-drawer. node.dag's Owns line updated. - DECISIONS.md B1 "Encoded in" column re-pointed: drop the deleted std/primitive.dag; `Hash` + canonical-form clause + content_hash fold all → std/node.dag. The clause + fold stay the T-1 closeout (as B1 already states); the `Hash` type lands now. - STRUCTURE.md §decomposition "Hash re-homing" rewritten from "should be updated" (deferral) to done. Verified: v2 seed indexes all 68 v4 modules, 0 diagnostics. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * v4: T-3 summary line — surface the T-2 dependency Addresses openai-pro APPROVE_WITH_COMMENTS on PR #3152 (non-blocking). The execution-graph summary said T-3 `[needs T-1]`, but the scalar/ numeric files consume algebra.dag — logic.dag needs BooleanAlgebra, nat.dag needs Monoid/Semiring (T-2). A worker dispatched off the summary alone could start the algebra-consuming scalar files before T-2 lands. The detailed T-3 dependency section was already correct; this aligns the summary line to it: `[needs T-1, T-2; diagnostic/ cardinality are T-1-only, the scalar/numeric stack needs T-2]`. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1 participant
Summary
The remaining design decisions, now ratified and encoded (this PR evolved from a proposal-ledger into the encoding PR — read it as substrate changes, not "just a doc").
src/v4/DECISIONS.md— the canonical decisions ledger. Part 1 RATIFIED (A1–A4, U1, U2, C1–C5, K-1, IR-1, B2-OMNI, L-1, B1) with where-each-is-encoded pointers. Part 3 is CLOSED — every fork ratified.InferredTree=Node+ frozen flatInferredFacts;cardinality/effectsNOT fields — single authority), B2-OMNI (the O(N+M) pivot; Shape A/B kept distinct), B1 (Hashcarrier inprimitive.dag;content_hashfold + canonical-form clause are node.dag-contract items the T-1 worker encodes per relay).extdeps/languages/verilog.dag(T-4.9, concurrency vs the 5 behaviors),extdeps/formats/spice.dag(T-4.10, Shape B, no control flow),test/claim/boundary/english_ingest_fail_closed.dag(T-4.11, fail-closed ingest). STRUCTURE.md File tree + counts re-enumerated (invariant Add SVG viz, test helpers, and makegen scaffold #1).Disjoint from #3149 (substrate review-fixes) and #3151 (T-1
node.dag). The B1 canonical-form clause +content_hashare routed to #3151 (the T-1 worker ownsnode.dag).Test plan
cargo fmt --all --checkclean🤖 Generated with Claude Code