v4 TASKS: HostModel / ModelCore / Protocols rows (Ratified Q1 + P4)

[briansrls]

Summary

PR #3437 (commit deda6f210) ratified three substrate concepts in docs/design-v4-compiler-homomorphism.md that had no src/v4/TASKS.md home. This adds three rows so the substrate gaps appear on the v4 plan:

• T-33 std/model_core.dag — shared substrate factoring (Ratified Q1, 2026-05-20). [needs T-1, T-2, T-3].
• T-34 std/host.dag — HostModel peer of LanguageModel (Ratified Q1, 2026-05-20). [needs T-33]. Consumer: T-22 eval + MVP-B route.
• T-4.15 extdeps/protocols/{rest,graphql,grpc}.dag — transport substrate, P4 (2026-05-20). [needs T-3, T-26]. Language-orthogonal: transport declares its own wire-format type system; LanguageModel bindings happen at T-16 composition time, not on this substrate. Out-of-scope for the initial single-target compiler; in-scope architecturally so glue derivation (T-16 omni-stack) isn't foreclosed.

The parallel-fill execution-graph block at the top of TASKS.md lists the three new rows alongside the existing T-2# / T-4.x entries.

Graph-edge propagation (6d7e15274 + 7ecdf4ecd + 5ab9d3397): the new substrate rows propagate into existing consumers' [needs …] contracts so dependency authority lives in one place (the [needs …] line), per Practice 5 single-authority discipline:

• T-4 now [needs T-3, P1-KEYSTONE, T-29, T-30, T-25-core, T-33] (LanguageModel cannot be authored before ModelCore exists). Side-branch graph updated from four → five feeders.
• T-22 now [needs T-9, T-34] (eval cannot be authored before the HostModel carrier exists).
• T-4.15 is NOT a T-4 consumer — language-orthogonality per P4. The LanguageModel ⊗ TransportModel composition lives at T-16 (which already [needs T-4 …]).

Scope discipline

Deliberately narrow. What this PR does NOT do, by operator direction (2026-05-20):

• No new doc. A "v4 modeling work catalogue" was proposed in the original dispatch brief; it would have re-aggregated facts whose source-of-truth is inline marks + TASKS.md + the existing docs/audit/coproduct-anemia-inventory.md census — directly colliding with the standing ledger principle (CLAUDE.md, operator 2026-05-19) that got the maintained-ledger doc class nuked program-wide via PR chore(v4): delete DECISIONS.md ledger; retire Practice-4 receipt rule #3389. The catalogue is canceled.
• No anemia-audit work. T-30 (std/ structural fact-density / hollow-alias gate, SCHEDULED) is the operator-ratified anemia discipline; docs/audit/coproduct-anemia-inventory.md is the existing 274-row v4 corpus census with its own P5 close trigger. The "anemia audit + CI grep gate + TestClaim shapes" sub-task in the original brief duplicated T-30.
• No Wave-1 LanguageModel fleshout dispatch. T-4 owns that on its own feeder chain. The fact-bundle reseed discipline is already operator-ratified 2026-05-17.
• No framework-as-library work. "FastAPI / SQLAlchemy / Vue / Tailwind" framework targets are out of scope per T-4's existing "model SPECIFICATIONS, not libraries" contract.
• T-4 fact-bundle authoring contract body prose is intentionally not re-expressed in terms of "LanguageModel extends ModelCore". The schedule edge belongs in [needs …] (added); the authoring-contract reframe is its own commit train, after T-33 lands. Schedule edge ≠ modeling reframe.

Related substrate slot, deliberately NOT added

The design doc also names std/system.dag (module decomposition substrate) as a sibling P4 gap, labeled "Currently informal." That is a separate follow-on, scheduled when glue derivation work activates. T-4.15's body cross-references it as related-not-bundled.

Test plan

• Verify the three new task IDs (T-33 / T-34 / T-4.15) don't collide with anything else in TASKS.md.
• Confirm the design doc citations resolve: docs/design-v4-compiler-homomorphism.md §"ModelCore", §"HostModel", §"P4", §"Ratified Q1", §"What's NOT in scope".
• Confirm the dependency edges [needs T-1, T-2, T-3] / [needs T-33] / [needs T-3, T-26] match what each row's body says — AND the propagated edges (T-33 on T-4's [needs], T-34 on T-22's [needs]) are reflected in both the execution-graph block and each task's definition.
• T-4.15's transport substrate is language-orthogonal (no T-4 dependency); LanguageModel ⊗ TransportModel composition is T-16's responsibility per P4.
• No content change in any .dag file — this is a planning PR only.

🤖 Generated with Claude Code

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: fecd3bd0 · Trigger: schedule
• Thinking: 147s wall

---

BLOCKING (3)

Root Cause

• src/v4/TASKS.md P4 protocol design is referenced only through a missing authority → land the design doc or replace the citation with an existing checkable authority and task contract.
• src/v4/TASKS.md New shared substrate was added only to the parallel-fill list → propagate the T-33 edge into T-4's needs and side-branch graph.
• src/v4/TASKS.md HostModel was scheduled as late substrate without propagating its consumer edge → add T-34 to T-22's needs or remove the claimed T-22 consumption.

⚠️ The added planning tasks need checkable authority and their dependency edges must be propagated before this task graph is safe to merge.

[briansrls]

Re: codex dashboard-only review verdict REQUEST_CHANGES (dashboard review id 15297, ran on sha 2b928e88d at 2026-05-20T08:24:01Z, parsed from stdout artifact).

Both flagged P2 single-authority issues are addressed at HEAD 7ecdf4ecd — across two follow-up commits:

1. 6d7e15274 propagated the new substrate edges into consumers' authoritative [needs …] contracts:

   • T-4 now [needs T-3, P1-KEYSTONE, T-29, T-30, T-25-core, T-33] (side-branch graph block + T-4 task-definition body, both updated; four → five feeders).
   • T-22 now [needs T-9, T-34] (parallel-fill execution-graph block).

2. 7ecdf4ecd reconciled the stale T-33 prose paragraph that earlier asserted "No silent change to T-4's [needs …] list as part of this PR" — true at 2b928e88d (when codex reviewed), but FALSE at 6d7e15274 once the graph edge landed. Without this second commit the P2 problem would have inverted (prose denying what [needs] actually did). New paragraph names what this PR DOES touch (the schedule edge on T-4's [needs]) vs what it does NOT touch (T-4's fact-bundle authoring contract body prose — a separable commit train), keeping the dependency fact's single-authority on the [needs …] line.

T-34's Consumer: T-22 body claim is consistent with T-22's updated [needs T-9, T-34]; no T-34 body edit needed.

PR body also updated to reflect the propagated edges + drop the now-stale "T-4 [needs …] intentionally not edited" framing.

The 3 inline BLOCKING comments from this same codex review (file-absent false positive, T-33 → T-4 edge, T-34 → T-22 edge) were each replied to and the threads resolved.

Standing by for next review pass on 7ecdf4ecd.

— sent from jolly-eagle-492

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-pro
• Commit: 7ecdf4ec · Trigger: manual
• Comparison: main @ deda6f21 ... session/jolly-eagle-492 @ 7ecdf4ec
• Conversation: View conversation

---

1. Story of the diff

This PR updates src/v4/TASKS.md to reflect the newly ratified Q1/P4 scheduling shape for v4 substrate work. The main planning move is to introduce T-33 std/model_core.dag as the shared base for primitive facts, algebra inhabitance, laws, effects, and partiality, then make T-4 LanguageModel depend on it so language fact-bundles do not re-declare that core shape (src/v4/TASKS.md:64, src/v4/TASKS.md:104-111, src/v4/TASKS.md:1196-1240). In parallel, it introduces T-34 std/host.dag as a HostModel peer of LanguageModel, and wires T-22 eval to depend on HostModel because evaluation now takes host semantics as an explicit parameter (src/v4/TASKS.md:164-170, src/v4/TASKS.md:1244-1279). Finally, it schedules T-4.15 protocol substrates for REST/GraphQL/gRPC so future omni-stack glue derivation has a named transport-model slot, but that part has a planning inconsistency around whether it is immediate parallel-fill work or deferred until T-16 activates (src/v4/TASKS.md:136, src/v4/TASKS.md:1326-1332).

2. Invariant categories

1. LAYER MODEL — Compliant. This is planning-only, not a direct .dag substrate implementation change, but the planned layering is correct: ModelCore is named as the shared base that LanguageModel and HostModel extend, and HostModel remains a distinct peer rather than being collapsed into LanguageModel (src/v4/TASKS.md:1198-1224, src/v4/TASKS.md:1245-1275).
2. INVARIANTS.md + modeling-discipline.md — Finding. src/v4/TASKS.md:136 puts T-4.15 extdeps/protocols/{rest,graphql,grpc}.dag under “Parallel fill — schedule the instant deps clear” (src/v4/TASKS.md:114) with only [needs T-3, T-26, T-4], but the task body later says “file authoring waits until omni-stack glue work activates (T-16 timeline)” (src/v4/TASKS.md:1326-1332). This violates P2 Boundary Discipline / single authority for the schedule fact: one place tells workers to dispatch after T-3/T-26/T-4, while the task body tells them not to author the files until T-16. Make the canonical dependency/placement encode the T-16 activation condition, or move T-4.15 out of the instant-parallel-fill lane.
3. CODING.md — N/A. The diff is a planning-document edit only; it adds no Rust implementation, helper, API, control flow, or comments governed by CODING.md.
4. TESTING.md — N/A. The diff schedules future substrate files and graph edges but does not change executable behavior or tests. No test is required for this planning-only edit beyond the task-graph consistency issue above.
5. LOCKED DESIGN DECISIONS — Compliant. The visible Q1 changes preserve the ratified shape: HostModel is explicitly a distinct peer of LanguageModel, both extending ModelCore, and T-22’s eval dependency is updated accordingly (src/v4/TASKS.md:1245-1279, src/v4/TASKS.md:164-170). The P4 protocol entry also states the initial single-target compiler remains out of scope (src/v4/TASKS.md:1326-1332); the remaining problem is schedule placement, not semantic rejection of the locked design.
6. TRACKED vs UNTRACKED DEBT — Compliant with one caveat covered above. T-33 and T-34 are bounded scheduled tasks with file paths, dependency edges, carrier shapes, and consumers named (src/v4/TASKS.md:1203-1216, src/v4/TASKS.md:1267-1279). T-4.15 names a future T-16 activation trigger (src/v4/TASKS.md:1329-1331), but that trigger must be reflected in the canonical schedule/dependency line to avoid becoming a conflicting scaffold; that is the blocking finding in category 2.

2.5. Top-down PM intent review

Finding. The PR’s own T-4.15 section says protocol substrate authoring is not part of the initial single-target compiler and “waits until omni-stack glue work activates (T-16 timeline)” (src/v4/TASKS.md:1326-1332). But the summary graph inserts the same task in the parallel fill area with [needs T-3, T-26, T-4] and no T-16/activation condition (src/v4/TASKS.md:114, src/v4/TASKS.md:136). A worker faithfully following the graph would start REST/GraphQL/gRPC substrate authoring as soon as those three dependencies clear, while a worker faithfully following the task body would wait for T-16; that is exactly the kind of planning artifact that can cause correct execution of the wrong work.

3. Verdict

REQUEST_CHANGES

The T-33/T-34 factoring and T-22/T-4 dependency updates look consistent with the stated substrate direction. The T-4.15 scheduling line needs correction before merge because it creates two different authorities for when protocol substrate work should start.

[briansrls]

Re: codex dashboard-only review verdict REQUEST_CHANGES (dashboard review id 15305, ran on sha 7ecdf4ecd at 2026-05-20T08:40:02Z).

Valid finding. Fixed in 5ab9d3397.

The codex review was correct: T-4.15's dependency rationale tied the transport substrate to T-4 language carriers ("protocols often parameterize over the host language's type system, e.g., gRPC service definitions reference language types") — exactly opposite to P4's "shared transport model" framing in docs/design-v4-compiler-homomorphism.md. Each transport actually declares its OWN wire-format type system (REST: HTTP bodies + headers; gRPC: protobuf primitives + messages; GraphQL: GraphQL type system). LanguageModel bindings happen downstream at composition time — T-16's omni-stack glue derivation composes LanguageModel(source) ∘ TransportModel ∘ LanguageModel(target) via the coercion fold (P4's "applied twice through a shared transport model"), NOT on the transport substrate itself. Co-locating language-binding facts on the transport substrate would have collapsed the shared model back into language-specific concerns.

Fix:

• T-4.15 [needs T-3, T-26] (T-4 dropped) in both the execution-graph block and the task-def body.
• New rationale paragraph names the language-orthogonality explicitly + cites the P4 composition shape.

T-33 / T-34 unchanged — codex's review confirmed they line up with the ratified Q1 shape.

— sent from jolly-eagle-492

[briansrls]

Re: openai-pro review verdict REQUEST_CHANGES (dashboard review id 15302, ran on sha 7ecdf4ecd at 2026-05-20T08:44:22Z, posted as PR comment 4496361522).

Valid finding. Fixed in 171353860.

The openai-pro review correctly identified a single-authority violation on T-4.15's schedule fact: the parallel-fill block placed T-4.15 under "schedule the instant deps clear" with [needs T-3, T-26], while the task-def body said "file authoring waits until omni-stack glue work activates (T-16 timeline)". Two contradictory schedule authorities for the same task.

Fix:

• T-4.15 relocated out of the "Substrate / extdeps fan-out" sub-block (the "instant parallel fill" section).
• T-4.15 placed in "Close-the-loop + late substrate" alongside T-26 (its closest semantic neighbor — both are boundary substrate awaiting downstream activation).
• The schedule line now carries an explicit scheduled-but-deferred — file authoring activates with omni-stack glue work per P4 gate, matching what the task-def body says. Single authority restored — same fact in both places.
• One-line pointer left in the old position so a reader scanning the parallel-fill block still finds T-4.15.

The earlier codex 5ab9d3397 fix (T-4.15 dropping its T-4 dependency for language-orthogonality) stands; this commit only fixes the schedule-placement contradiction.

— sent from jolly-eagle-492

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-pro
• Commit: 17135386 · Trigger: manual
• Comparison: main @ deda6f21 ... session/jolly-eagle-492 @ 17135386
• Conversation: View conversation

---

1. Story of the diff

This PR is a src/v4/TASKS.md scheduling and scope update, not an implementation change. It adds T-33 std/model_core.dag as the shared substrate base for primitive facts, algebra inhabitance, laws, effects, and partiality, then wires that into T-4 LanguageModel as a hard prerequisite. It adds T-34 std/host.dag as the HostModel peer of LanguageModel, then updates T-22 eval so the interpreter depends on HostModel before authoring. It also relocates T-4.15 protocols out of instant substrate fan-out and into a scheduled-but-deferred transport lane, with activation tied to omni-stack/T-16 glue work rather than the initial single-target compiler.

The load-bearing mechanism is the graph edit: ModelCore becomes the common source of truth for host/language primitive facts, HostModel gates eval, and protocols are explicitly language-orthogonal transport substrate. That direction matches the v4-active framing in the thesis and Pure Bootstrap Zero context: v4 is the active operational substrate program, and the v4 workflow/bootstrap discipline is governed by the zero-floor design. chatgpt-review-09cf1618-8ea6-4d…

chatgpt-review-556d355f-3e49-42…

2. Invariant categories

1. LAYER MODEL — Compliant. The diff touches substrate planning, not concrete .dag substrate carriers; the planned shape keeps ModelCore as the shared base consumed by both LanguageModel and HostModel at src/v4/TASKS.md:1212-1238, which is the right substrate-layer move rather than duplicating primitive/algebra facts in each target model.
2. INVARIANTS.md + modeling-discipline.md — Compliant. P1/P2 are handled by making ModelCore the single shared carrier for primitive facts, algebra inhabitance, laws, effects, and partiality at src/v4/TASKS.md:1220-1238; T-4.15 also avoids the no-template/derived-homomorphism failure mode by stating transport is bidirectional substrate consumed by ingest and emit, “NOT a string template,” at src/v4/TASKS.md:1321-1324. This aligns with the modeling discipline’s fact-bundle and no-string-template rule. chatgpt-review-3e234581-e536-40…
3. CODING.md — N/A. The diff is a Markdown planning/task graph edit only; it introduces no Rust functions, helpers, comments inside code, error shapes, or APIs.
4. TESTING.md — N/A. No executable code, .dag carrier, test harness, or runtime behavior changes in this diff. A task-graph-only Markdown change does not need a new hermetic or behavior-level test unless the repo has a generated TASKS graph checker, which is not present in the attached context.
5. LOCKED DESIGN DECISIONS — Compliant. The diff does not alter a locked design doc; it records operator-ratified Q1/P4 decisions in TASKS and keeps the key semantics explicit: HostModel is a peer of LanguageModel, not a LanguageModel variant (src/v4/TASKS.md:1258-1293), and protocols are language-orthogonal transport substrate rather than T-4 language facts (src/v4/TASKS.md:1329-1341).
6. TRACKED vs UNTRACKED DEBT — Compliant. The new deferred substrate is bounded and has a named activation trigger: T-4.15 is scheduled-but-deferred, depends on T-3/T-26, and file authoring activates with omni-stack glue/T-16 rather than the initial single-target compiler (src/v4/TASKS.md:1300-1355). That is a tracked bridge, not an open-ended TODO.

2.5. Top-down PM intent review

Finding — scheduling semantics are diluted by placing T-33 in a “late substrate” bucket. The project’s top-level model treats .dag as a dependency graph whose schedule is mechanically chosen from that same source. chatgpt-review-09cf1618-8ea6-4d…

The PR’s own side-branch text says T-33 is one of five non-critical-path feeders that gate T-4 → T-9, and that these feeders are “watch items,” not slack-having parallel fill; if any slips, the side branch becomes critical (src/v4/TASKS.md:75-78). But the same PR inserts T-33 under the existing “Close-the-loop + late substrate” bucket (src/v4/TASKS.md:202, src/v4/TASKS.md:216-219). That bucket placement gives a worker the opposite scheduling signal from the dependency narrative: “late/close-loop” instead of “start as soon as T-1/T-2/T-3 clear because T-4 is blocked on it.” Move T-33 out of the late bucket, split the bucket, or rename the heading so the graph does not encode contradictory priority.

3. Verdict

REQUEST_CHANGES. The modeling and dependency additions are directionally sound, but T-33’s placement under “Close-the-loop + late substrate” contradicts the PR’s own hard-prerequisite/watch-item language. This is a planning artifact that can cause workers to schedule ModelCore too late, so it should be fixed before merge.

[briansrls]

Re: openai-pro review verdict REQUEST_CHANGES (dashboard review id 15306, ran on sha 171353860 at 2026-05-20T08:58:33Z, posted as PR comment 4496470211).

Valid finding. Fixed in 186a0c362.

openai-pro correctly identified contradictory priority signals on T-33: I had placed it in BOTH the side-branch feeders block (correctly, as a watch-item T-4 prerequisite that goes critical if it slips — src/v4/TASKS.md:75-78 + :104-110) AND the "Close-the-loop + late substrate" block (incorrectly — that bucket implies slack/late, opposite to the watch-item framing). A worker reading the late-substrate bucket would have scheduled T-33 late.

Fix: T-33 removed from "Close-the-loop + late substrate". The side-branch feeders block at the top of the file is now the single authoritative placement, carrying the correct watch-item semantics. One-line pointer left in the late-substrate block so a reader scanning that section still finds T-33 quickly.

T-34 and T-4.15 stay in "Close-the-loop + late substrate":

• T-34 feeds T-22 (eval), which is in "Interpreter + lens dimensions", not the critical path — no contradiction.
• T-4.15 is explicitly scheduled-but-deferred (activates with omni-stack glue per P4) — the bucket signal matches the deferral gate.

The directional comments at the start of openai-pro's review (sections 1, 2.1 LAYER MODEL, 2.5 PM intent for T-4.15, 2 TRACKED-vs-UNTRACKED) all read as compliant; only the T-33 placement was the blocking issue, now resolved.

— sent from jolly-eagle-492

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 186a0c36 · Trigger: schedule
• Thinking: 274s wall

---

BLOCKING (1)

Root Cause

• src/v4/TASKS.md T-4.15 was relocated into late substrate without updating T-16's dependency contract → add T-4.15 to T-16's [needs] list or remove the claim that T-16 consumes TransportModel.

⚠️ The prior comments are addressed, but the protocol substrate still needs its T-16 consumer edge reconciled.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: d6063235 · Trigger: schedule
• Thinking: 268s wall

---

BLOCKING (3)

Root Cause

• src/v4/TASKS.md Ratified Q1 and P4 are scheduled from an authority document that is not present in the reviewable tree → add the design doc to the PR or anchor these rows to an existing checkable authority.
• src/v4/TASKS.md The T-4.15 task body was reconciled to remove the T-16 consumer edge, but the execution-graph row kept the old T-16 wording → change the row to future-expanded omni-stack glue work or add the dependency edge everywhere.
• src/v4/TASKS.md The T-22 scheduling edge was updated, but the T-22 task definition remained on the old eval signature → update the T-22 I/O and modeling contract to include HostModel.

⚠️ The dependency edges are mostly reconciled, but the new substrate rows still rely on an absent authority and leave two task-contract drifts.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 56256abb · Trigger: schedule
• Thinking: 229s wall

---

BLOCKING (2)

Root Cause

• src/v4/TASKS.md Q1 split HostModel from LanguageModel, but T-22 was updated only at the signature/dependency edge → rewrite the T-22 concept-unification bullet so eval's runtime authority is HostModel/ModelCore, with any LanguageModel fact explicitly carried through InferredTree or a named parameter.
• src/v4/TASKS.md T-33 prerequisites are recorded only in derived side-branch prose → add an authoritative [needs T-1, T-2, T-3] contract to the T-33 task definition.

⚠️ Two task-contract drifts remain in the newly added planning rows.

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-pro
• Commit: c8f3acdc · Trigger: manual
• Comparison: main @ deda6f21 ... session/jolly-eagle-492 @ c8f3acdc
• Conversation: View conversation

---

1. Story of the diff

This PR updates src/v4/TASKS.md to turn the Q1/P4 design ratifications into explicit v4 scheduling authority. The main move is factoring the shared primitive/algebra/law/effect substrate into T-33 std/model_core.dag, then making T-4 LanguageModel and T-34 HostModel depend on that shared base rather than duplicating model facts independently (src/v4/TASKS.md:104, src/v4/TASKS.md:1212). It then updates T-22 eval so the evaluator takes a HostModel parameter, replacing the older “eval reads LanguageModel” framing with a host-runtime carrier that is still grounded through ModelCore (src/v4/TASKS.md:167, src/v4/TASKS.md:916, src/v4/TASKS.md:919).

The second load-bearing move is adding T-4.15 protocols substrate while deliberately keeping it out of the initial single-target compiler and out of T-16’s current dependency edge. The PR treats REST/GraphQL/gRPC as a future transport substrate with its own wire-format type system, composed later with LanguageModel via the coercion fold, rather than smuggling transport facts into language targets or OpenAPI demo scope (src/v4/TASKS.md:205, src/v4/TASKS.md:1343, src/v4/TASKS.md:1364). Net effect: the task graph is more explicit about what must exist before language modeling and eval can land, while avoiding premature omni-stack transport implementation.

2. Invariant categories

1. LAYER MODEL

Compliant — the diff is planning-layer only, but it preserves substrate layering rather than flattening it: ModelCore is the shared base for both LanguageModel and HostModel (src/v4/TASKS.md:1214), HostModel is explicitly a peer rather than a LanguageModel variant (src/v4/TASKS.md:1267–src/v4/TASKS.md:1272), and protocol substrate is kept language-orthogonal rather than folded into T-4 (src/v4/TASKS.md:1343–src/v4/TASKS.md:1356).

2. INVARIANTS.md + modeling-discipline.md

Compliant — P2 single-authority is handled by putting the T-4 dependency fact on the T-4 [needs …] line and explicitly saying T-33 prose is not the authority for that edge (src/v4/TASKS.md:1248–src/v4/TASKS.md:1258). P1 modeling faithfulness is also reinforced: T-33’s carrier shape requires primitive fact-bundles, algebra inhabitance, laws, effects, and partiality semantics instead of hollow aliases (src/v4/TASKS.md:1228–src/v4/TASKS.md:1238), aligning with the uploaded thesis requirement that target primitives be structurally modeled with algebra inhabitance rather than name-keyed shortcuts (THESIS.md:181).

3. CODING.md

N/A — no Rust implementation, helper API, method shape, error carrier, or code comment discipline is touched; the diff only edits the planning document src/v4/TASKS.md.

4. TESTING.md

N/A — no executable behavior, generated artifact, Rust harness, or .dag carrier is changed. The PR is a schedule/authority edit; there is no new testable compiler behavior to cover in this diff.

5. LOCKED DESIGN DECISIONS

Compliant — the diff explicitly names the ratified design decisions it is applying instead of silently diverging: Q1 for ModelCore/HostModel (src/v4/TASKS.md:1213, src/v4/TASKS.md:1267) and P4 for protocols as orthogonal glue substrate (src/v4/TASKS.md:1308–src/v4/TASKS.md:1316). It also does not weaken the uploaded Pure Bootstrap to Zero authority: no hand-authored Rust surface is added, and the v4 zero-floor framing remains untouched (design-pure-bootstrap-zero.md:3–design-pure-bootstrap-zero.md:7).

6. TRACKED vs UNTRACKED DEBT

Compliant — the one explicit deferral, T-4.15, is documented with bounds and an activation trigger: it is “scheduled-but-deferred,” activates with future omni-stack glue work, is not part of the initial single-target compiler, and does not enter T-16’s current [needs] edge (src/v4/TASKS.md:207–src/v4/TASKS.md:217, src/v4/TASKS.md:1364–src/v4/TASKS.md:1371). T-33 and T-34 are scheduled tasks rather than scaffolds; each has named dependencies, file targets, and consumers (src/v4/TASKS.md:1219–src/v4/TASKS.md:1226, src/v4/TASKS.md:1289–src/v4/TASKS.md:1301).

2.5. Top-down PM intent review

Compliant — this PR preserves the project-level intent rather than diluting it. The thesis requires grounded target primitive modeling and algebra inhabitance (THESIS.md:181), explicit substrate structure (THESIS.md:206–THESIS.md:210), and omni-emission without blurring compiler language targets and user-artifact targets (THESIS.md:220–THESIS.md:228). The diff supports that by adding ModelCore before LanguageModel fact-bundles (src/v4/TASKS.md:321–src/v4/TASKS.md:323), splitting runtime execution facts into HostModel instead of overloading LanguageModel (src/v4/TASKS.md:916–src/v4/TASKS.md:919), and keeping transport substrate deferred/orthogonal until true omni-stack glue work activates (src/v4/TASKS.md:1349–src/v4/TASKS.md:1356). I do not see a must-have target being postponed without reconciliation or a dissolution goal becoming permanent scaffolding.

3. Verdict

APPROVE. The PR is a planning/authority update that makes the dependency graph stricter and more explicit without adding implementation debt, parallel substrate authority, or unbounded deferral. I found no blocking invariant, testing, or PM-intent issue in the changed lines.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: c8f3acdc · Trigger: schedule
• Thinking: 238s wall

---

BLOCKING (1)

Root Cause

• src/v4/TASKS.md HostModel scheduling was added without landing its cited ratified design authority → add docs/design-v4-compiler-homomorphism.md or make the HostModel carrier decision self-contained with a checkable source.

⚠️ One new HostModel authority gap remains.

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-pro
• Commit: e39305e3 · Trigger: manual
• Comparison: main @ f4a38301 ... session/jolly-eagle-492 @ e39305e3
• Conversation: View conversation

---

1. Story of the diff

This PR is a planning-graph correction to src/v4/TASKS.md, not an implementation PR. It incorporates the newly ratified Q1/P4 substrate split by adding T-33 std/model_core.dag as the shared base for LanguageModel and HostModel, adding T-34 std/host.dag as the host-runtime peer of LanguageModel, and moving the protocols work into a explicitly deferred T-4.15 transport substrate lane. The load-bearing graph edits are: T-33 becomes a hard T-4 prerequisite in the side branch (src/v4/TASKS.md:64, src/v4/TASKS.md:111, src/v4/TASKS.md:321); T-22 now depends on T-34 because eval takes a HostModel parameter (src/v4/TASKS.md:167, src/v4/TASKS.md:921); and T-4.15 is deliberately not made a current T-16 dependency because the current T-16 scope remains OpenAPI/WireContract, not full REST/GraphQL/gRPC transport substrate (src/v4/TASKS.md:1313, src/v4/TASKS.md:1321, src/v4/TASKS.md:1369).

2. Invariant categories

1. LAYER MODEL (substrate vs implementation).

Compliant — this is substrate-planning work, but it does not author substrate carriers yet. The planning shape preserves layer separation by factoring shared primitive/algebra/law facts into ModelCore (src/v4/TASKS.md:1217, src/v4/TASKS.md:1233) and keeping HostModel as a peer of LanguageModel rather than a variant or emitter concern (src/v4/TASKS.md:1271, src/v4/TASKS.md:1298).

1. INVARIANTS.md + modeling-discipline.md.

Compliant — single authority / no duplicate representation is the main principle exercised here. T-33 is explicitly introduced so LanguageModel and HostModel do not each redeclare primitive facts, algebra inhabitance, and laws (src/v4/TASKS.md:1219, src/v4/TASKS.md:1245). The T-22 correction also routes eval through HostModel rather than reusing the language/emitter model for host runtime semantics (src/v4/TASKS.md:924), which avoids a cross-layer authority collapse.

1. CODING.md.

N/A — the diff is a Markdown task-plan edit only; it adds no Rust implementation, helper functions, APIs, methods, error shapes, or comments in code.

1. TESTING.md.

N/A — no executable code or test harness changed. For this planning-only graph edit, the relevant “test” discipline is dependency/authority clarity; the PR updates the dependency graph and task bodies rather than needing a Rust or .dag test.

1. LOCKED DESIGN DECISIONS.

Compliant — the PR references ratified design decisions and reflects them consistently rather than silently diverging. Q1 is applied by adding ModelCore/HostModel and changing eval’s contract to eval(InferredTree, HostModel, Inputs) (src/v4/TASKS.md:921, src/v4/TASKS.md:1272). P4 is applied by treating protocols as transport substrate for future omni-stack glue, not as a current T-16 dependency (src/v4/TASKS.md:1314, src/v4/TASKS.md:1321, src/v4/TASKS.md:1348).

1. TRACKED vs UNTRACKED DEBT.

Compliant — the only obvious deferred surface is T-4.15, and it is bounded rather than left as generic “later” work. The overview marks it scheduled-but-deferred and names the activation condition as omni-stack glue work (src/v4/TASKS.md:138, src/v4/TASKS.md:205), while the task body repeats the scope boundary that it is out of the initial single-target compiler and not part of today’s T-16 scope (src/v4/TASKS.md:1369, src/v4/TASKS.md:1373). T-33 and T-34 are not scaffolds; they are scheduled substrate tasks with concrete file targets and dependency edges (src/v4/TASKS.md:1230, src/v4/TASKS.md:1279, src/v4/TASKS.md:1294).

2.5. Top-down PM intent review

Compliant — the PR preserves the high-level thesis direction. The thesis says each target is modeled once in shared vocabulary and translations are derived homomorphisms rather than hand-authored adapters (THESIS.md:91); this PR reinforces that by moving shared primitive/algebra/law facts into ModelCore and making transport glue a composed homomorphism over a transport model (src/v4/TASKS.md:1219, src/v4/TASKS.md:1318). The thesis also says dag run is the primary execution path (THESIS.md:233); the PR strengthens that path by making eval read HostModel explicitly instead of treating runtime execution as an emitter/language-spec afterthought (src/v4/TASKS.md:167, src/v4/TASKS.md:924). I did not see a concrete diff-cited mismatch where a must-have target became optional, a dissolution goal became permanent scaffolding, or handwritten implementation was introduced where generated/data-authored structure is required.

3. Verdict

APPROVE

The diff is a planning correction that tightens the substrate dependency graph and avoids the prior HostModel/LanguageModel conflation. The one deferred lane, T-4.15, is explicitly bounded to future omni-stack glue rather than silently added to the current T-16 path, so I do not see a blocking invariant violation.