docs: OGAR as a semantic transpiler (v3 sink) + ClassView Askama bitmask field-view#139
Conversation
… compiled) Capture the consumer-side pattern: render a ClassView as an HTML field view the way Redmine renders model fields with ERB, but compiled, type-safe, JSON-free, and driven by a u64 field mask instead of per-field conditionals. One generic Askama loop over a mask-filtered, generated FieldDesc[] table (no if-noise). The render-mask = the read-mask: the same selector that chooses which facets to decode chooses which fields to show (the render-side twin of the compiled-ClassView spine). Discipline: field/idx/bit from ONE generated source (I-LEGACY-API-FEATURE-GATED). Residual novel fields -> generic Vec view or build.rs codegen. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RhpwkHGgia2TuDFvdnuQdE
…bstrate Capture the synthesis before it dilutes (PROPOSAL, pending 5+3 hardening): - OGAR = semantic transpiler; v3 substrate = the sink (s1.5 'sinks into OGAR'). - Two front-ends: serialization (syntax incidental) vs semantic-application (Odoo/Redmine/OpenProject) where syntax is PRESERVED (semantic->semantic). - Three arms: THINK/Class+fieldview structure-preserving; DO/ActionDef re-imagined as classid-keyed adapters = the God-object/SoC dividend; membrane/KausalSpec. - The universal selector: one generated bitmask = read=query=render=version=auth. - classid/HHTL cascade = the semantic zoom + renderer dispatch by signature. - living ractor mailbox-owned SoA; JSON-free end to end. - Honest edges flagged for grading (generated-bits iron rule, Biolink upper layer, renderer scale ceiling). All claims [G]/[H]/[S]-gradable. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RhpwkHGgia2TuDFvdnuQdE
- ractor = RUNTIME ownership guarantee (mailbox serializes at runtime); NOT a compile-time borrow proof. n actors x up-to-64k partitions run concurrently, race-free by single-owner-per-partition. - field count is a representation parameter ([u64;N] -> 256+), NOT a blocker. - transpile is a landing zone (a destination, ahead of code): claims tagged CODED-today vs LANDING-ZONE rather than downgraded for being unbuilt. - payoff stated: Odoo ERP at the cost of an import; past Palantir Foundry (whole-app transpile incl. DO arm -> ractor StateMachines, not rebuilt). - factual fixes kept: DO arm = ractor StateMachine (not thin adapter); selector = N3 presence FieldMask doing 3 jobs (read/render/auth-projection); query rides a 2nd physical-tenant basis; version = radix lookup; auth verdict/scope = I-K8 folds. Two FieldMask bases must be type-separated. - Firewall: Arrow/HTML egress is the OUTER crossing (ADR-022 s5 + D-KV-RENDER), inner path moves keys; no inner serialization. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RhpwkHGgia2TuDFvdnuQdE
…Writer = throughput Correct the concurrency framing into two distinct mechanisms: - ractor: compile-time ownership guarantee (mailbox-as-owner, Rust move semantics, UB becomes a compile error; E-CE64-MB-4). The SAFETY. n compile-time-owned partitions update in parallel, race-free by construction. - Lance BatchWriter: the THROUGHPUT, and where cognition hides. On initiation, while one SoA batch lowers (persists to Lance), the next batch already thinks (computes) -- compute/persist overlap on a double-buffered SoA; cost ~ max(compute, write), not compute + write. ractor is no longer the straw puppet carrying the concurrency story; it is the compile-time safety, the BatchWriter overlap is the latency-hidden throughput. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RhpwkHGgia2TuDFvdnuQdE
…aph-llm / rig) - DO arm: ActionDef -> StateMachine REALIZED AS an rs-graph-llm workflow graph; control-flow maps node-for-node => DO is lossless; F1 becomes a graph-equivalence check. - New section 6 Orchestration & cognition: ActionDef -> rs-graph-llm graph -> ractor (compile-time execution ownership) -> Lance BatchWriter (overlap); rig = RAG + LLM boundary + thinking-replay; LLM API = removable training wheels; replay distils LLM traces into offline DeepNSM cognition. - Edges: pin the lifecycle owner; use the REAL rs-graph-llm (not the graph-flow stub); offline-replay fidelity is LANDING-ZONE. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RhpwkHGgia2TuDFvdnuQdE
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7bdb0583f3
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| |---|---|---|---| | ||
| | **read** (facet decode) | N3 presence `FieldMask` | compiled reader | **CODED** (`class_view.rs`) | | ||
| | **render** (field view) | same N3 `FieldMask`, bit-gated loop | compiled template | **CODED** (`render_rows`) | | ||
| | **auth field-projection** | same N3 `FieldMask`, OR-folded | role × lo-u16 grant lattice | **CODED** (`lance-graph-rbac::authorize_scoped`, `permission.rs::projection`) | |
There was a problem hiding this comment.
In this repo's current canon, the mask-bearing auth path is still not shipped: README.md:76 says lance-graph-rbac::authorize is [H]/gated, and docs/CLASSID-RBAC-KEYSTONE-SPEC.md:197-203 says the projecting Allow { scope, mask } return remains CONJECTURE; a repo-wide rg authorize_scoped only finds this new doc. Marking auth field-projection as CODED here will lead consumers to depend on a scope/projection API that the existing contract still says is pending.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed. Downgraded auth-field-projection from CODED to [H] cross-repo lag: the authorize_scoped/field_mask code lives in the sibling lance-graph-rbac, but OGAR canon (CLASSID-RBAC-KEYSTONE-SPEC I-K8, README [H]) still grades the projecting Allow{scope,mask} CONJECTURE, so the doc no longer claims it as shipped-from-OGAR. Also added the part that is real and shippable today — verb × class action authorization (I-K8 axis-1) — as the DO handler's link to the existing lance-graph-rbac gate, with field-projection (axis-4) named as the separate pending leg.
Generated by Claude Code
| 1. **Field count is a representation parameter, not a ceiling.** A presence mask | ||
| is `[u64; N]` — `[u64; 4]` = 256 fields, wider as needed; Odoo's 109-field | ||
| `account.move` is a `[u64; 2]`, not a wall. Widen the basis; the algebra is |
There was a problem hiding this comment.
The current implementation and integration plan still enforce a single-u64 mask: crates/ogar-class-view/src/lib.rs:353-364 has field_basis_fits_in_one_u64_mask, and docs/INTEGRATION-MAP.md:111 / :345-346 call the 109-field Odoo case a real >64-field blocker until X7/F14. Saying the presence mask "is [u64; N]" and that account.move is not a wall hides the failure mode where fields beyond 64 are silently dropped in today's render path.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed — kept both truths. The doc now states the widening to [u64; N] is a cheap ClassView-index change (no architectural wall), and that shipped code is still single-u64 (field_basis_fits_in_one_u64_mask guard; INTEGRATION-MAP X7/F14), so fields beyond 64 silently drop in the current render path until the widening lands. Framed as a tracked limitation (X7/F14), not hidden and not an architectural ceiling.
Generated by Claude Code
- auth field-projection: downgrade from CODED to [H] cross-repo lag (coded in
sibling lance-graph-rbac, but OGAR canon/CLASSID-RBAC-KEYSTONE-SPEC still
grades the projecting Allow{scope,mask} CONJECTURE). Don't depend on it as
shipped-from-OGAR. (codex P2 #1)
- field-count: keep both truths — widening to [u64;N] is a cheap ClassView
index (no architectural wall), but shipped code is single-u64
(field_basis_fits_in_one_u64_mask guard; X7/F14), so fields >64 drop TODAY
until the widening lands. Tracked limitation, not hidden. (codex P2 #2)
- action authorization (verb x class) rides the existing lance-graph-rbac
gate (I-K8 axis-1) = the easiest real link; field-projection (axis-4) is the
separate pending leg. Added to the s3 table and s6 orchestration layer.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RhpwkHGgia2TuDFvdnuQdE
…st rbac reuse; add Tuersteher AGI-gatekeeper payoff - §3 action-auth row + prose: required_role is a compile-time const on the ActionDef in lance-graph-ogar/src/actions.rs ([G]-tested), most-shipped of the auth jobs; lance-graph-rbac supplies only the runtime grant map. - §6 Authorization bullet: RBAC hardcoded on the class = the Tuersteher floor (def-match -> RBAC -> Libet -> MUL in commit); cognition decides above it, never lowers it. - §6 punchline: capability-bounding not alignment-bounding -> thinking can upgrade to AGI, OGAR still gatekeeps every action structurally (closed DO surface + unforgeable commit gate). The legible pitch for compliance buyers. - cross-ref ACTIONHANDLER-TURSTEHER.md + actions.rs receipt.
Extend the Redmine-ERB ClassView bitmask doc (PR #139) with the operator's 2026-06-29 refinements, now tested in ruff: - 'Wide classes — the quadruplet and bucket chaining': the u64 mask is ONE bucket (64 fields); a wide class (Odoo account.move ~109 fields) chains up to 4 buckets = the [u64; 4] quadruplet = 256, and the selected() loop is bucket-agnostic (idx/64, idx%64). Clean separation overflow automation; > 256 is a god object that splits into a second ClassView, never a wider mask. Pinned + tested in ruff_spo_address::soc (FIELD_MASK_CAP = 256, field_mask_buckets, 109-field -> Duplication). lance-graph FieldMask is the eventual matching expansion. - 'Simple rules': if it's a template it's probably a ClassView; deduplicate routes (N field-subset routes = one templated render with N masks); < 256 is clean, >= 256 is the god-object SoC split signal. - Cross-refs to ruff_spo_address::soc and canonical_node::GUIDS_PER_NODE (the node-level twin of the same clean/SoC-over-packed doctrine). Co-Authored-By: Claude <noreply@anthropic.com>
1 participant
What
Two related docs capturing an architecture synthesis before it diluted, then hardened by the 5+3 pass (5 research savants + 3 brutally-honest reviewers) and corrected by operator review.
docs/CLASSVIEW-FIELDVIEW-ASKAMA-BITMASK.md— the render-side pattern: render aClassViewas an HTML field view the Redmine-ERB way, but compiled, type-safe, JSON-free, and driven by a presenceFieldMaskinstead of per-fieldifs. One generic Askama loop over a mask-filtered, generatedFieldDesc[];selected()/unselected(); the render-mask is the read-mask.docs/OGAR-SEMANTIC-TRANSPILER-CAPSTONE.md— the capstone: OGAR is a semantic transpiler; the v3 substrate is its sink. Tagged CODED-today vs LANDING-ZONE (a destination is not weaker for being unbuilt).Headline
An Odoo ERP at the cost of an import — in the same substrate as the anatomy and the genome. Whole-app transpilation into a compiled, living substrate (THINK arm and DO arm), not data-integration-into-an-ontology. Past Palantir Foundry, not chasing it.
What the 5+3 hardening + operator review changed
BatchWriteroverlap = the throughput (while one SoA batch lowers, the next already thinks — cost ≈max(compute, write)). ractor is not the straw puppet carrying concurrency.ActionDef→StateMachinerealized as an rs-graph-llm workflow graph → behaviour maps node-for-node, so DO is lossless; gate F1 becomes a graph-equivalence check. New §6 Orchestration & cognition: rs-graph-llm (action handler) → ractor (compile-time execution) → Lance (overlap); rig = RAG + LLM boundary + thinking-replay; LLM API = removable training wheels whose replay distils into offline DeepNSM cognition.FieldMaskgenuinely does three jobs (read + render + auth-field-projection, shipped/tested), not five —queryrides a second physical-tenant basis,versionis a radix lookup, auth verdict/scope are RBAC keystone I-K8 folds. The twoFieldMaskbases must be type-separated (FieldMask<Logical/Physical>) — a real latent defect surfaced by the pass.[u64; N](256+), not a wall.StructuralSignatureis a group-by key, no renderer exists; coordinate-free graphs need address assignment) → LANDING-ZONE; "holy grail / genomics simply lowers in" → roadmap.D-KV-RENDER).Status / follow-ups (not in this PR)
These are PROPOSAL / landing-zone docs, not yet canon-pinned. If accepted, the mirror duty (separate change): new D-ids for the transpiler/selector/front-end framings; a render-egress row in
INTEGRATION-MAP+THE-FIREWALL§4; and the first brick — render an existing OpenProjectWorkPackage/ medcarePatientrow throughClassView::render_rows(class, FieldMask), proving one N3 mask drives projection and render (~60 LOC; coded surface, not genomics).Doc-only. No code, no canon-ledger edits, no PII labels, no model identifiers.
🤖 Generated with Claude Code
https://claude.ai/code/session_01RhpwkHGgia2TuDFvdnuQdE
Generated by Claude Code