diff --git a/AGENTS.md b/AGENTS.md
index 3760d09..251d0f5 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -55,6 +55,7 @@ Full diagram and concurrency model: [docs/architecture.md](docs/architecture.md)
 
 | Area | Read |
 |---|---|
+| **Architectural invariants & deny-list (read before any non-trivial proposal or review)** | **[docs/invariants.md](docs/invariants.md)** |
 | Architecture, L1/L2 framing, concurrency model | [docs/architecture.md](docs/architecture.md) |
 | Storage layout, `__manifest` schema, URI schemes, S3 env vars | [docs/storage.md](docs/storage.md) |
 | `.pg` schema language, types, constraints, annotations, migration planning | [docs/schema-language.md](docs/schema-language.md) |
@@ -83,7 +84,7 @@ Full diagram and concurrency model: [docs/architecture.md](docs/architecture.md)
 
 ## Always-on rules (load these into your working memory)
 
-These invariants need to be in scope on every change — they're the ones that quietly break if forgotten.
+These invariants need to be in scope on every change — they're the ones that quietly break if forgotten. The full architectural invariants and deny-list live in [docs/invariants.md](docs/invariants.md); §IX (deny-list) is the fastest first-pass when reviewing any change.
 
 1. **`__manifest` is the atomic-publish boundary.** Multi-dataset commits flip via a single `ManifestBatchPublisher` write. Don't introduce code paths that publish per sub-table outside the batch publisher — you'll lose snapshot isolation across tables.
 2. **`nearest($x.vec, $q)` requires a `LIMIT`.** The compiler enforces it, but if you're touching the query lowering or executor, don't break this rule. ANN without a limit is unbounded.
@@ -96,6 +97,29 @@ These invariants need to be in scope on every change — they're the ones that q
 9. **Indexes are built on the branch head, not on a snapshot.** Reads always see the current index state. Lazy fork: a branch that hasn't mutated a sub-table reuses the source's index until the first write.
 10. **Stable type IDs survive renames.** Schema migration uses `stable_type_id` (kind+name hashed at first sight). Don't mint new IDs on rename.
 
+### Deny-list (fast-pass review filter — full reasoning in [docs/invariants.md §IX](docs/invariants.md))
+
+If a proposal fits one of these, the burden is on the proposer to justify why this case is the exception:
+
+- Synchronous-inline index updates for indexes expensive to build (vector ANN, FTS) — use the reconciler pattern.
+- Custom WAL / transaction manager / buffer pool — Lance owns these.
+- Job queue for state derivable from manifest — reconciler pattern instead.
+- Per-feature lowering for shapes that share a structure (interfaces, wildcards, alternation) — use one mechanism.
+- Eager materialization of cross-products in multi-hop — factorize; flatten only when needed.
+- Ad-hoc IN-list filtering when SIP fits.
+- String-flattened SQL filter generation when structured pushdown is available.
+- In-process-only `Dataset` impls — `Send + Sync`, remote descriptors.
+- Cost-blind plan choice — lowering-order execution is not a planner.
+- Hidden statistics — if a metric matters for plan choice, it must be exposed through the trait surface.
+- Side-channels for query semantics — search modes, mutations, polymorphism are first-class IR concepts.
+- Discarding rank in retrieval — score and rank propagate as columns.
+- State that drifts from the manifest — derive from observable state.
+- Cloud-only correctness fixes — correctness is always OSS.
+- Forking the codebase for Cloud — trait-extension only.
+- Hand-rolling something Lance already does — check the spec first.
+- Mutating in place state that should be immutable (Lance fragments, index segments) — new segments instead.
+- Silent failures — OOM, timeout, partial result must all be surfaced and bounded.
+
 ---
 
 ## Quick-reference flows
@@ -172,6 +196,8 @@ omnigraph policy explain --actor act-alice --action change --branch main
 
 When you change something user-visible, **update the relevant `docs/<area>.md` in the same change**. Pointers from this file to that doc must keep working — CI enforces cross-link integrity via `scripts/check-agents-md.sh`.
 
+When proposing or reviewing a non-trivial change, walk [docs/invariants.md](docs/invariants.md) — at minimum the §IX deny-list and §X review checklist. Add to the deny-list when a new anti-pattern surfaces; relaxing an invariant requires the same review process as code.
+
 Rules:
 
 1. **Update in the same PR.** New endpoint, query function, CLI flag, env var, constant, schema construct, or invariant: update both the source code and the doc in the same change. Never split documentation drift into a follow-up.
diff --git a/docs/invariants.md b/docs/invariants.md
new file mode 100644
index 0000000..eecf4dc
--- /dev/null
+++ b/docs/invariants.md
@@ -0,0 +1,162 @@
+# Architectural Invariants & Policies
+
+**Type:** Reference / standing document
+**Status:** Living — updated as decisions accrue
+**Audience:** anyone proposing, reviewing, or implementing a change to any part of OmniGraph
+
+This document captures the invariants that hold across all of OmniGraph — storage, engine, server, schema, indexing, observability, and the OSS / Cloud product split. New RFCs, designs, and implementations are checked against this list.
+
+These are not query-engine-specific. They apply to every layer.
+
+> **Note on numbering.** §IV (Additivity / migration) and §VIII (OSS / Cloud kernel-product split) are referenced from §X but not yet drafted in this revision. Pending upstream fill-in.
+
+## How to use
+
+- **Writing an RFC or design proposal:** walk through the relevant sections and state how the proposal upholds each invariant — or why a documented exception is justified.
+- **Reviewing a PR or design:** scan for invariants the change might violate. The deny-list (§IX) is the fastest first pass.
+- **Debating a tradeoff:** invoke the relevant invariant and check whether the tradeoff respects it. Invariants here are load-bearing; breaking one is rare and requires explicit justification in the proposal.
+- **Updating this document:** add to the deny-list freely. Removing or relaxing an invariant requires the same review process as any other architectural decision.
+
+## I. Substrate boundaries — what we don't build
+
+The most important question for any new component is "does the substrate already do this?" If yes, we don't.
+
+1. **Lance is the storage substrate.** We do not build a competing storage format. We build above Lance via a trait. Where we want format-level behavior Lance doesn't have, we propose it upstream or work around it.
+   *Check:* Does this proposal introduce a parallel storage format, custom on-disk pages, custom serialization?
+
+2. **No own-format WAL or transaction manager.** Lance manifest MVCC plus (eventually) MemWAL is the durability story. We do not write durability code.
+   *Check:* Does this proposal track its own write log, recovery state, or transaction journal?
+
+3. **No own buffer pool.** Lance + `object_store` + Lance's scan-aware cache cover our needs.
+   *Check:* Does this proposal cache Arrow batches or pages outside Lance's cache?
+
+4. **The runtime substrate provides relational machinery.** Whether we choose DataFusion (current prior) or a custom executor, we do not rebuild joins, aggregations, parallelism, spill. We build only graph- or multi-modal-shaped operators on top.
+   *Check:* Are we extending the substrate via traits, or reimplementing parts of it?
+
+5. **Don't replace Lance's index lifecycle.** New fragments enter without index coverage; reads union indexed and scan paths via `fragment_bitmap`. Our reconciler observes the same primitive; we don't replicate it.
+   *Check:* Does this proposal maintain its own version of "what's indexed and what isn't" parallel to Lance's?
+
+## II. Layering — the seams that hold
+
+6. **The IR is the contract between frontend and backend.** Frontends emit IR; planner / executor consume it. No frontend logic leaks downward; no executor concerns leak upward.
+   *Check:* Does the proposal add to the IR, or to a layer? If to a layer, does it cross another layer's concern?
+
+7. **Capabilities and statistics flow upward; data flows downward.** Lower layers expose what they can do (capabilities) and what they know (statistics). Upper layers consume both. Methods alone are insufficient — methods without capability advertisement force one-size-fits-all plans.
+   *Check:* When adding a method to a layer trait, did we also expose the capability so the planner can reason about it?
+
+8. **One trait boundary per layer.** Crossing a layer means going through its trait. Direct calls to lower-layer concrete types from upper layers are forbidden.
+   *Check:* Does this code call `lance::Dataset` directly outside engine-storage? Call planner internals from the executor?
+
+9. **No god modules.** Single-module concerns: storage, IR, planner, executor, frontend, reconciler, schema, policy. Each crate has a reference test suite that runs without the others.
+   *Check:* Does this PR add a concern to a crate that already owns a different one?
+
+## III. Distributability — kernel stays remote-friendly
+
+These are technical constraints, independent of whether we ship a distributed product. They preserve the architectural seam.
+
+10. Storage trait is `Send + Sync`. No in-process-only assumptions in `Dataset` impls.
+11. `Dataset` impls accept remote descriptors (URI, snapshot ref, fragment ID) without requiring an open in-process Lance handle.
+12. **IR is location-neutral.** No IR operator embeds an assumption about where data lives.
+13. Cost model accepts a network-cost term as a future additive component. No place hard-codes "all cost is local I/O."
+14. Reconciler trait admits alternate implementations. In-process tokio is the OSS default; a separable worker fleet is the distributed shape.
+
+## IV. *(Additivity / migration — placeholder, not yet drafted)*
+
+## V. Honesty — cost, observability, calibration
+
+20. **Estimate-vs-actual logging on every estimator.** Cost models drift; calibration is a continuous process, not a one-off.
+21. **Coverage and lag are first-class metrics.** Index coverage, reconciler lag, cost-model accuracy — surfaced through the storage trait's `capabilities()` and a unified observability API.
+22. **Honest failure modes.** Cost-model misses degrade gracefully (spill, partial-result, bounded abort). No silent OOM.
+23. **Per-query budgets propagate through operators.** Memory cap, wall-clock timeout, max-rows-scanned, max-fragments-scanned. Operators respect them; budgets exposed via explain.
+24. **Plans are explainable.** Every executed query can be inspected as IR + physical plan + cost annotations. No "you'd have to read the source to know what this does."
+
+## VI. Patterns — use the unified mechanism
+
+When two features look similar, they probably share a mechanism. Use it.
+
+25. **Reconciler pattern for derivable state.** Index coverage, statistics, anything derivable from manifest state — reconciled, not job-queued.
+26. **Polymorphism via Union, not per-feature lowering.** Interfaces / wildcards / alternation on nodes and edges share one IR (`Polymorphism<T>`) and one lowering (Union of per-type concrete plans).
+27. **Mutations wrap read subplans.** Insert / Update / Delete / Merge are operators that consume read-shaped subplans. Same planner, same cost model, same storage trait.
+28. **SIP for cross-operator selectivity propagation.** Producers publish ID bitmaps; downstream scans consume them through structured pushdown. Don't ad-hoc IN-lists.
+29. **Factorize multi-hop, flatten only at projection.** Lists carry multiplicity through intermediate operators. Flatten is inserted by the planner where required, not eagerly.
+30. **Stable row IDs as dense graph IDs.** Don't maintain parallel string→u32 maps. Lance's stable row IDs are the substrate's identity layer; we use them directly.
+31. **Rank and score are columns.** Retrieval operators emit `_score`, `_rank`. Fusion operators consume rank-bearing batches. Don't discard rank in pipeline-twice merges.
+32. **Policy as predicates.** Authorization decisions are filter expressions injected into the planner, not enforcement at the API boundary.
+
+## VII. Quality gates — every change passes
+
+33. **Tests at every boundary.** `MemStorage` for engine tests; planner-only tests; executor-only tests with a stub storage. No layer tested only via end-to-end.
+34. **Reference implementation per trait.** Every trait has a primary impl (Lance for storage) and at least a test impl.
+35. **Documented capability surface.** New capabilities are documented with what they advertise, who consumes them, how the planner uses them.
+36. **Benchmark before optimization.** New optimizations land with a benchmark that motivates them; if the motivating workload doesn't exist, the feature waits.
+
+## VIII. *(OSS / Cloud kernel-product split — placeholder, not yet drafted)*
+
+## IX. Anti-patterns — deny list
+
+If a proposal fits one of these, the burden is on the proposer to justify why this case is the exception.
+
+- Synchronous-inline index updates for indexes expensive to build (vector ANN, FTS). Reconciler pattern instead.
+- Custom WAL / transaction manager / buffer pool. Lance owns these.
+- Job queue for state derivable from manifest. Reconciler pattern instead.
+- Per-feature lowering for shapes that share a structure (interfaces, wildcards, alternation). Use one mechanism.
+- Eager materialization of cross-products in multi-hop. Factorize; flatten only when needed.
+- Ad-hoc IN-list filtering when SIP fits.
+- String-flattened SQL filter generation when structured pushdown is available.
+- In-process-only `Dataset` impls. `Send + Sync`, remote descriptors.
+- Cost-blind plan choice. Lowering-order execution is not a planner.
+- Hidden statistics. If a metric matters for plan choice, it must be exposed through the trait surface.
+- Side-channels for query semantics. Search modes, mutations, polymorphism — all first-class IR concepts.
+- Discarding rank in retrieval. Score and rank propagate as columns.
+- State that drifts from the manifest. Derive from observable state.
+- Cloud-only correctness fixes. Correctness is always OSS.
+- Forking the codebase for Cloud. Trait-extension only.
+- Hand-rolling something Lance already does. Check the spec first.
+- Mutating in place state that should be immutable (Lance fragments, index segments). New segments instead.
+- Silent failures. OOM, timeout, partial result — all surfaced and bounded.
+
+## X. Review checklist (use against any non-trivial change)
+
+Print this when reviewing an RFC or PR. Each line is **yes / no / N/A**.
+
+- Does it respect the substrate? (§I)
+- Does it cross only one trait boundary per layer? (§II)
+- Are capabilities and stats exposed for any new behavior? (§II.7)
+- Storage trait stays `Send + Sync` and remote-friendly? (§III)
+- Additive, not rewrite? Feature-flagged where behavior changes? (§IV)
+- Any new estimator has estimate-vs-actual logging? (§V.20)
+- Coverage / lag / budget metrics surfaced? (§V.21–23)
+- Failure modes graceful, bounded, observable? (§V.22)
+- Reuses an existing pattern (reconciler, Union, mutation-wrap-read, SIP, factorize) where applicable? (§VI)
+- Tests at every boundary, not just end-to-end? (§VII.33)
+- Reference impl + test impl for any new trait? (§VII.34)
+- If commercial-relevant: kernel/product split preserved? (§VIII)
+- None of the deny-list patterns apply? (§IX)
+
+## XI. Living document policy
+
+This document is updated when:
+
+- A new architectural decision establishes a new invariant — add it.
+- An existing invariant is challenged and either reaffirmed (with the case sharpened) or revised (with explicit migration of any affected code).
+- A new anti-pattern surfaces in review and deserves a place on the deny-list — add it.
+
+Updates require the same review process as code. Adding to the deny-list (§IX) is cheap; removing or relaxing an invariant (§I–VIII) requires explicit justification in the proposal.
+
+When an invariant is contested in the moment, the resolution path is: (a) state the case in the relevant RFC or PR; (b) link it from this document; (c) update this document if the resolution changes the rule.
+
+## XII. Source / origin
+
+These invariants were extracted from the architectural decisions in:
+
+- **MR-737** — Query Engine v2 RFC (the kernel scope and seams)
+- **MR-738** — OSS / Cloud strategy (the commercial overlay)
+- The schema migration program (**MR-694** family — additive evolution, safety tiers)
+- The policy program (**MR-722 / MR-725** — predicate pushdown)
+- The reconciler / index-lifecycle work (**MR-737 §5.16, MR-688, MR-679, MR-680**)
+- The factorization and SIP work (**MR-737 §5.2, §5.3** — Kuzu / Ladybug inspiration)
+- The polymorphic-bindings framing (**MR-737 §5.13** — one mechanism for eight cells)
+
+General precedent: Lance + LanceDB Enterprise architecture; ClickHouse merge subsystem; Kubernetes controllers; Postgres autovacuum; the FDAL stack (Flight + DataFusion + Arrow + Lance).
+
+Adding a new invariant here means we've learned something — either from a hard call we made and want to preserve, or from a mistake we don't want to repeat. Both are worth recording.