@@ -18,13 +18,13 @@ the membrane / Valhalla cheap addresses / everything else underneath).
 | D-LGJ-MMV-3 | Valhalla value-class promotions where the three-truths lab shows a measured facade-path win; each reversible | Queued (evidence-gated) |
 | D-LGJ-MMV-4 | Temporal read binding via `LanceVersion`/`VersionRange`/`TemporalPov` (contract vocabulary, never the Kanban-entangled types) | Queued (substrate-first per F6) |
 | D-MRL-2c | *proposed:* lower family-(1) agreement topology onto mask ops | **Proposed** — needs its own F-RLR-2 insufficiency argument; MUST NOT be conflated with family (3) |
-| D-MRL-G2 | **the blocking question §14's arithmetic owes** (v4.2 §17.3): what row space does a mask span, times how many classes, rebuilt how often? | **BLOCKING (2026-09-03).** KILL: if a cube is per-class over the GLOBAL row space, `C × 9 × N/8` at 10M rows × 100 classes ≈ **1.1 GB**, and §14.2's headline 11.25 MB is struck. Either the numbers are understated by the class count or the L8 strike is unfounded — both cannot hold. A pricing defect, NOT a novelty defect |
+| D-MRL-G2 | **the blocking question §14's arithmetic owes** (v4.2 §17.3): what row space does a mask span, times how many classes, rebuilt how often? | **BLOCKING (2026-09-03).** KILL: if a cube is per-class over the GLOBAL row space, `C × 9 × N/8` at 10M rows × 100 classes ≈ **1.05 GiB**, and §14.2's headline **10.73 MiB** is struck. Either the numbers are understated by the class count or the L8 strike is unfounded — both cannot hold. A pricing defect, NOT a novelty defect |
 | D-MRL-2a | mint `BELNAP_JOIN` at a core loco slot with a lowering table | **BLOCKED (2026-09-03, v4.1 §16.1)** — the carrier is deleted. `EpistemicBassin24`/`support_mask`/`refute_mask`: zero occurrences in lance-graph; the `0x87..0x8B` band was retracted 2026-09-02 *with the model it encoded* (`ogar-loco/src/lib.rs:596-599`). Unblocks only if a live carrier for the join semantics exists |
 | D-MRL-1a | `TERNLOG` + `TERNARY_MATCH` ops | **AMENDED (2026-09-03, v4.1 §16.2)** — no mint: consume `ogar_loco::TERNLOG` = `FnIndex(0x86)`, which ships with this op's description verbatim and **zero consumers**. Wave 1 becomes its first consumer, discharging the vacancy `E-A-RULED-HOME-NEEDS-A-FIRST-CONSUMER-OR-IT-IS-A-VACANCY-1` names |
 | ⊘ D-MRL-2c (SUPERSEDED by the row above, v4.2) | *proposed, not scheduled:* lower family-(1) agreement topology (`witness_fabric` — ⊘ the "119 scalar sites" here is the count v4.2 withdrew as unreproducible; kept per append-only, NOT a live figure) onto mask ops | **Proposed (2026-09-03, v4.1 §16.3)** — needs its own F-RLR-2 insufficiency argument. MUST NOT be conflated with family (3): the six-families ruling forbids the impersonation |
 | D-MRL-0f | **OSM voxel probe** (`mask-risc-lowering-v1.md` §15) — build the 3-state-per-tier voxel cube over a real OSM tile population addressed by `ogar_osm::GEO_V3_FACET`; measure the cardinality question **at the arity of the population actually probed** — ⊘ v4.2: 27-vs-8 is the CANONICAL THREE-TIER KEY; `GEO_V3_FACET` has FOUR tier rails, so the OSM case is **81 cells vs 16**. The struck text said 27-vs-8 unqualified, which would have measured the wrong population, primitive memory vs prediction, cull fraction per viewport | **Queued (2026-09-03, v4)** — falsifier of §14. KILL: the real stream needs patterns finer than the tier byte-pair ⇒ §3c.1's "OSM is the native case" is struck |
 | D-MRL-0g | **Frame-sequence probe** (§15.1b) — ordered viewport path against the horizontal rail cache under a real capacity bound; frame-to-frame overlap, hit COUNT, `cold − warm`, prefetch-resolvable fraction | **Queued (2026-09-03, v4)** — the most favourable honest stream for the trie. KILL: fails G9 here ⇒ Wave 1c dies and is not retried on a friendlier stream |
-| D-MRL-0..3 | `mask-risc-lowering-v1.md` — verbs → mask RISC → trie memo over immutable versions (photolithography). Waves: 0 probes (`DOWN[x]` falsifier, rail-group hit rate, cold build cost), 1 `lgj-abi` (TERNLOG/TERNARY_MATCH ops, survivor skip, trie, tier masks), 2 loco verbs with a lowering table, 3 Java glove | **v4.2 COUNCIL-CORRECTED (2026-09-03)** — 5+3 verification council (the review CodeRabbit/Bugbot could not give): 42 savant findings, 2 reviewer BLOCKs, 2 external P2s. 8 claims struck-and-regraded in place, incl. `witness_fabric` 379→**3,439 lines with its "landed after the council" excuse retracted**, `3¹²` 4.2 GB→**4.05 GiB**, blasgraph six→**seven** semirings, §7's D-MRL-2a row ⊘'d as a **live mint directive**, and five model identifiers §10 kept after the v3 BLOCK fixed only §8. New blocking row D-MRL-G2. Prior: **v4 AMENDED** — §14 voxelmasking (vertical axis enumerated, not cached: 9 primitives = 72 KB at 65,536 rows; strikes S5-Q4/L8/F7/F12/F10 on that axis only), §3c reuse map (OSM native; weather splits wind→voxel / humidity×13→blasgraph), §14.6 culling, §14.7 prefetch. Prior: **RATIFIED v3** — 5+3 council complete: 5 savants/40 findings → draft v2 (24 amendments) → 3 reviewers (2 BLOCK, 14 FIX) + 2 external Codex P1s → v3. Both BLOCKs resolved (model identifiers struck from the artifact; the cost ladder regraded `[E]` throughout). Waves QUEUED behind Wave 0, which now gates all of Wave 1 |
+| D-MRL-0..3 | `mask-risc-lowering-v1.md` — verbs → mask RISC → trie memo over immutable versions (photolithography). Waves: 0 probes (`DOWN[x]` falsifier, rail-group hit rate, cold build cost), 1 `lgj-abi` (TERNLOG/TERNARY_MATCH ops, survivor skip, trie, tier masks), 2 loco verbs with a lowering table, 3 Java glove | **v4.2 COUNCIL-CORRECTED (2026-09-03)** — 5+3 verification council (the review CodeRabbit/Bugbot could not give): 42 savant findings, 2 reviewer BLOCKs, 2 external P2s. 8 claims struck-and-regraded in place, incl. `witness_fabric` 379→**3,439 lines with its "landed after the council" excuse retracted**, `3¹²` 4.2 GB→**4.05 GiB**, blasgraph six→**seven** semirings, §7's D-MRL-2a row ⊘'d as a **live mint directive**, and five model identifiers §10 kept after the v3 BLOCK fixed only §8. New blocking row D-MRL-G2. Prior: **v4 AMENDED** — §14 voxelmasking (vertical axis enumerated, not cached: 9 primitives = 72 KiB at 65,536 rows; strikes S5-Q4/L8/F7/F12/F10 on that axis only), §3c reuse map (OSM native; weather splits wind→voxel / humidity×13→blasgraph), §14.6 culling, §14.7 prefetch. Prior: **RATIFIED v3** — 5+3 council complete: 5 savants/40 findings → draft v2 (24 amendments) → 3 reviewers (2 BLOCK, 14 FIX) + 2 external Codex P1s → v3. Both BLOCKs resolved (model identifiers struck from the artifact; the cost ladder regraded `[E]` throughout). Waves QUEUED behind Wave 0, which now gates all of Wave 1 |
 
 W4.2 (the write seam) deliberately has NO row — blocked with GridLake
 (plan F5) and substrate-side by the F7 measurement; a row here would be
 
@@ -467,19 +467,28 @@ of them precomputed at version seal.
 
 ### §14.2 — The arithmetic (the whole argument)
 
-A mask over `N` rows costs `N/8` bytes. At the fixture size `N = 65_536` that is **8 KB**.
+A mask over `N` rows costs `N/8` bytes. At the fixture size `N = 65_536` that is **8 KiB**.
 
 | quantity | at N = 65,536 | at N = 10,000,000 |
 |---|---|---|
-| one mask | 8 KB | 1.25 MB |
-| **9 primitives** (3 tiers × 3 states) | **72 KB** | **11.25 MB** |
-| full 27-cell cube, if materialised | 216 KB | 33.75 MB |
+| one mask | 8 KiB | **1.19 MiB** |
+| **9 primitives** (3 tiers × 3 states) | **72 KiB** | **10.73 MiB** |
+| full 27-cell cube, if materialised | 216 KiB | **32.19 MiB** |
 
 **Only the 9 primitives are built.** The other 18 cells are an AND of three primitives —
 the exact op `plan_eval` already composes (`exports.rs`, the existing chain: n ops, one
 crossing, monotone `V(k+1) ⊆ V(k)`). So the cube is a **9-entry array indexed by a base-3
 digit triple**, not a 27-entry store.
 
+> **⊘ v4.3 — UNITS ARE BINARY (KiB / MiB / GiB) THROUGHOUT.** The v4.2 council
+> BLOCKed the mixed convention and deferred it as *"a decision, not a correction"*.
+> Decided here: **binary**, because every quantity in this section is `N/8` bytes over
+> a power-of-two row count and a power-of-three cell count — the decimal reading was
+> only ever an artifact of writing `1.25 MB` for `1,250,000` bytes while writing
+> `72 KB` for `73,728`. The struck decimal figures were **10M: 1.25 / 11.25 / 33.75 MB**;
+> the same bytes in binary are **1.19 / 10.73 / 32.19 MiB**. No byte count changed —
+> only its name. The 65,536 column was already binary and is unchanged.
+
 ### §14.3 — What this DELETES, and why that is the point
 
 Five open items in v3 exist only because hit rate on the vertical axis was unmeasured. Under
@@ -503,8 +512,8 @@ Enumeration works **at tier granularity and nowhere else**.
 
 | granularity | cells | primitive cost at N = 65,536 |
 |---|---|---|
-| 3 tiers | 3³ = 27 | 72 KB — **build it** |
-| 6 rails (`6×(u8:u8)`) | 3⁶ = 729 | 144 KB primitives; cube 5.7 MB — borderline |
+| 3 tiers | 3³ = 27 | 72 KiB — **build it** |
+| 6 rails (`6×(u8:u8)`) | 3⁶ = 729 | 144 KiB primitives; cube 5.70 MiB — borderline |
 | 12 facet bytes | 3¹² = **531,441** | cube = **4.05 GiB** (⊘ v4.2 §17: the struck text said *"≈ 4.2 GB"*, which is neither convention — 4.05 GiB / 4.35 GB) — **never** |
 
 So the two axes **split their mechanism**, which is the v4 correction to v3's single-cache
@@ -583,8 +592,8 @@ test of the trie, and a failure there is close to decisive against it.
 
 Do real queries use **must-not-match** per tier, or only match / don't-care?
 
-- with must-not-match: 3 states, 27 cells, **9 primitives (72 KB)**
-- without: 2 states, 8 cells, **6 primitives (48 KB)**
+- with must-not-match: 3 states, 27 cells, **9 primitives (72 KiB)**
+- without: 2 states, 8 cells, **6 primitives (48 KiB)**
 
 This is not a design choice to make in a document — it is a **measurement on a real query
 stream**, and it is the primary output of D-MRL-0f below.
@@ -613,7 +622,7 @@ tiers (heel / hip / twig / leaf), each a `256×256` tile with x and y bound lite
 Note the honest discrepancy, recorded rather than smoothed: the canonical key carries **three**
 named tiers (HEEL/HIP/TWIG); the geo facet carries **four** rails as tiers (adding leaf). The
 cube is therefore `3³ = 27` on the key and `3⁴ = 81` on the geo facet (⊘ v4.2 §17: the struck text read "27 primitives-worth", a stale 3³ carry-over — it is 81 cells /
-addressing, still 12 primitives = 96 KB at N = 65,536). D-MRL-0f measures on the geo facet and
+addressing, still 12 primitives = 96 KiB at N = 65,536). D-MRL-0f measures on the geo facet and
 reports which arity the real stream uses.
 
 ### §3c.2 — Weather SPLITS, and the split is the finding
@@ -853,7 +862,7 @@ The council's sharpest finding, and it is **not** a novelty objection:
 
 > **§14.2 prices ONE cube; §14.3's L8 strike asserts "a per-class array".** If a cube
 > is per-class over the global row space the cost is `C × 9 × N/8` — at 10M rows and
-> 100 classes ≈ **1.1 GB, not 11.25 MB**. Either the headline numbers are understated
+> 100 classes ≈ **1.05 GiB, not 10.73 MiB** (v4.3 binary units). Either the headline numbers are understated
 > by the class count, or the L8 strike is unfounded. **Both cannot hold.**
 
 Three unstated premises underneath, in dependency order — **G2 is the root**: