diff --git a/docs/research/epistemic-awareness/2026-10-29-epistemic-basin-metathought-pr-scratchpad.md b/docs/research/epistemic-awareness/2026-10-29-epistemic-basin-metathought-pr-scratchpad.md
new file mode 100644
index 000000000..438323649
--- /dev/null
+++ b/docs/research/epistemic-awareness/2026-10-29-epistemic-basin-metathought-pr-scratchpad.md
@@ -0,0 +1,92 @@
+# Epistemic basins → Global Sudoku → 0+63 Thought continuations → 64×64×16 metacognitive frontier
+
+**Date:** 2026-10-29  
+**Status:** BRAINSTORM / research scratchpad / proposed PR sequence. **Not** an approved ABI, runtime, source-writing mechanism, cognitive claim, or merge authorization.  
+**Scope:** new *integration frontier*, not a replacement for existing DR/DT, Thought-Git, Pearl, NARS, EAW, CFA, NSL or TurboVec plans.  
+**Coordination:** Check GitHub main/open PRs and EXO-50 immediately before each new PR. This scratchpad intentionally adds only one file; do not edit shared board files in concurrent sessions.
+
+## 0. The proposed loop, in one breath
+
+Global Sudoku uses existing **HHTL16 Cartesian/Morton coordinate access** to visit relevant **DR64/DR256 focus windows** and **fold directly**, rather than requiring a globally materialized DR256 WHERE selection (1,024 u64 words, 8 KiB) for every thought. A *chunk* is a source-bounded epistemic basin: local evidence, open questions, contradictions, provenance and a **derived** uncertainty/selection signal. Pearl's existing proof gates test the relevant relations and **only then** may its existing writers produce a lawful Epi5 topology/certification change. DR's next selection can consume the derived entropy/pressure ranking, without treating pressure as truth.
+
+Context progresses chunk by chunk. Alternative progressions may occupy the **0+63 Thought-Git** frontier: T0 is sealed givens, T1..T63 speculative overlays, not 64 duplicated worlds. Metacognitive evaluation can compare a bounded **64 origins × 64 possible continuations × 16 declared evaluation lenses = 65,536 logical addresses**. This is a **hypothesis** about a useful logical axis product, *not* a currently ratified DR layout, a mandated full matrix, a 16-lens Tenant16 encoding, 65,536 CE64 edges, 64 globally independent evidence roots, or subjective awareness.
+
+**Metacognitive target:** detect correlated blind spots, unsupported transitions and genuinely discriminating next experiments across candidate interpretations, using local DR folds and source-root witnesses. Never turn an ensemble vote or Markov likelihood into proof.
+
+## 1. Four different spaces. Do not alias them.
+
+| Space | Size | What it means | What must not follow |
+|---|---:|---|---|
+| Global Sudoku ordinal field | 65,536 possible ordinals (2^16) | Existing finite hydrated knowledge-field model | Not a demand for one global materialized bitmap or 65,536 node objects |
+| DR256² global WHERE universe | 65,536 Boolean cells; **8 KiB** if dense | Existing #1790/#1791/#1836 comparison and demanded-selection result | The **1,024 u64** selection words are not required as an intermediate for every fold |
+| DR256³ logical Cartesian comparison | 256³ = 16,777,216 cells; **2 MiB** at one bit/cell | Larger *logical* ternary coordinate space | Not automatically CE64 SPO24, physical DR³, or a required allocation. NFR-OP-2 third-axis type is OPEN |
+| Metacognitive proposal | 64 × 64 × 16 = **65,536** coordinate combinations | Thought-origin / possible continuation / declared evaluation lens, **roles subject to owner proof** | Same cardinality as the 64K field does **not** mean same identities or semantics |
+
+A DR64² relation is a 64×64, 512-byte Boolean register. A proposed DR64³ logical space is 64³, 32 KiB if physically dense at one bit per cell. HHTL16 Morton8x8 with 4×4 refinement nibbles is already used in D-DRCA-1. It is the **spatial coordinate/router** for the Cartesian focus, **not a new 8+8+8 semantic SPO packing**. DR owns coordinate mapping and fold projections; **DT** owns ClassId-bound schema/method/meaning, **Quack** query binding, **WORLD** sealed source/evidence, **Thought-Git** speculative overlays.
+
+## 2. Evidence already on main; do not reimplement it
+
+1. **Global Sudoku / pressure:** [D-GSO plan](https://github.com/AdaWorldAPI/lance-graph/blob/main/.claude/plans/2026-10-06-global-sudoku-replayable-orchestration-v1.md), [#1351 entropy × topology](https://github.com/AdaWorldAPI/lance-graph/pull/1351), [#1398 crossword attention](https://github.com/AdaWorldAPI/lance-graph/pull/1398). Shannon entropy chooses *where to investigate*, not *what is certified*.
+2. **No persistent masks as default:** [resident fold plan](https://github.com/AdaWorldAPI/lance-graph/blob/main/.claude/plans/2026-10-08-resident-projection-fold-mask-v1.md), [#1790 global-mask control](https://github.com/AdaWorldAPI/lance-graph/pull/1790), [#1791 on-demand / L1 control](https://github.com/AdaWorldAPI/lance-graph/pull/1791), [DR256 WHERE capstone](https://github.com/AdaWorldAPI/lance-graph/blob/main/.claude/plans/dr256-where-capstone-v1.md). #1791 measured resident-fold **timing** slowdown 11.8/13% on clustered repeated rebuilds, but had **no hardware L1 counters**; retaining a single G was within ~1.5%. It does **not** prove all maskless access wins.
+3. **Coordinate routing and ownership:** [#1834 HHTL16 DRCA-1](https://github.com/AdaWorldAPI/lance-graph/pull/1834), [#1835 DR/DT amendment](https://github.com/AdaWorldAPI/lance-graph/pull/1835), [#1836 guarded WHERE→DR#3 projection](https://github.com/AdaWorldAPI/lance-graph/pull/1836), [DR math binding lab](https://github.com/AdaWorldAPI/lance-graph/blob/main/crates/dr-math-binding-probe/src/hhtl16.rs), [D-NFR native status](https://github.com/AdaWorldAPI/lance-graph/blob/main/.claude/plans/d-native-repr-contract-v1.md). DR64² and borrowed/read-guarded projections are LAB, physical 3rd axis not ratified. Unknown ≠ absent ≠ false.
+4. **Pearl and Epi5:** [pearl::hydrate / reason / revise](https://github.com/AdaWorldAPI/lance-graph/blob/main/crates/lance-graph-planner/src/pearl.rs), [EpistemicState5](https://github.com/AdaWorldAPI/lance-graph/blob/main/crates/lance-graph-contract/src/epistemic_state5.rs), [#1840 ledger](https://github.com/AdaWorldAPI/lance-graph/pull/1840), [#1843 grounded mediator adapter](https://github.com/AdaWorldAPI/lance-graph/pull/1843), [#1842 scent](https://github.com/AdaWorldAPI/lance-graph/pull/1842), [#1844 scent learner](https://github.com/AdaWorldAPI/lance-graph/pull/1844). Bits 59..60 are Topology2, 61..63 Certification3; only evidence-qualified existing owners write them.
+5. **Thought/counterfactual:** [#1628 0+63 Thought basics](https://github.com/AdaWorldAPI/lance-graph/pull/1628), [#1674 replay receipt](https://github.com/AdaWorldAPI/lance-graph/pull/1674), [#1813 and #1815 CFA plans](https://github.com/AdaWorldAPI/lance-graph/pull/1815), [#1824 EAW world isolation](https://github.com/AdaWorldAPI/lance-graph/pull/1824), [#1827 temporal frontier](https://github.com/AdaWorldAPI/lance-graph/pull/1827), [#1828 feature-delta ranking](https://github.com/AdaWorldAPI/lance-graph/pull/1828). Speculation is not source evidence. Thought slot is not immutable commit identity, and CE64 W=53..58 is a **writer breadcrumb**, not necessarily the current reader's slot on inherited givens.
+6. **Existing parallel work, do not duplicate:** [#1847 / S1–S7 handoff](https://github.com/AdaWorldAPI/lance-graph/blob/main/.claude/plans/2026-10-29-dr3-epiphany-session-scratchpad.md) already owns DR³ exact residuals, CE64×Thought typed projection, root-disjoint epiphany labels, NSL learner and TurboVec retrieval. This scratchpad covers **chunk-basin + attention + metacognitive progression** rather than restating S1–S7.
+7. **MARKOV CORRIGENDUM, binding boundary:** [#1737 corrected coresearch note](https://github.com/AdaWorldAPI/lance-graph/blob/main/.claude/board/entries/2026-10-23-coresearch-thought-git-markov-cc2.md). **WORLD/Thought-Git substrate is deterministic, not probabilistic.** Probabilistic models are explicit *counterfactuals*, require cold law/compiler certification when claimed, and cannot be a stochastic truth or merge authority. Never use Markov probabilities, stationary mass, votes or epsilon squash to change Epi5 or eliminate exact replay. Predictive context ranking may remain provisional as a read-only priority hypothesis.
+
+## 3. Proposed small, non-overlapping PRs
+
+**Labels below are planned work-package identifiers, not existing PR numbers.** Every worker must check the current main, outstanding draft PRs and EXO-50 before claiming an owner. Keep each implementation PR on its own main-based branch; provide independent oracle, negative/disable controls, real cold-preparation cost and an explicit ADOPT-LAB / DEFER / REFUSE verdict. Do not change canonical board/index from parallel workers.
+
+| Order / label | Proposed *new* PR, single question | Primary write scope / owner | Depends on / acceptance gate |
+|---|---|---|---|
+| **B0 / D-EBM-0** | **Freeze roles + oracle**: define chunk/basin, source-witnessed context, origin/continuation/lens identity; distinguish all four spaces; specify an independent scalar global-Sudoku→Pearl→Thought oracle | New docs/test-fixture file under epistemic-awareness; council/contract owner | First. Do not invent the 16-lens ABI or ratify third DR axis. Freeze seeded fixtures, source roots, witness mismatch and missingness before other worker code |
+| **B1 / D-EBM-1** | **Address-only global→local walker**: visit HHTL16 Morton-addressed focused windows for a global candidate, with no demanded global bitmap; compare direct sealed map and legal rebinding | New lab example in dr-math-binding-probe, **DR owner** | B0 + #1834/#1836. Every selected address maps to the correct source identity; out-of-focus, stale, duplicate, foreign coordinate refused; no post-bind raw value gather; no new global namespace |
+| **B2 / D-EBM-2** | **Pearl hydration through focused DR folds**: read-only adapter that uses the existing sealed-chain/ledger and existing pearl::hydrate. Distinguish local-no-path from globally-not-present | New test file in lance-graph-planner, **Pearl owner** | B0+B1; exact parity with full independent path enumerator (including cross-window intermediates), K3 unknown, direction, cycles, missing links; **no new Epi5 writer** |
+| **B3 / D-EBM-3** | **L1 and allocation shootout**: compare #1791 dense/reused G, local DR only, scoped Morton walk, CSR/SoA; measure real address and cold bind overhead, not just fold | New drf-focus-probe benchmark, **performance owner** | B1; same candidates/answers, vary densities, 0/1/2/4/8/64 thought reuses, scattered/clustered, full/empty, pin CPU; report wall-time and cache counters **only if available**; no performance claim absent measurements |
+| **B4 / D-EBM-4** | **Pure epistemic basin projection**: chunk-level source root set, read-only Epi5/topology, open alternatives, contradiction, entropy and no-source-evidence status, each with valid revision | New cognitive-shader-driver example or stand-alone lab, **basin owner** | B0+B2; independent recomputation from original chunks; duplicate roots do not inflate support; equal entropy/different topology stays distinct; unset/UNKNOWN never becomes zero/false |
+| **B5 / D-EBM-5** | **Deterministic next-chunk context transition**: advance compact context using admitted fold and nonlocal source-address references; optional Markov-shaped proposal remains read-only and explicitly counterfactual | New deepnsm-v2 or context lab example, **context owner**, coordinate with DeepNSM current scope | B4; deterministic replay, same inputs same context, backwards/out-of-order references, long-range dependencies; if compressed context is insufficient retrieve addressed evidence or refuse; no bundled history or Markov truth engine |
+| **B6 / D-EBM-6** | **0+63 contextual continuation overlays**: each possible next chunk binds to a sealed WORLD frame plus sparse branch-specific delta and immutable parent/provenance; no 63 copies | New thought-git-probe example/test, **Thought-Git owner** | B4+B5; slot 0 sealed, slots 1..63 speculative, 64 invalid; ABA/generation, sibling leak, source-root reuse, replay of evicted branch, frame revocation, protected intermediates; no spontaneous source writes |
+| **B7 / D-EBM-7** | **Conditional counterfactual continuation**: proposed interventions across next chunks, with conditional-law proof/refusal; borrow existing CFA/EAW/pearl owners | New CFA lab test, **counterfactual owner** | B2+B6; original sealed chain untouched, intervention not confused with observation, proof compiler conditional claim separate from empirical causal certification, no unsupported CE64/Epi5 promotion |
+| **B8 / D-EBM-8** | **64×64×16 lazy metacognitive address view**: typed (origin, continuation, lens) with no dense 65,536-CE64 population or 16-lens claim; one terminal receives a single folded result | New isolated dr-math-binding-probe example **after** S1 DR³ owner review; **DR math owner** | B0+B5+B6; logical product 65,536; origin=0 denotes givens not a hypothesis; lens types explicitly declared; cross-frame/foreign map/colliding slot refuses; scalar oracle parity; no third-axis ABI by stealth |
+| **B9 / D-EBM-9** | **Metacognitive failure/insight oracle**: discover shared unsupported premise, circular echoes, contradictory siblings, hidden common roots and a distinguishing next experiment; no majority-as-evidence | New NARS/ledger test module, **epistemic owner** | B7+B8 + #1840/#1843; correlated-copy fixture (47 apparent votes, one real root), independent-root control, honest unchanged uncertainty, causal dependence vs derivation dependence, exact replay and no value-equality squash |
+| **B10 / D-EBM-10** | **Cognitive pressure → DR selection arbitration**: rank next focus by declared entropy/contradiction/topology/evidence-competence and expected information gain **as a hypothesis**; do not create a second scent learner | New evaluation example beside C3b/C3c, **attention owner** | B3+B4+B9; matched-budget comparison vs FIFO, uniform, #1842 scent, #1844 C3c; true earned Epi5 changes per unit cost, ground truth from independent oracle, no leak from future chunks, no automatic certificate |
+| **B11 / D-EBM-11** | **Council cross-layer integration decision**: compile tested costs, guard matrix, actual proof status, viable bounded end-to-end path and permanent refusal cases | New read-only report/docs, **integrator** | B1–B10 receipts; may decide LAB only, DEFER or REFUSE. No production rollout, CE64 layout change, auto-merge or new persistent world without explicit owner decision |
+
+### Scheduling, without collisions
+
+- **Wave A:** B0 only: freeze fixture, roles and baselines. This may be a short docs-only PR.
+- **Wave B:** B1 (DR) and B4 *fixture preparation only* (basin) may begin independently after B0; B4 integration waits on B2. B3 starts after B1, alongside B2.
+- **Wave C:** B5 after B4; B6 after B5; B7 after B6; keep them in different crates/branches.
+- **Wave D:** B8 after B5/B6 and DR owner signoff; B9 after B7/B8; B10 after B3/B9; B11 last.
+- **Concurrent but separate:** #1847 S1–S7 exact-residual/epiphany learner work and ndarray/TurboVec9 quality probes. Consume their checked results, never write their owned source paths from this lane.
+- At every handoff: state *source SHA and frame*, exact changed files, data/oracle, bench host, cold/hot allocations, all failed controls, and **what was not proven**.
+
+## 4. One small executable acceptance path before scaling
+
+**Fixture**: two short chunks from a frozen public or synthetic passage stream; three Thought states (T0 givens, T1/T2 divergent interpretations); two potential continuations; two declared lenses (e.g., grounding and novelty), with independent scalar ground truth. At least one cross-chunk relation, one contradictory reading, one duplicated-root "consensus", one missing path, one valid independent root, one unchanged source with changed thought, one stale frame and one UNKNOWN.
+
+1. Complete deterministic baseline **without** DR; know the right contextual step and Pearl/Epi5 outcome from the independent sealed oracle.
+2. Execute addressed local DR folds **without** a global selection bitmap. If focus is incomplete, expand legally or refuse.
+3. Add basin/entropy **read-only**; ensure entropy changes ordering only.
+4. Add sparse Thought forks; verify zero cross-branch writes and exact replay.
+5. Add a metacognitive lens that catches shared-root false agreement; measure whether it selects the discriminating question.
+6. Expand up to **64×64×16 only as an addressed logical comparison**, with no implied allocation; check the dense-reference oracle in test-only infrastructure if needed.
+
+**Counterexample gates**: a 47-vote echo must not defeat a single independent refutation; identical low entropy with different Epi5 topology must route differently; local absence is not global non-existence; invalid thought slot/ABA must refuse; probabilistic next-chunk proposals may never overwrite exact root or Pearl certificate; one stale DT/DR/Codebook generation poisons the reuse; corpus chronology and source authority may not be replaced with sequence order.
+
+## 5. Required measurement and no-go policy
+
+**Report costs separately:** global→local bind, address generation, cursor/rebind, local fold, source-dependent reads, optional mask construction, branch overlay/replay, selection ranking, Pearl experiment, and write/seal (if explicitly authorized). For every strategy show throughput/latency, working-set bytes and allocations, *actual hardware* L1D misses only when measured, correct/refused counts, and verified discriminating discoveries at equal work budget.
+
+**Immediate no-go** if: (a) CE64 59..63 are updated from a prediction, entropy, or vote; (b) invented 64×64×16 cells become source evidence; (c) markovian probabilities become a deterministic Thought-Git merge law; (d) the same source root is counted as independent confirmation across children; (e) a global DR256 mask silently becomes mandatory; (f) an incorrect/stale ClassView/DT/DR coordinate is treated as the same identity; (g) a dense DR³ or 65,536-CE64 allocation appears before a consumer and cost gate; (h) an existing unmerged worker's file or EXO-50 decision is overwritten.
+
+**Out of scope**: new CE64/SPO24 packing, 512³ predicate ontology, native U9 TurboVec index, revision of 34 shipped recipe IDs, replacement of Quack/DT/DR owners, philosophical consciousness claim. Those are separate questions.
+
+## 6. First worker handoff (paste-ready)
+
+**B0 owner prompt:** "In AdaWorldAPI/lance-graph create a *docs/test-fixture-only* independent draft PR for **D-EBM-0**, based on current main. Read this scratchpad, the global Sudoku plan, DR/DT ownership amendment, #1834/#1836 DR labs, #1791 L1 comparison, #1843 Pearl ledger adapter, #1737 Markov corrigendum, #1815 CFA, #1824 EAW and #1847 S1–S7 scratchpad. Freeze the smallest two-chunk/three-thought/two-continuation/two-lens independent scalar fixture, type every axis, identify existing owner for each component, show exact no-mask vs demanded-mask controls, and pre-register adversarial cases. **Do not implement a third DR axis, a probability-driven substrate, a new CE64 writer, a global bitmap or an additional ledger.** Before writing, inspect current GitHub PRs and EXO-50 for concurrent work. Make one new file/branch only, no shared board edits, no merge/force push. Return testable acceptance criteria and the precise B1/B2 handoff."
+
+---
+
+**One-line north star:** *The whole corpus is addressable, only the current question is hot; 63 speculative interpretations explore the next chunk, and 64×64×16 compares their blind spots without minting truth or materializing a universe.*