docs(D-CFA-0): counterfactual flow analysis capstone + thought dependency scratchpad

docs/research/counterfactual-behavior/README.md

@@ -0,0 +1,49 @@
+# D-CFA-0 — Counterfactual Flow Analysis: research-only PR stack
+
+**2026-10-27. Status: PREREGISTERED EXPLORATION / DOCS ONLY.** This is a first-round question map, not implementation approval, shipped DR64³, proof, new contract, benchmark, runtime, or permission to mutate business systems. The user will refine fixture and interface details in a subsequent review.
+
+## Claim to falsify
+Can a versioned, typed ON × WHEN × DO **logical incidence relation** be subjected to independent Thought-Git ablations, such that counterfactual outcome differences reveal conditional behavioral dependencies *not established by static SPO alone*, while rejecting false dependencies caused by coincidental co-activation?
+
+The 64³ address space (262,144 possible incidence cells; 32 KiB as a hypothetical fully dense Boolean representation) is **not** the 64-thought axis (0 immutable givens + up to 63 speculative slots) and not a deployed tensor. Each cell can reference multiple independently identified methods. Codebook/local ordinals must not masquerade as global identities.
+
+## Existing authority: do not duplicate
+| Concern | Existing owner / source | Non-goal |
+|---|---|---|
+| source harvesting | AdaWorldAPI/ruff `ruff_spo_triplet`, OGAR; D-SAG-6 #1760 | a new AST or Rust semantic oracle |
+| source-attested registration | DT64 Q0 and #1794–#1799 | a physical DT64 executor |
+| relational execution | Quack / RelAlg / R2IL / mask-risc / DR-borrow | new database or hot evaluator |
+| speculative history | D-TGA / `thought-git-probe`; #1605 #1739 #1684 | 63 durable Lance branches or side effects |
+| counts and association | `lance-graph-arm-discovery`; #1313 #1359 #1584 | second learning crate/count engine |
+| fixed-point/matching | Datafrog / Ascent / MORK differential arms; D-SAG-6 | unchecked recursive truth |
+| numerical models | NSL experiment; ONNX external target; ndarray kernels; #1802 | ONNX training inside DT64 |
+| law admission | D-DRMC; egglog; optional DialEgg/MLIR | learned correlation = proof |
+
+Canonical existing plans: [DT64 Q0](../../../.claude/plans/d-dt64-q0-borrowed-quack-binding-seam-v1.md), [SAG-6](../../../.claude/plans/d-sag-6-rust-codegraph-math-logic-tracert-v1.md), [Thought Git](../../../.claude/plans/thought-git-api-v1.md). Existing notes are authority, not rewritten here.
+
+## Planned review chain (draft-branch stack)
+| Gate | Document | Research question | Minimum kill test |
+|---|---|---|---|
+| CFA-0 | this file + `thought-dependency-scratchpad.md` | ownership and evidence grades consistent? | conflating a thought slot with cube coordinate is rejected |
+| CFA-1 | `cfa-1-typed-incidence.md` | can harvested signatures bind an ON/WHEN/DO presence relation without silent method loss? | two implementations in one cell survive; unknown guards do not fire |
+| CFA-2 | `cfa-2-ablation-replay.md` | can replay/ablation preserve intermediate and higher-order outcomes? | A,B redundant jointly; squash must not hide protected divergent intermediate |
+| CFA-3 | `cfa-3-arm-logic.md` | can ARM/RelAlg/Datafrog separate co-occurrence and dependency with held-out controls? | perfect co-activation but no influence never promoted |
+| CFA-4 | `cfa-4-nsl-onnx.md` | can optional trained ranking prioritize ablations without claiming semantic authority? | held-out precision and ablation cost do not improve => drop |
+| CFA-5 | `cfa-5-compiler-analysis.md` | can candidate behavioral laws become verified effect-safe transformations? | side-effectful reorder or unproven law refuses |
+
+**Status of every gate:** research question only; no tests run, performance measured, code added, type invented, or source write authorized.
+
+## Non-negotiable semantics
+- **Fact vs hypothesis:** source-attested structure, population observation, statistical candidate, model-relative counterfactual effect, logical entailment, and independently checked law are separately labeled.
+- **Intervention:** disabling an ON event, changing WHEN, or suppressing DO is not interchangeable. Unknown/unauthorized/side-effectful scenarios may refuse.
+- **Identity:** semantic IDs and source revision, typed method/argument signatures, frame/focus/scope, codebook generation, permission/knownness, program and result witnesses are never collapsed into a u6 ordinal.
+- **History:** 0 = sealed givens; 1..63 = resident speculative branches. Replay and protected intermediate read requirements survive squash (#1739).
+- **Memory:** no forced dense cube; source data stays owner-borrowed; scratch is bounded and intermediate materialization must be costed.
+- **No automatic causal law:** A changes B in one replay is a *model-conditioned outcome*, not universal necessity, world causation, or legal compiler rewrite.
+- **External effects:** speculative arms create proposals only; no execution of source business actions.
+- **Computation:** native first. ONNX is a compatibility/prediction leaf, MLIR is an optional cold export, DialEgg/egglog is cold optimization, not hot execution.
+
+## Questions reserved for next round
+Which real executable fixture is first: synthetic Odoo action family, OGAR ON/WHEN/DO, or a small Rust method codegraph? Which exact outcome projection distinguishes visible outputs from internal trace deltas? Which causal grades and practical time budget suffice? What precise branch-level D-TGB semantics are available on the pinned main? None is pre-decided by creating this stack.
+
+See [thought-dependency-scratchpad.md](thought-dependency-scratchpad.md) for breadcrumbs and dependency arrows.

docs/research/counterfactual-behavior/thought-dependency-scratchpad.md

@@ -0,0 +1,57 @@
+# D-CFA thought dependency scratchpad — first-round breadcrumbs
+
+**Status: evolving research notes, not normative ABI.** Append discoveries here on a future pass; do not treat proposed edges as implemented capabilities. The diagram is a *dependency of research claims*, not a runtime call graph.
+
+```mermaid
+flowchart TD
+  A["Ruff / OGAR source facts\nroles, typed identities, uncertainty"] --> B["CFA-1: DT64 ON × WHEN × DO\nlogical incidence, method multimap"]
+  B --> C["CFA-2: Thought-Git\nT0 baseline + T1…T63 interventions"]
+  C --> D["Typed outcome deltas\ntraces, witnesses, protected intermediates"]
+  D --> E["CFA-3: ARM Discovery\nco-occurrence vs counterfactual influence"]
+  E --> F["Datafrog / MORK / RelAlg\nconditional closure, hyperedges, disagreement"]
+  F --> G["CFA-5: Egglog / DialEgg\nguarded equivalence proposals"]
+  E --> H["CFA-4: Optional NSL\nlearn candidate priority"]
+  H --> I["ONNX leaf\nrank interventions; not law authority"]
+  I --> C
+  G --> J["Existing proof + effect + source-parity gates"]
+  J --> K["Versioned recipe / proposed method"]
+  K -. "new observations, not self-certified truth" .-> D
+```
+
+## Breadcrumb keys
+- **B0 source:** factual role bits in `AdaWorldAPI/ruff/crates/ruff_spo_triplet/src/role.rs` (READS, DERIVES, EFFECTS, CONSTRAINS, REQUIRES, AUTHORIZES, ENABLES, TRIGGERS). OGAR Trigger/Applicability/Capability facets are not a source of populated guards without deployed config.
+- **B1 binding:** DT64-Q0 is typed witness/binding and a borrowed input seam, not the new S1 ON/WHEN/DO cube. Coordinate equality without identity witness is not source equality.
+- **B2 thought:** `0` is sealed; `1..=63` are speculative slots, **not** the 64 DO axis; slot generation/reuse must not become method identity.
+- **B3 intervention:** ablate registered *method identities* where cell multiplicity matters, as well as axis cells; distinguish suppressing ON, altering WHEN, blocking DO. Causal interpretation requires specified intervention semantics.
+- **B4 protected history:** #1739 demonstrates that final-state squash can erase intermediate contradictions, divergences, and write multiplicity unless protected witnesses survive.
+- **B5 discovery:** ARM co-occurrence/count evidence (#1313/#1359) is separate from Thought-Git ablation evidence; repeated correlated thoughts are not independent evidence.
+- **B6 logic:** Datafrog/Ascent can serve cold differential oracles; MORK matches patterns. A closure derived from candidate edges is still candidate-grade.
+- **B7 learned prioritization:** NSL may rank which ablations to spend on; graph adjacency need not mean positive similarity (prevention is signed/directional). ONNX prediction != semantic proof.
+- **B8 compiler bridge:** MemorySSA offers may-def/use analogies; DialEgg can explore *admitted* MLIR rewrites, but Rust effect/borrow/drop/panic/unsafe constraints cannot be optimized away by observed equivalence.
+- **B9 feedback:** revisions and independently admitted rules travel by existing provenance/recipes; no new runtime, no source writes and no unguarded supersession.
+
+## Falsifier seeds for review
+| Seed | Small counterexample | Mandatory classification |
+|---|---|---|
+| K1 | A,B always co-fire but never read/write each other | association only, no required dependency |
+| K2 | A,B independently produce same needed value | both single ablations silent, joint ablation changes outcome |
+| K3 | C requires A AND B | conjunction hyperedge, not independent A→C or B→C |
+| K4 | A disables B only for guard G | conditional prevention, not global inhibition |
+| K5 | two methods share a cube coordinate in same base revision | both identities retained; no implicit last-wins |
+| K6 | unknown WHEN condition and incomplete authorization | UNKNOWN/REFUSED, not false or true |
+| K7 | final state equal but an intermediate side effect/exception differs | cannot squash or certify equivalence |
+| K8 | two thoughts replay identical source roots | do not double count independent evidence |
+| K9 | same cube ordinals but different source version/codebook | typed refusal on stale binding |
+| K10 | learned ONNX predictor ranks wrong on held-out family | no silent deployment or authority |
+
+## Source anchors
+- #1605 Thought-Git porcelain/history; #1604 candidates; #1684 replay/commit witness; #1739 protected squash
+- #1799 DT64 capstone/Q0; #1797 Thought-Git/edge projections; #1747 DR64³ research/supersession
+- #1313 ARM population evidence; #1359 SPOFC do/learn; #1584 learned association measures
+- #1760 Rust Codegraph → Mathematics → Logic → Tracert
+- LLVM MemorySSA: https://llvm.org/docs/MemorySSA.html
+- DialEgg CGO 2025: https://2025.cgo.org/details/cgo-2025-papers/44/DialEgg-Dialect-Agnostic-MLIR-Optimizer-using-Equality-Saturation-with-Egglog
+- TensorFlow NSL: https://github.com/tensorflow/neural-structured-learning
+
+## Read-out template (fill after actual probe, NOT NOW)
+Fixture + pinned versions | source/coverage | baseline observable | intervention semantics | thought slot/receipt | exact counterfactual diff | independent oracle | competing explanation | grade (observed/model/inferred/proven) | falsifier result | cost | decision/refusal | next breadcrumb.