feat(unblock): sustained 10min zero-wall loop + 10-agent SHIM-CD-01 first thin SIP probe design (VectorSteerer minimal guarded diff + full review package)

docs/steering_chelation_rag_dag_research/README.md

@@ -0,0 +1,32 @@
+# Steering Node + Chelation + Adaptive RAG-DAG + MicroSLM Research Program
+
+This directory contains the canonical artifacts for the 2026 research program that converges ChelatedAI's five strongest threads:
+
+- Adaptive spectral chelation and self-healing correction
+- TTS (Translation-Transport-Steering) inference-time vector relocation nodes
+- Model-Scope sparse feature steering and hook infrastructure
+- Computational-storage block graphs, drive-node speculative racing, and repo graph memory
+- EGGROLL-style hyperscale evolution strategies + OPSD on-policy self-distillation patterns (from the 2026-05 Loop 01 swarm)
+
+**Into a single substrate**: a live-mutable RAG-DAG where chelation variance acts as the primary "reconsider this neighborhood / spawn reroutes" signal, steering nodes can cast multiple vector reroutes or propose new token routes, a 2-4 GB micro SLM learns the route policy, and drive-node / low-rank population dispatch provides the graph execution and convergence mechanism.
+
+## Canonical Documents (start here)
+- `STEERING_CHELATION_RAGDAG_MICROSLM_RESEARCH_PLAN.md` — full 10-loop program definition, thesis, scope locks, success criteria.
+- `STEERING_CHELATION_BHS_RESEARCH_RUBRIC.md` — BHS v3.3 extensions specific to route metrics, micro-SLM gates, drive-node evidence, and carried-debt re-audit requirements.
+- `STEERING_CHELATION_10_LOOP_BHS_PROGRAM.md` — loop definitions and execution rules.
+- `shim_nodes_mtp_lookahead_nomenclature.md` — **Major new primitive**: formal definition of Shim Nodes, Shim Vectors, Compounding Cascades, MTP Shim Lookahead, Precomputed Shims, Shim Backdoors, SE-RDAG, and integration points with existing surfaces (FeatureDirectionBank, TTS, Model-Scope, OPSD, block graphs).
+- `loop_01/` — Loop 1 (Deep Research & Mapping) artifacts as they land. Includes kickoff brief and literature starter.
+
+## Relationship to Other Work
+- Re-uses (does not duplicate) the entire `docs/chelation_opsd_research/` OPSD 10-loop program and its Loop 01 synthesis.
+- Builds directly on `docs/model-scope-steering-architecture-2026-05-01.md`, `docs/evolution-strategies-hyperscale-chelatedai-analysis.md`, `docs/COMPUTATIONAL_STORAGE_DRIVE_NODES.md`, and the TTS pipeline.
+- All new work must follow the brutal-honesty convention (CLAUDE.md + conventions/brutal-honesty-rulebook.md). No pattern or micro-SLM configuration is promoted without a complete evidence chain and Tier B review scoring 100.
+
+## Quick Navigation
+- For the "why now" and architecture vision: read the top-level plan.
+- For exact trigger phrases and what "evidence" means in this program: read the rubric.
+- For the immediate next work (Loop 1): read `loop_01/00_kickoff_brief.md`.
+
+This program exists because the repo already has the raw material for something more powerful than another incremental chelation adapter or another static GraphRAG. The convergence work is the point.
+
+*Program kickoff — 2026-05*

docs/steering_chelation_rag_dag_research/STEERING_CHELATION_10_LOOP_BHS_PROGRAM.md

@@ -0,0 +1,76 @@
+# STEERING + CHELATION + ADAPTIVE RAG-DAG + MICROS LM
+## 10-Loop BHS Research Program Definition
+
+**Version**: Kickoff (extends the successful CHELATION_OPSD_10_LOOP_BHS_PROGRAM structure)
+
+## Program Identity
+**Full Name**: Steering Node Chelation for Adaptive RAG-DAGs with Hyperscale Graph Convergence and MicroSLM Proof-of-Concept  
+**Short**: Steering-Chelation-RAGDAG-MicroSLM Program  
+**Repository Home**: `docs/steering_chelation_rag_dag_research/`  
+**Governance**: BHS v3.3 (see `STEERING_CHELATION_BHS_RESEARCH_RUBRIC.md` and parent conventions)  
+**Cross-References**: All prior OPSD Loop 01 artifacts, Model-Scope steering architecture (2026-05-01), EGGROLL hyperscale analysis, Computational Storage Drive Nodes doc, TTS pipeline, full AEP archive.
+
+## The 10 Loops (Canonical)
+
+1. **Deep Research & Mapping**  
+   Literature (LogicRAG dynamic DAGs, SAE-RSV + Matryoshka SAEs for steering, latest OPSD/SDPO variants, spectral embedding methods, hyperscale ES extensions) + ruthless audit of the *five connected substrates* (Chelation, TTS Steering Nodes, Model-Scope, Comp-Storage graphs/drive nodes, EGGROLL/OPSD optimizer surfaces). Produce master pain-point → technique mapping and Tier S/A/B upgrade patterns.
+
+2. **Architecture Design**  
+   Concrete specs for `RerouteDAG` abstraction, multi-cast steering node interface, micro-SLM I/O contract (chelation signals + sparse features → route proposals), drive-node dispatch contract for candidate routes, artifact card + promotion schema extensions. Multiple viable patterns sketched with pseudocode and dependency DAGs.
+
+3. **Loss Function & Training Regime Variants**  
+   OPSD asymmetric privileged-diagnostic for reroute traces, route-cohesion auxiliary losses, structural health regularizers, quantization-aware objectives, low-rank population (EGGROLL-style) vs gradient hybrids. Target: stable on-policy self-improvement of the route policy without KL shocks in embedding or DAG space.
+
+4. **Stability, KL Control & Route Forgetting Mitigation**  
+   The hardest practical problem. Mechanisms to prevent accepted reroutes from destroying previously reliable retrieval facts or base model compatibility. Includes retention replay families specific to route histories, anchoring to privileged successful traces, and bounded actuator constraints extended to DAG mutations.
+
+5. **Sample Efficiency & Data Filtering**  
+   MIS-PO / hard-negative / attribution-style filtering applied to reroute proposal traces. How to decide which noisy neighborhoods or failed routes are worth spending micro-SLM capacity and multi-path speculation budget on. Budget-aware collection policy as first-class citizen.
+
+6. **Quantization-Aware & Low-Rank / Bounded Variants**  
+   Everything (micro-SLM head, steering actuators, route proposal generators) must survive the same INT8/BoundedAdapter floor that production chelation already targets. Low-rank route deltas, Matryoshka-style nested steering features, block-graph friendly representations for drive-node paths.
+
+7. **Self-Edit Directive + Steering Node + DAG Mutation Integration**  
+   Extend `SelfEditDirective` (and the self-healing ledger) so that generated directives can propose *live DAG topology changes* and multi-vector reroute sets, not just embedding adapter corrections. Close the loop between diagnostics → directive → steering node execution → outcome fitness → distillation back into the policy.
+
+8. **Evaluation Framework & Benchmark Design**  
+   Extend synthetic collapse fixture and road-course harnesses to DAG/reroute tasks. New "route acceptance under noise" family of benchmarks. Drive-node latency parity surfaces (where in scope). Full transfer testing (BEIR + new DAG reasoning tasks). Artifact card + verifier card automation for every candidate.
+
+9. **Implementation of Top Patterns + Tests + MicroSLM Smoke**  
+   Ship the 2-3 highest-ranked patterns from Loops 2-8 as working, evidence-backed slices. First real training of a 2-4 GB class micro-SLM (or its steering head) against the substrate. Full BHS evidence packages. Smoke pipeline that exercises the entire chain on a fresh checkout.
+
+10. **Comparative Analysis, Recommendations & Final Upgrade Roadmap**  
+    What actually delivered lift under BHS gates. What was rejected and why (with data). Concrete ship/no-ship decisions for main. Updated productionization plan for the winning substrate (how it wires into AntigravityEngine / Model-Scope / existing vector store). Clear statement of remaining carried debt and recommended next program (if any).
+
+## Loop Execution Rules (identical spirit to OPSD program)
+- Each loop is executed by one or more specialized agents (literature, substrate audit, architecture, loss design, etc.) + integration lead for synthesis.
+- Every loop ends with a `loop_N/NN_synthesis_and_prioritization.md` (or equivalent) that contains the ranked patterns, BHS self-assessment, updated program score, and explicit carried-debt re-audit.
+- Code or architecture artifacts from a loop only become "official" for the program after the synthesis document is written and the loop is closed.
+- Parallel swarm execution is encouraged for speed (as proven in OPSD Loop 01), but integration/synthesis is serial and owned by the orchestrator.
+- No loop may be declared complete until the BHS rubric items for that loop's deliverables are satisfied (evidence, brutal honesty, no L1-L13 violations).
+
+## Entry Criteria for Starting Loop N
+- Loop N-1 synthesis + all supporting agent docs committed.
+- Explicit "Loop N Kickoff Brief" (one-pager) that names the 3-5 concrete questions the loop must answer and the minimal evidence surface required to close it.
+- Carried debt items from prior loops either closed or explicitly carried with mitigation plan.
+
+## Exit Criteria for the Whole Program (Loop 10 close)
+- At least one pattern or micro-SLM configuration has a complete, independently reviewed BHS evidence chain (artifact card, replay, holdout, quant gate, rollback demo, route-specific metrics) scoring 100.
+- The program can state with evidence which of the original thesis claims are supported, refuted, or still open.
+- Clear recommendation: "Ship X to main as the new default reroute substrate", "Retain Y as experimental in `computational_storage_poc/` or a feature branch", "Pivot Z because fundamental blocker discovered".
+- Updated docs for consumers (engineers who will actually use the new surfaces).
+
+## Relationship to Prior Work (explicit, non-duplicative)
+This program *does not* restart chelation or OPSD research. It treats the 2026-05 OPSD Loop 01 outputs as *input substrate* (the pain-point mappings and candidate upgrade patterns for the chelation/self-edit layer are directly reusable). It adds the missing "graph + steering node + micro SLM + drive dispatch + hyperscale ES" dimensions that turn a per-vector correction system into a live reroutable reasoning graph.
+
+It also does not duplicate the Model-Scope steering architecture doc or the computational-storage POC; those are the *raw material* being converged.
+
+## Scope Locks (non-negotiable for this tranche)
+- Full agent harnesses / long-horizon planning loops: out of scope (inspiration only, per frontier adaptive overlay decisions).
+- Claiming real hardware LLM inference on spinning rust or SSDs: scope-locked to existing computational-storage transport + dispatch contracts + emulation. New RP2040 evidence only if actually captured on hardware.
+- Mutating base 2-4 GB model weights in the proof: forbidden without explicit exception + full rollback evidence. Adapters, steering heads, and feature banks only.
+- "It worked in simulation therefore it ships": never. Every surface must have a smoke/replay path that a fresh checkout can run.
+
+**This document is the canonical program definition.** Update it only at loop closeouts with a new version note and diff summary.
+
+*Program kickoff — 2026-05*

docs/steering_chelation_rag_dag_research/STEERING_CHELATION_BHS_RESEARCH_RUBRIC.md

@@ -0,0 +1,49 @@
+# BHS Research Rubric — Steering Node + Chelation + RAG-DAG + MicroSLM Program
+
+**Baseline**: Inherits and extends `docs/chelation_opsd_research/CHELATION_OPSD_BHS_RESEARCH_RUBRIC.md` (v from 2026-05-15 OPSD swarm). All BHS v3.3 rules from `docs/conventions/brutal-honesty-rulebook.md` and CLAUDE.md apply with no exceptions.
+
+**Program-Specific Severity and Evidence Rules** (additions only):
+
+## Route-Specific Metrics (must appear in every candidate report / artifact card)
+- **Reroute Acceptance Rate under Controlled Noise**: % of synthetic or real noisy neighborhoods where the system elects to cast ≥1 reroute and at least one improves the downstream metric vs baseline.
+- **Route Cohesion Score**: Topology / isomer-style metric over the *proposed route set* (not just final top-k). Penalizes semantically divergent or high-variance route families.
+- **Rollback Success Rate**: After accepting a reroute or micro-SLM update, ability to revert to pre-reroute baseline behavior on replay with zero or bounded regression.
+- **Quantization Survival Delta**: NDCG / route acceptance delta when all actuators + micro-SLM head run under the same INT8/Bounded constraints as production.
+- **Budget-Adjusted Lift**: Primary metric must be reported both raw and normalized by extra retrieval / steering / ES-population tokens or drive-node dispatches used.
+
+## Micro-SLM Specific Gates (before any training run is considered "evidence")
+- Base model compatibility: frozen weights + adapter/steering head only. Any run that mutates the 2-4 GB core without explicit exception + rollback evidence is L4 (partial as complete).
+- Legacy case retention: on a fixed set of "old base cases" (SciFact clean, NFCorpus, plus 3-5 curated retrieval facts from prior road-courses), the micro-SLM configuration must not regress below pre-registered tolerance without the actuator being disabled.
+- Training data provenance card: every trace used for OPSD-style privileged vs student must carry (source, collapse-severity or route-failure label, success/failure outcome, checksum of prompt+context that produced it).
+
+## Drive-Node / Graph Substrate Evidence
+- Any claim involving `computational_storage_poc/` dispatch or block-graph candidate evaluation must include parity evidence between software replay and the dispatch path (existing `test_computational_storage_*` discipline).
+- Speculative multi-path claims must report both latency model and correctness parity; "faster in simulation" without parity is not evidence.
+- RP2040 or real hardware only for transport contract verification per existing retention policy. Emulation + mock_array is the ceiling for this program unless new hardware evidence is captured.
+
+## Promotion / Rejection Language (mandatory phrasing)
+- "Promoted to Tier S candidate": only after full artifact card + replay + holdout + quant gate + BHS_TIER_B = 100.
+- "Shows directional promise on X but failed Y gate — retained as guarded research": the honest default for most early loops.
+- "Rejected for this tranche — fundamental instability under Z condition": when a pattern repeatedly produces unrecoverable regressions or KL-shock analogs.
+
+## Carried Debt from Prior Sessions (must be re-audited in Loop 1)
+- No default chelation profile has survived multi-task confirmation (Sessions 32-34).
+- Adaptive overlay / learned gate work is strong on instrumentation but weak on promotion.
+- Computational-storage drive-node claims remain scope-locked to transport + software parity.
+- Model-Scope steering is in shadow/advisory mode only; no production steering actuator has a full evidence chain yet.
+
+Every Loop N synthesis must contain an explicit "Carried Debt Re-audit" section that says for each prior debt item: "Still open / Partially addressed by <new surface> / Closed by evidence in <artifact>".
+
+**BHS Research Score for this Program**:
+- Starts at program kickoff with the honest score of the *connected prior surfaces* (OPSD Loop 01 synthesis was strong; Model-Scope and TTS are instrumented but not yet promoted; EGGROLL mapping is analysis only).
+- Each loop must publish an updated program-level BHS Research Score (0-100) with the same 5-iteration / Tier B discipline as PRs where relevant.
+- A score < 70 at the end of Loop 5 triggers mandatory scope reduction or pivot review before continuing.
+
+**Trigger Phrases for This Program** (use in every agent dispatch and review):
+- "Be brutally honest about the reroute."
+- "If I disable the steering node / micro-SLM head, what visible behavior on the DAG changes?"
+- "Show me the route cohesion and rollback evidence, not the test."
+- "What fraction of the claimed lift survives when we force the same quantization floor and token budget as the baseline?"
+- "Is this a new route or just a prettier way to describe an old chelation adapter?"
+
+*This rubric is living. Update it in every loop closeout.*