Add presentation materials for technical talk on transformer training

Five standalone HTML visuals covering the RSA pipeline, training pipeline,
gate decomposition, dataset anatomy, and model tier sizing. Slide deck brief
describing a 35-slide two-act deck for a data science audience. Handoff note
documenting what was created this session.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

presentations/slide_deck_brief.md

@@ -0,0 +1,363 @@
+# Slide Deck Brief — whetstone_RSA: Tiny Models for Bounded Decisions
+
+**Audience:** Whetstone_RSA, ex data science instructor. Technical depth expected.
+Wants to understand the actual transformer training approach, not just the concept.
+
+**Format:** Slide deck, dark theme preferred. One idea per slide, minimal text.
+Visuals are in the same directory as this file (HTML files — embed or screenshot them).
+
+**Length:** ~30–35 slides across two acts.
+
+---
+
+## Act I — The Problem and the Pattern (Slides 1–17)
+
+These slides are already written in `rsa_tiny_models_talk.md`. Keep them as-is.
+They establish the motivation: fuzzy input → bounded output → deterministic execution.
+Key points already covered:
+- Why regex is brittle
+- Why LLMs are oversized for bounded tasks
+- What RSA does (contract layer between user and tool)
+- The core pipeline: input → RSA gate → typed contract → deterministic software
+- Where this pattern shows up (command palettes, search bars, workflow routing)
+
+**Visual for Act I:** `visual_rsa_pipeline.html`
+
+---
+
+## Act II — The Actual Training Work (Slides 18–35)
+
+This is the new material. It covers the real implementation, dataset, and results.
+
+---
+
+### Slide 18: The Case Study
+
+**Title:** First Real Target — whetstone_DSL
+
+**Content:**
+- whetstone_DSL is a C++ IDE with an AST-first codegen pipeline
+- It uses an LLM to make many bounded decisions bundled together
+- Goal: replace those bundles with one tiny specialist per decision gate
+- First gate selected: `prereq_op_selector`
+  - Decides which prerequisite operations a taskitem needs before it can execute
+
+**Speaker notes:** This is not a toy example. These gates run on real pipeline decisions.
+
+---
+
+### Slide 19: What a Gate Is
+
+**Title:** The Gate Contract
+
+**Content:**
+```
+input:   task title · requirements · acceptance criteria · constraints
+output:  needs_resolve_dependencies: bool
+         needs_architect_review: bool
+         needs_validate_intake: bool  (always true — dropped)
+         confidence: float
+         abstain: bool
+```
+
+**Visual idea:** Simple box diagram: structured text in, typed struct out.
+
+**Speaker notes:** The output is not generated text. It is a typed struct with known fields.
+
+---
+
+### Slide 20: The Training Pipeline Overview
+
+**Title:** From Run Artifacts to Specialist
+
+**Visual:** `visual_training_pipeline.html` — use full slide.
+
+**Speaker notes:** Six-step pipeline. Real data from real pipeline runs. The training
+framework (Fabricate) lives on the GPU desktop — can demo it live there.
+
+---
+
+### Slide 21: Step 1 — The Corpus
+
+**Title:** Source Data: 1000 Pipeline Runs
+
+**Content:**
+- whetstone_DSL was run 1000 times on real and synthetic projects
+- Each run produces JSON log artifacts recording every taskitem decision
+- `extract_gate_rows.py` parses these into per-gate supervision rows
+
+**Visual idea:** Stack of JSON files on left → extraction tool → rows on right.
+
+---
+
+### Slide 22: Step 2 — Gate Extraction
+
+**Title:** Isolating One Decision at a Time
+
+**Content:**
+- Each row contains: decision_context · decision_label · decision_provenance
+- Only the information available at decision time is allowed in the context
+- Rows with schema-drift labels are rejected (24 of 1321 total)
+- Result: 1297 clean rows for prereq_op_selector
+
+**Key extraction rule:** Never leak downstream fields into the context.
+A gate must only see what was available when it would actually fire.
+
+---
+
+### Slide 23: The Discovery — Gate Decomposition
+
+**Title:** needs_validate_intake Is Always True
+
+**Content:**
+- Analysis of 1297 rows: needs_validate_intake = True in all 1297 of them
+- A label with zero variance cannot be learned — it is a constant, not a gate
+- Real gate decomposes into two independent binary classifiers:
+  - needs_resolve_dependencies (76.9% positive rate)
+  - needs_architect_review (19.8% positive rate)
+
+**Visual:** `visual_gate_decomposition.html`
+
+**Speaker notes:** This is a real finding, not a design choice made upfront. The data
+told us the gate shape was wrong. That's the process working correctly.
+
+---
+
+### Slide 24: Dataset Anatomy
+
+**Title:** What 1297 Rows Actually Looks Like
+
+**Visual:** `visual_dataset_anatomy.html` — use full slide.
+
+**Speaker notes:** The ~17 unique templates is the important number. High row count
+does not mean high diversity. This is the honest accounting.
+
+---
+
+### Slide 25: The TSV Format
+
+**Title:** Fabricate Training Format
+
+**Content:**
+```
+label <TAB> hop <TAB> text
+```
+- label: 0 (no) or 1 (yes) — one file per binary gate
+- hop: always 0 (unused routing field in Fabricate)
+- text: composed from title · reasons · acceptance criteria · constraints
+
+**Four files produced:**
+- prereq_resolve_train.tsv (1102 rows)
+- prereq_resolve_eval.tsv (195 rows)
+- prereq_architect_train.tsv (1102 rows)
+- prereq_architect_eval.tsv (195 rows)
+
+**Split:** 85% train / 15% eval, random seed 42.
+
+---
+
+### Slide 26: The Model — Tiny Transformer
+
+**Title:** ~213K Parameters. ~800KB on Disk.
+
+**Content:**
+- Framework: Fabricate (custom tiny transformer training tool)
+- Architecture: transformer encoder with binary classification head
+- ~213K parameters per specialist
+- Serialized checkpoint: ~800KB
+- Training: GPU desktop, 4000 steps per specialist
+- Inference: CPU-friendly, local, no network dependency
+
+**Visual idea:** Size comparison — model checkpoint vs. a JPEG photo (~800KB each).
+
+**Speaker notes:** This is the whole point. A model smaller than an average image file,
+making a bounded yes/no decision. Can demo inference live on the GPU desktop.
+
+---
+
+### Slide 27: Training Results
+
+**Title:** Baseline Results @ 4000 Steps
+
+**Content:**
+
+| Specialist | Accuracy | Steps | Positive rate |
+|---|---|---|---|
+| prereq_resolve (needs_resolve_dependencies) | 69.4% | 4000 | 76.9% |
+| prereq_architect (needs_architect_review) | 75.0% | 4000 | 19.8% |
+
+**Important context:**
+- A majority classifier on needs_resolve would score ~77% (model is below naive baseline)
+- A majority classifier on needs_architect would score ~80% (model is learning minority class)
+- These are weak baselines — expected given the 17-template corpus
+
+**Visual idea:** Bar chart with majority-classifier baseline marked as a reference line.
+
+---
+
+### Slide 28: Why the Baselines Are Weak (And Why That's Fine)
+
+**Title:** The Corpus Was Not Designed for This
+
+**Content:**
+- The taskitem schema was designed to record execution outputs, not discriminative inputs
+- 1297 rows, but only ~17 unique text patterns — the model has almost no signal variation
+- The labels are correct; the features are not rich enough yet
+
+**The fix is not more data from the same corpus.**
+The fix is adding RSA-native fields to the taskitem schema upstream.
+
+**Speaker notes:** This is an architectural finding. The pipeline methodology is validated.
+Signal quality is a schema problem, not a model problem.
+
+---
+
+### Slide 29: What Would Make the Signal Better
+
+**Title:** Better Input = Better Gate
+
+**Content:**
+Each taskitem needs to carry:
+- `architectural_surface` — what part of the system is touched
+- `cross_component_deps` — explicit dependency flags
+- `security_sensitive` — flag for security-relevant changes
+- `uncertainty_score` — explicit uncertainty signal
+
+With these fields, some gates that currently need ML may become deterministic rules.
+The RSA entropy threshold shifts as the contract gets richer.
+
+**Visual idea:** Taskitem schema before vs. after adding RSA fields.
+
+---
+
+### Slide 30: The Entropy → Model Size Hypothesis
+
+**Title:** Match the Model to the Decision Surface
+
+**Visual:** `visual_model_tiers.html`
+
+**Speaker notes:** The hypothesis is measurable and falsifiable. The experiment is:
+train each tier on the same gate, record the smallest one that clears the target.
+
+---
+
+### Slide 31: Gate Entropy Factors
+
+**Title:** What Makes a Gate Hard?
+
+**Content:**
+Factors that increase required model tier:
+- output_cardinality (more labels = harder)
+- class_imbalance
+- lexical_variation in inputs
+- context_width_required
+- label_boundary_fuzziness
+- world_knowledge_dependence
+- slot_interdependence
+
+**Target deliverable:**
+```
+gate_id | entropy_score | smallest_passing_tier | accuracy | latency_ms
+```
+
+**Speaker notes:** We don't have this table yet. It's the next experiment.
+
+---
+
+### Slide 32: What's Next — SPRINT-004
+
+**Title:** Pipeline Decision Audit
+
+**Content:**
+Before training more gates, map every decision in the full whetstone_DSL pipeline:
+- Start at RegisterArchitectIntakeTools.h (pipeline entry point)
+- At every stage: what decision is being made?
+- What inputs are available?
+- What is the optimal method: deterministic / RSA / SLM / LLM / human?
+
+`generate_taskitems` alone bundles 10+ separable decisions.
+
+**Output:** `docs/case_studies/whetstone_dsl_pipeline_decision_map.md`
+
+**Speaker notes:** Train data collection strategy follows from the decision map.
+Don't collect data for the wrong gate shape.
+
+---
+
+### Slide 33: What Is Already Deterministic
+
+**Title:** Most of Whetstone Is Already Right
+
+**Content:**
+Once the AST and target language are fixed, most current code generation is deterministic:
+- Target generator dispatch — deterministic
+- AST concept dispatch — deterministic
+- Environment gating — deterministic
+- C++ raising profile validation — deterministic
+
+RSA is highest value in:
+- taskitem generation
+- execution contract shaping
+- routing decisions
+- pre-codegen profile decisions
+
+**Speaker notes:** This is a feature, not a gap. Don't replace deterministic software
+with models. Find the actual ambiguity boundary.
+
+---
+
+### Slide 34: The C++ Runtime
+
+**Title:** whetstone_RSA Runtime Scaffold
+
+**Content:**
+A C++ library wrapping the gate contracts:
+- `GateDefinition` · `GateEvidence` · `GateDiagnosis` · `GateSuitability`
+- `ProbeRegistry` + `DiagnosisEngine`
+- First concrete probe: `confidence_threshold_probe`
+- Builds with CMake, demo executable runs
+
+This is the deployment layer — gates are loaded here at runtime.
+
+**Visual idea:** Stack diagram: Fabricate specialist → RSA runtime → whetstone_DSL pipeline.
+
+---
+
+### Slide 35: The Summary
+
+**Title:** What This Is
+
+**Content:**
+1. Real, running pipeline decisions in whetstone_DSL
+2. Real extracted training data (1297 rows, two binary gates)
+3. Real trained specialists (~800KB each, ~213K params)
+4. Honest baselines (weak corpus, clear reason, clear fix)
+5. A measurable hypothesis: entropy → tier selection
+6. A C++ runtime that loads and runs them
+
+**Not a chatbot. Not a search engine.**
+A bounded decision contract layer, with real data and real results.
+
+---
+
+## Appendix Slides (optional, for Q&A)
+
+- Rejection taxonomy (schema drift, sibling-uniform labels)
+- Full gate inventory for whetstone_DSL
+- The decision_contract_ontology (input_structure, output_topology, policy_stability, etc.)
+- Backend selection taxonomy (transformer vs. gradient boosting vs. Bayes — kept separate from gate identity)
+- The abstain/escalate policy design
+
+---
+
+## Files in this directory
+
+| File | Description |
+|---|---|
+| `rsa_tiny_models_talk.md` | Act I source (31 slides, already complete) |
+| `slide_deck_brief.md` | This file |
+| `visual_rsa_pipeline.html` | The core RSA pattern — fuzzy in, contract out |
+| `visual_training_pipeline.html` | Six-step pipeline from run artifacts to trained specialist |
+| `visual_gate_decomposition.html` | prereq_op_selector decomposed into two binary gates |
+| `visual_dataset_anatomy.html` | Dataset stats, class balance, TSV format sample |
+| `visual_model_tiers.html` | Entropy → model tier sizing diagram |

presentations/visual_dataset_anatomy.html

@@ -0,0 +1,140 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="UTF-8">
+<title>Dataset Anatomy</title>
+<style>
+  * { box-sizing: border-box; margin: 0; padding: 0; }
+  body { background: #0f1117; font-family: 'Segoe UI', system-ui, sans-serif; display: flex; align-items: center; justify-content: center; min-height: 100vh; }
+  .slide { width: 960px; padding: 56px 72px; }
+  h2 { color: #e2e8f0; font-size: 1.5rem; font-weight: 400; letter-spacing: 0.05em; text-transform: uppercase; margin-bottom: 48px; text-align: center; }
+
+  .top-stats { display: flex; gap: 20px; margin-bottom: 36px; }
+  .stat-card { flex: 1; background: #161b27; border-radius: 10px; padding: 22px 20px; text-align: center; border: 1.5px solid #2d3748; }
+  .stat-num { font-size: 2.2rem; font-weight: 700; margin-bottom: 4px; }
+  .stat-label { font-size: 0.72rem; letter-spacing: 0.1em; text-transform: uppercase; color: #6b7280; }
+  .c1 .stat-num { color: #60a5fa; }
+  .c2 .stat-num { color: #4ade80; }
+  .c3 .stat-num { color: #f87171; }
+  .c4 .stat-num { color: #fbbf24; }
+
+  .mid { display: flex; gap: 24px; margin-bottom: 28px; }
+
+  .class-balance { flex: 1; background: #161b27; border-radius: 10px; padding: 22px 22px; border: 1.5px solid #2d3748; }
+  .section-title { font-size: 0.65rem; letter-spacing: 0.12em; text-transform: uppercase; color: #6b7280; margin-bottom: 16px; }
+
+  .bar-row { display: flex; align-items: center; gap: 12px; margin-bottom: 14px; }
+  .bar-name { font-size: 0.82rem; color: #94a3b8; width: 200px; flex-shrink: 0; }
+  .bar-track { flex: 1; height: 22px; background: #1e2433; border-radius: 4px; overflow: hidden; position: relative; }
+  .bar-fill { height: 100%; border-radius: 4px; display: flex; align-items: center; padding-left: 10px; font-size: 0.72rem; font-weight: 600; white-space: nowrap; }
+  .bar-fill.pos-resolve   { background: linear-gradient(90deg, #1d4ed8 0%, #3b82f6 100%); color: #bfdbfe; width: 76.9%; }
+  .bar-fill.neg-resolve   { background: #1e2433; color: #4a5568; width: 23.1%; }
+  .bar-fill.pos-architect { background: linear-gradient(90deg, #5b21b6 0%, #8b5cf6 100%); color: #ddd6fe; width: 19.8%; }
+  .bar-fill.neg-architect { background: #1e2433; color: #4a5568; width: 80.2%; }
+  .bar-pct { font-size: 0.75rem; color: #6b7280; width: 40px; text-align: right; flex-shrink: 0; }
+
+  .corpus-issue { flex: 0 0 300px; background: #1a1500; border: 1.5px solid #f59e0b; border-radius: 10px; padding: 22px; }
+  .corpus-issue .section-title { color: #92400e; }
+  .issue-item { display: flex; gap: 10px; margin-bottom: 12px; align-items: flex-start; }
+  .issue-bullet { color: #f59e0b; flex-shrink: 0; font-size: 0.85rem; margin-top: 2px; }
+  .issue-text { font-size: 0.82rem; color: #d97706; line-height: 1.5; }
+  .issue-text strong { color: #fbbf24; }
+
+  .tsv-preview { background: #161b27; border-radius: 10px; padding: 22px 24px; border: 1.5px solid #2d3748; }
+  .tsv-header { display: flex; gap: 0; margin-bottom: 14px; }
+  .col-head { font-size: 0.65rem; letter-spacing: 0.1em; text-transform: uppercase; color: #6b7280; padding: 0 12px; border-right: 1px solid #2d3748; }
+  .col-head:first-child { padding-left: 0; width: 70px; }
+  .col-head:nth-child(2) { width: 60px; }
+  .col-head:last-child { flex: 1; border-right: none; }
+  .tsv-row { display: flex; gap: 0; padding: 7px 0; border-bottom: 1px solid #1a1f2e; font-family: 'Courier New', monospace; }
+  .tsv-row:last-child { border-bottom: none; }
+  .col-val { font-size: 0.78rem; padding: 0 12px; border-right: 1px solid #1e2433; white-space: nowrap; overflow: hidden; text-overflow: ellipsis; }
+  .col-val:first-child { padding-left: 0; width: 70px; font-weight: 700; }
+  .col-val:nth-child(2) { width: 60px; color: #4a5568; }
+  .col-val:last-child { flex: 1; border-right: none; color: #94a3b8; }
+  .v1 { color: #4ade80; }
+  .v0 { color: #6b7280; }
+</style>
+</head>
+<body>
+<div class="slide">
+  <h2>Dataset Anatomy — prereq_op_selector</h2>
+
+  <div class="top-stats">
+    <div class="stat-card c1"><div class="stat-num">1000</div><div class="stat-label">runs extracted</div></div>
+    <div class="stat-card c2"><div class="stat-num">1297</div><div class="stat-label">accepted rows</div></div>
+    <div class="stat-card c3"><div class="stat-num">24</div><div class="stat-label">schema-drift rejected</div></div>
+    <div class="stat-card c4"><div class="stat-num">~17</div><div class="stat-label">unique text templates</div></div>
+  </div>
+
+  <div class="mid">
+    <div class="class-balance">
+      <div class="section-title">Class balance</div>
+      <div class="bar-row">
+        <div class="bar-name">needs_resolve_dependencies = yes</div>
+        <div class="bar-track"><div class="bar-fill pos-resolve">997 rows</div></div>
+        <div class="bar-pct">76.9%</div>
+      </div>
+      <div class="bar-row">
+        <div class="bar-name">needs_resolve_dependencies = no</div>
+        <div class="bar-track"><div class="bar-fill neg-resolve">300 rows</div></div>
+        <div class="bar-pct">23.1%</div>
+      </div>
+      <div style="height:10px"></div>
+      <div class="bar-row">
+        <div class="bar-name">needs_architect_review = yes</div>
+        <div class="bar-track"><div class="bar-fill pos-architect">257 rows</div></div>
+        <div class="bar-pct">19.8%</div>
+      </div>
+      <div class="bar-row">
+        <div class="bar-name">needs_architect_review = no</div>
+        <div class="bar-track"><div class="bar-fill neg-architect">1040 rows</div></div>
+        <div class="bar-pct">80.2%</div>
+      </div>
+    </div>
+
+    <div class="corpus-issue">
+      <div class="section-title">⚠ Corpus limitation</div>
+      <div class="issue-item">
+        <div class="issue-bullet">▸</div>
+        <div class="issue-text">Only <strong>~17 unique text templates</strong> across 1297 rows. High row count masks low actual diversity.</div>
+      </div>
+      <div class="issue-item">
+        <div class="issue-bullet">▸</div>
+        <div class="issue-text">The taskitem schema records <strong>outputs</strong>, not discriminative inputs. RSA-relevant fields were never stored.</div>
+      </div>
+      <div class="issue-item">
+        <div class="issue-bullet">▸</div>
+        <div class="issue-text">Baselines are <strong>intentionally weak</strong>. Pipeline is correct; signal improves with schema changes upstream.</div>
+      </div>
+    </div>
+  </div>
+
+  <div class="tsv-preview">
+    <div class="section-title">Fabricate TSV format — prereq_resolve_train.tsv (sample rows)</div>
+    <div class="tsv-header">
+      <div class="col-head">label</div>
+      <div class="col-head">hop</div>
+      <div class="col-head">text (title · reasons · accepts · constraints)</div>
+    </div>
+    <div class="tsv-row">
+      <div class="col-val v1">1</div><div class="col-val">0</div>
+      <div class="col-val">title: Execution Readiness Primary | reasons: high_confidence_clear_path | accepts: at least one tool selection event; no malformed mcp payloads</div>
+    </div>
+    <div class="tsv-row">
+      <div class="col-val v0">0</div><div class="col-val">0</div>
+      <div class="col-val">title: Intake Foundation Primary | reasons: high_confidence_clear_path | accepts: deterministic taskitems with strict execution contracts | constraints: primary language go</div>
+    </div>
+    <div class="tsv-row">
+      <div class="col-val v1">1</div><div class="col-val">0</div>
+      <div class="col-val">title: Intake Foundation Primary | reasons: high_confidence_clear_path | accepts: scenario 1 given when then; scenario 2 given when then | constraints: language runtime constraints</div>
+    </div>
+    <div class="tsv-row">
+      <div class="col-val v0">0</div><div class="col-val">0</div>
+      <div class="col-val">title: Execution Readiness Primary | reasons: high_confidence_clear_path | accepts: at least one tool execution event; no unbounded loops</div>
+    </div>
+  </div>
+
+</div>
+</body>
+</html>

presentations/visual_gate_decomposition.html

@@ -0,0 +1,146 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="UTF-8">
+<title>Gate Decomposition</title>
+<style>
+  * { box-sizing: border-box; margin: 0; padding: 0; }
+  body { background: #0f1117; font-family: 'Segoe UI', system-ui, sans-serif; display: flex; align-items: center; justify-content: center; min-height: 100vh; }
+  .slide { width: 960px; padding: 56px 72px; }
+  h2 { color: #e2e8f0; font-size: 1.5rem; font-weight: 400; letter-spacing: 0.05em; text-transform: uppercase; margin-bottom: 48px; text-align: center; }
+
+  .top { display: flex; justify-content: center; margin-bottom: 0; }
+  .origin { background: #1a2040; border: 1.5px solid #4a5568; border-radius: 10px; padding: 24px 36px; text-align: center; min-width: 320px; }
+  .origin-label { font-size: 0.65rem; letter-spacing: 0.12em; text-transform: uppercase; color: #6b7280; margin-bottom: 8px; }
+  .origin-name { font-size: 1.1rem; font-weight: 600; color: #e2e8f0; margin-bottom: 6px; }
+  .origin-sub { font-size: 0.82rem; color: #6b7280; }
+
+  .discovery { margin: 28px 0; background: #1a1f2e; border: 1px dashed #7c3aed; border-radius: 8px; padding: 16px 24px; display: flex; align-items: flex-start; gap: 14px; }
+  .disc-icon { font-size: 1.2rem; flex-shrink: 0; margin-top: 2px; }
+  .disc-text { font-size: 0.85rem; color: #94a3b8; line-height: 1.6; }
+  .disc-text strong { color: #a78bfa; }
+  .disc-text code { color: #c4b5fd; background: #1e1435; padding: 1px 5px; border-radius: 3px; font-family: 'Courier New', monospace; font-size: 0.8rem; }
+
+  .branches { display: flex; gap: 32px; margin-top: 8px; position: relative; }
+
+  .branch { flex: 1; }
+  .branch-line { display: flex; justify-content: center; margin-bottom: 0; }
+  .branch-line svg { display: block; }
+
+  .gate-card { border-radius: 10px; padding: 26px 24px; }
+  .gate-card.resolve { background: #0c1f3a; border: 1.5px solid #3b82f6; }
+  .gate-card.architect { background: #1a0d2e; border: 1.5px solid #8b5cf6; }
+
+  .gate-label { font-size: 0.65rem; letter-spacing: 0.12em; text-transform: uppercase; margin-bottom: 8px; font-weight: 600; }
+  .resolve .gate-label { color: #60a5fa; }
+  .architect .gate-label { color: #a78bfa; }
+
+  .gate-name { font-size: 1rem; font-weight: 600; color: #e2e8f0; margin-bottom: 14px; }
+  .gate-q { font-size: 0.83rem; color: #94a3b8; font-style: italic; margin-bottom: 16px; line-height: 1.5; }
+
+  .stat-row { display: flex; gap: 10px; margin-bottom: 12px; flex-wrap: wrap; }
+  .chip { font-size: 0.75rem; padding: 4px 12px; border-radius: 20px; font-weight: 500; }
+  .chip-blue { background: #1e3a5f; color: #60a5fa; }
+  .chip-purple { background: #2e1065; color: #a78bfa; }
+  .chip-green { background: #052e16; color: #4ade80; }
+  .chip-amber { background: #2d1a00; color: #fbbf24; }
+
+  .acc-bar-wrap { margin-top: 14px; }
+  .acc-label { font-size: 0.72rem; color: #6b7280; margin-bottom: 6px; display: flex; justify-content: space-between; }
+  .acc-label span { font-weight: 600; }
+  .resolve .acc-label span { color: #60a5fa; }
+  .architect .acc-label span { color: #a78bfa; }
+  .acc-track { background: #1e2433; border-radius: 4px; height: 10px; overflow: hidden; }
+  .acc-fill { height: 100%; border-radius: 4px; }
+  .resolve .acc-fill { background: linear-gradient(90deg, #1d4ed8, #3b82f6); width: 69.4%; }
+  .architect .acc-fill { background: linear-gradient(90deg, #5b21b6, #8b5cf6); width: 75%; }
+
+  .divider { width: 1px; background: #2d3748; flex-shrink: 0; margin: 20px 0; }
+
+  .bottom-note { margin-top: 32px; display: flex; gap: 16px; }
+  .bn { flex: 1; background: #161b27; border-radius: 8px; padding: 16px 18px; border-left: 3px solid #4a5568; }
+  .bn-label { font-size: 0.65rem; letter-spacing: 0.1em; text-transform: uppercase; color: #6b7280; margin-bottom: 6px; }
+  .bn-text { font-size: 0.82rem; color: #94a3b8; line-height: 1.5; }
+</style>
+</head>
+<body>
+<div class="slide">
+  <h2>prereq_op_selector — Gate Decomposition</h2>
+
+  <div class="top">
+    <div class="origin">
+      <div class="origin-label">Original gate</div>
+      <div class="origin-name">prereq_op_selector</div>
+      <div class="origin-sub">Which prerequisite operations does this taskitem need?</div>
+    </div>
+  </div>
+
+  <div class="discovery">
+    <div class="disc-icon">🔍</div>
+    <div class="disc-text">
+      Analysis of 1297 accepted rows revealed <code>needs_validate_intake</code> is <strong>always True</strong> (1297/1297).
+      It is a constant, not a decision surface — no model can learn from a label with zero variance.
+      The effective gate is <strong>two independent binary classifiers</strong>.
+    </div>
+  </div>
+
+  <div class="branches">
+    <div class="branch">
+      <div class="gate-card resolve">
+        <div class="gate-label">Gate A — prereq_resolve</div>
+        <div class="gate-name">needs_resolve_dependencies</div>
+        <div class="gate-q">"Does this task require resolving cross-task dependencies before execution?"</div>
+        <div class="stat-row">
+          <span class="chip chip-blue">997 / 1297 positive</span>
+          <span class="chip chip-blue">76.9% base rate</span>
+        </div>
+        <div class="acc-bar-wrap">
+          <div class="acc-label">Baseline accuracy @ 4000 steps <span>69.4%</span></div>
+          <div class="acc-track"><div class="acc-fill"></div></div>
+        </div>
+        <div style="margin-top:12px; font-size:0.78rem; color:#4a5568;">
+          Note: high base rate — a majority classifier would score ~77%.<br>Baseline is learning something, but corpus is weak.
+        </div>
+      </div>
+    </div>
+
+    <div class="divider"></div>
+
+    <div class="branch">
+      <div class="gate-card architect">
+        <div class="gate-label">Gate B — prereq_architect</div>
+        <div class="gate-name">needs_architect_review</div>
+        <div class="gate-q">"Does this task require architect review before proceeding?"</div>
+        <div class="stat-row">
+          <span class="chip chip-purple">257 / 1297 positive</span>
+          <span class="chip chip-purple">19.8% base rate</span>
+        </div>
+        <div class="acc-bar-wrap">
+          <div class="acc-label">Baseline accuracy @ 4000 steps <span>75.0%</span></div>
+          <div class="acc-track"><div class="acc-fill"></div></div>
+        </div>
+        <div style="margin-top:12px; font-size:0.78rem; color:#4a5568;">
+          Note: a majority classifier here scores ~80%.<br>Baseline is learning from the minority class.
+        </div>
+      </div>
+    </div>
+  </div>
+
+  <div class="bottom-note">
+    <div class="bn">
+      <div class="bn-label">Why two separate models</div>
+      <div class="bn-text">Each gate fires independently. A task can need both, either, or neither. Coupling them into one multiclass model creates false label structure.</div>
+    </div>
+    <div class="bn">
+      <div class="bn-label">What the corpus lacks</div>
+      <div class="bn-text">Only ~17 unique text templates across 1297 rows. The taskitem schema was not designed to carry discriminative RSA fields. Better signal requires schema changes upstream.</div>
+    </div>
+    <div class="bn">
+      <div class="bn-label">Next step</div>
+      <div class="bn-text">Pipeline decision audit (SPRINT-004): walk the full whetstone_DSL pipeline and identify every decision point before collecting more training data.</div>
+    </div>
+  </div>
+
+</div>
+</body>
+</html>

presentations/visual_model_tiers.html

@@ -0,0 +1,135 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="UTF-8">
+<title>Model Tiers</title>
+<style>
+  * { box-sizing: border-box; margin: 0; padding: 0; }
+  body { background: #0f1117; font-family: 'Segoe UI', system-ui, sans-serif; display: flex; align-items: center; justify-content: center; min-height: 100vh; }
+  .slide { width: 960px; padding: 56px 72px; }
+  h2 { color: #e2e8f0; font-size: 1.5rem; font-weight: 400; letter-spacing: 0.05em; text-transform: uppercase; margin-bottom: 8px; text-align: center; }
+  .subtitle { text-align: center; color: #6b7280; font-size: 0.88rem; margin-bottom: 48px; }
+
+  .tiers { display: flex; align-items: flex-end; gap: 20px; margin-bottom: 40px; }
+  .tier { flex: 1; border-radius: 10px 10px 0 0; display: flex; flex-direction: column; justify-content: flex-end; position: relative; }
+
+  .tier-inner { padding: 20px 16px 24px; border-radius: 10px; border: 1.5px solid; }
+  .tier.det .tier-inner   { background: #0d1f16; border-color: #22c55e; }
+  .tier.tiny .tier-inner  { background: #0c1a30; border-color: #3b82f6; }
+  .tier.med .tier-inner   { background: #12122a; border-color: #8b5cf6; }
+  .tier.large .tier-inner { background: #2a1100; border-color: #f59e0b; }
+  .tier.llm .tier-inner   { background: #2a0f0f; border-color: #ef4444; }
+
+  .tier-badge { font-size: 0.62rem; letter-spacing: 0.14em; text-transform: uppercase; font-weight: 700; margin-bottom: 8px; }
+  .det   .tier-badge { color: #4ade80; }
+  .tiny  .tier-badge { color: #60a5fa; }
+  .med   .tier-badge { color: #a78bfa; }
+  .large .tier-badge { color: #fbbf24; }
+  .llm   .tier-badge { color: #f87171; }
+
+  .tier-name { font-size: 0.95rem; font-weight: 600; color: #e2e8f0; margin-bottom: 10px; line-height: 1.3; }
+  .tier-params { font-size: 0.8rem; margin-bottom: 8px; font-weight: 600; }
+  .det   .tier-params { color: #4ade80; }
+  .tiny  .tier-params { color: #60a5fa; }
+  .med   .tier-params { color: #a78bfa; }
+  .large .tier-params { color: #fbbf24; }
+  .llm   .tier-params { color: #f87171; }
+
+  .tier-detail { font-size: 0.76rem; color: #6b7280; line-height: 1.5; }
+
+  .marker { position: absolute; top: -32px; left: 50%; transform: translateX(-50%); white-space: nowrap; font-size: 0.7rem; color: #4ade80; background: #052e16; border: 1px solid #166534; padding: 3px 10px; border-radius: 20px; }
+  .marker.current { color: #60a5fa; background: #0c1a30; border-color: #1e3a5f; }
+
+  .axis { display: flex; align-items: center; margin-bottom: 8px; }
+  .axis-line { flex: 1; height: 2px; background: linear-gradient(90deg, #22c55e, #3b82f6, #8b5cf6, #f59e0b, #ef4444); border-radius: 1px; }
+  .axis-label-left  { font-size: 0.7rem; color: #4ade80; margin-right: 12px; white-space: nowrap; }
+  .axis-label-right { font-size: 0.7rem; color: #ef4444; margin-left: 12px; white-space: nowrap; }
+
+  .axis-labels { display: flex; justify-content: space-between; margin-bottom: 32px; }
+  .axis-labels span { font-size: 0.7rem; color: #4a5568; }
+
+  .hypothesis { background: #1a1f2e; border: 1px solid #2d3748; border-radius: 10px; padding: 24px 28px; }
+  .hyp-title { font-size: 0.65rem; letter-spacing: 0.12em; text-transform: uppercase; color: #6b7280; margin-bottom: 12px; }
+  .hyp-rule { font-size: 0.88rem; color: #94a3b8; line-height: 1.8; font-family: 'Courier New', monospace; }
+  .hyp-rule .dim { color: #4a5568; }
+  .hyp-rule .hi  { color: #22d3ee; font-weight: 600; }
+  .hyp-rule .note { color: #6b7280; font-family: 'Segoe UI', sans-serif; font-size: 0.8rem; margin-top: 10px; display: block; font-style: italic; }
+</style>
+</head>
+<body>
+<div class="slide">
+  <h2>Entropy → Model Tier</h2>
+  <div class="subtitle">The smallest model that clears the quality target for a given gate's entropy budget</div>
+
+  <div class="tiers">
+    <div class="tier det" style="position:relative;">
+      <div class="marker">← already valid</div>
+      <div class="tier-inner">
+        <div class="tier-badge">Tier 0</div>
+        <div class="tier-name">Deterministic Rules</div>
+        <div class="tier-params">0 params</div>
+        <div class="tier-detail">Perfect when input–output mapping is explicit and stable. No training cost. Fully auditable.</div>
+      </div>
+    </div>
+    <div class="tier tiny" style="position:relative;">
+      <div class="marker current">← trained baseline</div>
+      <div class="tier-inner">
+        <div class="tier-badge">Tier 1 · current</div>
+        <div class="tier-name">Tiny Specialist (Fabricate)</div>
+        <div class="tier-params">~213K params · ~800KB</div>
+        <div class="tier-detail">Binary/multiclass classifiers. Local, fast, inspectable. Target for bounded gates with sufficient entropy.</div>
+      </div>
+    </div>
+    <div class="tier med">
+      <div class="tier-inner">
+        <div class="tier-badge">Tier 2 · planned</div>
+        <div class="tier-name">Medium Specialist</div>
+        <div class="tier-params">~10× tier 1 · ~2–3M params</div>
+        <div class="tier-detail">For gates where tiny tier misses due to higher label boundary fuzziness or longer context dependence.</div>
+      </div>
+    </div>
+    <div class="tier large">
+      <div class="tier-inner">
+        <div class="tier-badge">Tier 3 · planned</div>
+        <div class="tier-name">Large Specialist</div>
+        <div class="tier-params">~16× tier 1 · ~3–4M params</div>
+        <div class="tier-detail">For gates with high lexical variation, world-knowledge dependence, or complex slot interdependence.</div>
+      </div>
+    </div>
+    <div class="tier llm">
+      <div class="tier-inner">
+        <div class="tier-badge">Tier 4 · escalation</div>
+        <div class="tier-name">SLM / LLM</div>
+        <div class="tier-params">billions of params</div>
+        <div class="tier-detail">Only when the gate is genuinely open-ended. Not a default. Escalation when specialists abstain.</div>
+      </div>
+    </div>
+  </div>
+
+  <div class="axis">
+    <span class="axis-label-left">low entropy</span>
+    <div class="axis-line"></div>
+    <span class="axis-label-right">high entropy</span>
+  </div>
+  <div class="axis-labels">
+    <span>simple lookup</span>
+    <span>binary classification</span>
+    <span>slot filling</span>
+    <span>complex routing</span>
+    <span>open reasoning</span>
+  </div>
+
+  <div class="hypothesis">
+    <div class="hyp-title">Current sizing hypothesis</div>
+    <div class="hyp-rule">
+      <span class="dim">if</span> entropy <span class="hi">&lt;= A</span>  <span class="dim">→</span>  deterministic<br>
+      <span class="dim">if</span> entropy <span class="hi">&lt;= B</span>  <span class="dim">→</span>  tiny specialist   (~213K params)<br>
+      <span class="dim">if</span> entropy <span class="hi">&lt;= C</span>  <span class="dim">→</span>  medium specialist (~2–3M params)<br>
+      <span class="dim">if</span> entropy <span class="hi">&gt;  C</span>  <span class="dim">→</span>  large specialist or escalate to SLM/LLM
+      <span class="note">Before moving up a tier: check schema, label set, missing context, and factorization options first.</span>
+    </div>
+  </div>
+
+</div>
+</body>
+</html>

presentations/visual_rsa_pipeline.html

@@ -0,0 +1,100 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="UTF-8">
+<title>RSA Pipeline</title>
+<style>
+  * { box-sizing: border-box; margin: 0; padding: 0; }
+  body { background: #0f1117; font-family: 'Segoe UI', system-ui, sans-serif; display: flex; align-items: center; justify-content: center; min-height: 100vh; }
+  .slide { width: 960px; padding: 60px 80px; }
+  h2 { color: #e2e8f0; font-size: 1.5rem; font-weight: 400; letter-spacing: 0.05em; text-transform: uppercase; margin-bottom: 56px; text-align: center; }
+  .pipeline { display: flex; align-items: center; gap: 0; }
+  .box { flex: 1; border-radius: 10px; padding: 28px 18px; text-align: center; }
+  .box.fuzzy  { background: #1e2433; border: 1.5px solid #4a5568; }
+  .box.gate   { background: #1a2744; border: 1.5px solid #3b82f6; }
+  .box.contract { background: #12293a; border: 1.5px solid #06b6d4; }
+  .box.det    { background: #12291e; border: 1.5px solid #22c55e; }
+  .box-label { font-size: 0.7rem; letter-spacing: 0.12em; text-transform: uppercase; margin-bottom: 10px; font-weight: 600; }
+  .fuzzy  .box-label { color: #9ca3af; }
+  .gate   .box-label { color: #60a5fa; }
+  .contract .box-label { color: #22d3ee; }
+  .det    .box-label { color: #4ade80; }
+  .box-main { font-size: 1.05rem; color: #e2e8f0; font-weight: 500; line-height: 1.4; min-height: 48px; display: flex; align-items: center; justify-content: center; }
+  .box-sub { font-size: 0.78rem; color: #6b7280; margin-top: 10px; line-height: 1.5; }
+  .arrow { flex: 0 0 40px; display: flex; align-items: center; justify-content: center; }
+  .arrow svg { display: block; }
+  .examples { margin-top: 52px; display: flex; gap: 24px; }
+  .ex { flex: 1; background: #161b27; border-radius: 8px; padding: 20px 22px; }
+  .ex-label { font-size: 0.65rem; letter-spacing: 0.1em; text-transform: uppercase; color: #6b7280; margin-bottom: 10px; }
+  .ex-text { font-size: 0.88rem; color: #94a3b8; line-height: 1.6; font-family: 'Courier New', monospace; }
+  .ex-text .hl { color: #22d3ee; }
+</style>
+</head>
+<body>
+<div class="slide">
+  <h2>The RSA Pattern</h2>
+  <div class="pipeline">
+    <div class="box fuzzy">
+      <div class="box-label">Input</div>
+      <div class="box-main">Fuzzy natural language</div>
+      <div class="box-sub">Varied phrasing<br>Partial information<br>Ambiguous intent</div>
+    </div>
+    <div class="arrow">
+      <svg width="40" height="24"><path d="M4 12 L30 12 M22 5 L30 12 L22 19" stroke="#4a5568" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round"/></svg>
+    </div>
+    <div class="box gate">
+      <div class="box-label">RSA Gate</div>
+      <div class="box-main">Tiny transformer specialist</div>
+      <div class="box-sub">Bounded output space<br>Confidence + abstain<br>~213K params / ~800KB</div>
+    </div>
+    <div class="arrow">
+      <svg width="40" height="24"><path d="M4 12 L30 12 M22 5 L30 12 L22 19" stroke="#4a5568" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round"/></svg>
+    </div>
+    <div class="box contract">
+      <div class="box-label">Contract</div>
+      <div class="box-main">Typed structured decision</div>
+      <div class="box-sub">action + slots<br>confidence score<br>escalation flag</div>
+    </div>
+    <div class="arrow">
+      <svg width="40" height="24"><path d="M4 12 L30 12 M22 5 L30 12 L22 19" stroke="#4a5568" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round"/></svg>
+    </div>
+    <div class="box det">
+      <div class="box-label">Execution</div>
+      <div class="box-main">Deterministic software</div>
+      <div class="box-sub">No model past this point<br>Ordinary code<br>Auditable + fast</div>
+    </div>
+  </div>
+  <div class="examples">
+    <div class="ex">
+      <div class="ex-label">Example — developer tools</div>
+      <div class="ex-text">
+        "show me the failed builds from yesterday"<br><br>
+        → <span class="hl">action:</span> search_builds<br>
+        → <span class="hl">status:</span> failed<br>
+        → <span class="hl">date:</span> yesterday<br>
+        → <span class="hl">confidence:</span> 0.93
+      </div>
+    </div>
+    <div class="ex">
+      <div class="ex-label">Example — whetstone_DSL gate</div>
+      <div class="ex-text">
+        taskitem: "Integrate auth middleware into API layer"<br><br>
+        → <span class="hl">needs_architect_review:</span> true<br>
+        → <span class="hl">needs_resolve_deps:</span> true<br>
+        → <span class="hl">confidence:</span> 0.87
+      </div>
+    </div>
+    <div class="ex">
+      <div class="ex-label">The design rule</div>
+      <div class="ex-text">
+        Learned behavior lives only<br>
+        where deterministic tooling<br>
+        is not yet enough.<br><br>
+        Everything after a bounded decision<br>
+        is ordinary software.
+      </div>
+    </div>
+  </div>
+</div>
+</body>
+</html>

presentations/visual_training_pipeline.html

@@ -0,0 +1,145 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="UTF-8">
+<title>Training Pipeline</title>
+<style>
+  * { box-sizing: border-box; margin: 0; padding: 0; }
+  body { background: #0f1117; font-family: 'Segoe UI', system-ui, sans-serif; display: flex; align-items: center; justify-content: center; min-height: 100vh; }
+  .slide { width: 1000px; padding: 56px 72px; }
+  h2 { color: #e2e8f0; font-size: 1.5rem; font-weight: 400; letter-spacing: 0.05em; text-transform: uppercase; margin-bottom: 52px; text-align: center; }
+
+  .pipeline { display: flex; flex-direction: column; gap: 0; }
+  .row { display: flex; align-items: stretch; gap: 0; }
+  .step { flex: 1; background: #161b27; border-radius: 10px; padding: 22px 20px; border: 1.5px solid #2d3748; }
+  .step-num { font-size: 0.65rem; letter-spacing: 0.14em; text-transform: uppercase; color: #4a5568; margin-bottom: 6px; }
+  .step-title { font-size: 1rem; font-weight: 600; color: #e2e8f0; margin-bottom: 8px; }
+  .step-detail { font-size: 0.8rem; color: #6b7280; line-height: 1.6; }
+  .step-detail code { color: #94a3b8; background: #1e2433; padding: 1px 5px; border-radius: 3px; font-family: 'Courier New', monospace; font-size: 0.78rem; }
+  .step-stat { display: inline-block; margin-top: 8px; font-size: 0.78rem; padding: 3px 10px; border-radius: 20px; font-weight: 600; }
+
+  .step.s1 { border-color: #4a5568; }
+  .step.s2 { border-color: #7c3aed; }
+  .step.s3 { border-color: #3b82f6; }
+  .step.s4 { border-color: #06b6d4; }
+  .step.s5 { border-color: #f59e0b; }
+  .step.s6 { border-color: #22c55e; }
+
+  .s1 .step-title { color: #9ca3af; }
+  .s2 .step-title { color: #a78bfa; }
+  .s3 .step-title { color: #60a5fa; }
+  .s4 .step-title { color: #22d3ee; }
+  .s5 .step-title { color: #fbbf24; }
+  .s6 .step-title { color: #4ade80; }
+
+  .stat-purple { background: #2e1065; color: #a78bfa; }
+  .stat-blue   { background: #1e3a5f; color: #60a5fa; }
+  .stat-cyan   { background: #0c2a3a; color: #22d3ee; }
+  .stat-amber  { background: #2d1a00; color: #fbbf24; }
+  .stat-green  { background: #052e16; color: #4ade80; }
+
+  .vgap { height: 12px; display: flex; justify-content: center; align-items: center; }
+  .vgap svg { display: block; }
+
+  .hgap { width: 12px; display: flex; align-items: center; justify-content: center; flex-shrink: 0; }
+  .hgap svg { display: block; }
+
+  .note { margin-top: 40px; background: #1a1f2e; border-left: 3px solid #f59e0b; border-radius: 0 8px 8px 0; padding: 16px 20px; }
+  .note-text { font-size: 0.83rem; color: #94a3b8; line-height: 1.6; }
+  .note-text strong { color: #fbbf24; }
+</style>
+</head>
+<body>
+<div class="slide">
+  <h2>From Run Artifacts to Trained Specialist</h2>
+  <div class="pipeline">
+
+    <div class="row">
+      <div class="step s1">
+        <div class="step-num">Step 1</div>
+        <div class="step-title">whetstone_DSL Run Corpus</div>
+        <div class="step-detail">
+          Real pipeline runs stored as JSON log artifacts.<br>
+          Each run records taskitem decisions, prerequisite ops, worker routing.
+          <span class="step-stat stat-purple">1000 runs extracted</span>
+        </div>
+      </div>
+      <div class="hgap">
+        <svg width="12" height="40"><path d="M6 4 L6 36 M2 28 L6 36 L10 28" stroke="#4a5568" stroke-width="1.5" fill="none" stroke-linecap="round" stroke-linejoin="round"/></svg>
+      </div>
+      <div class="step s2">
+        <div class="step-num">Step 2</div>
+        <div class="step-title">Gate Extraction</div>
+        <div class="step-detail">
+          <code>extract_gate_rows.py</code> parses each run.<br>
+          Isolates one gate's decision context and label per row.<br>
+          Rejects rows with schema drift or sibling-uniform labels.
+          <span class="step-stat stat-purple">1297 accepted / 24 schema-drift rejected</span>
+        </div>
+      </div>
+      <div class="hgap">
+        <svg width="12" height="40"><path d="M6 4 L6 36 M2 28 L6 36 L10 28" stroke="#4a5568" stroke-width="1.5" fill="none" stroke-linecap="round" stroke-linejoin="round"/></svg>
+      </div>
+      <div class="step s3">
+        <div class="step-num">Step 3</div>
+        <div class="step-title">Gate Decomposition</div>
+        <div class="step-detail">
+          Analysis revealed <code>needs_validate_intake</code> is always True — not a decision surface.<br>
+          Gate reduces to <strong style="color:#60a5fa">two independent binary classifiers</strong>.
+          <span class="step-stat stat-blue">2 gates from 1</span>
+        </div>
+      </div>
+    </div>
+
+    <div class="vgap">
+      <svg width="860" height="12"><path d="M430 2 L430 10 M426 6 L430 10 L434 6" stroke="#4a5568" stroke-width="1.5" fill="none" stroke-linecap="round" stroke-linejoin="round"/></svg>
+    </div>
+
+    <div class="row">
+      <div class="step s4">
+        <div class="step-num">Step 4</div>
+        <div class="step-title">TSV Generation</div>
+        <div class="step-detail">
+          <code>gen_prereq_op_rsa_data.py</code> converts rows to Fabricate TSV format.<br>
+          Format: <code>label &lt;TAB&gt; 0 &lt;TAB&gt; text</code><br>
+          Text composed from: title · reasons · acceptance criteria · constraints.
+          <span class="step-stat stat-cyan">85% train / 15% eval split · seed=42</span>
+        </div>
+      </div>
+      <div class="hgap">
+        <svg width="12" height="40"><path d="M6 4 L6 36 M2 28 L6 36 L10 28" stroke="#4a5568" stroke-width="1.5" fill="none" stroke-linecap="round" stroke-linejoin="round"/></svg>
+      </div>
+      <div class="step s5">
+        <div class="step-num">Step 5</div>
+        <div class="step-title">Fabricate Training</div>
+        <div class="step-detail">
+          Tiny transformer framework. GPU training on dedicated desktop.<br>
+          ~213K parameter model architecture.<br>
+          Binary classification head per gate.
+          <span class="step-stat stat-amber">4000 steps per specialist</span>
+        </div>
+      </div>
+      <div class="hgap">
+        <svg width="12" height="40"><path d="M6 4 L6 36 M2 28 L6 36 L10 28" stroke="#4a5568" stroke-width="1.5" fill="none" stroke-linecap="round" stroke-linejoin="round"/></svg>
+      </div>
+      <div class="step s6">
+        <div class="step-num">Step 6</div>
+        <div class="step-title">Specialist Checkpoint</div>
+        <div class="step-detail">
+          Serialized model: ~800KB on disk.<br>
+          Loaded at runtime by the RSA gate layer.<br>
+          Returns: label · confidence · abstain flag.
+          <span class="step-stat stat-green">~800KB per specialist</span>
+        </div>
+      </div>
+    </div>
+
+  </div>
+  <div class="note">
+    <div class="note-text">
+      <strong>Corpus caveat:</strong> The 1000-run dataset contains only ~17 unique text templates — the taskitem schema was designed to record outputs, not discriminative inputs. These baselines are intentionally weak. The pipeline is correct; the signal improves when the taskitem schema carries RSA-native fields.
+    </div>
+  </div>
+</div>
+</body>
+</html>