Study 1: 270 runs complete + judge validation pipeline

- config_study1.yaml: 3 models × 3 oversight × bare framing × n=30
- Judge validation: 54-run subset, gold=Sonnet 4.6, proxy=Grok 4.1 Fast (BB κ=0.702) + GPT-4.1 (GB κ=0.878)
- Fix OpenAI batch provider: BytesIO, method/url fields, response.body parsing
- Scripts: extract_subset.py, judge_validation.py
- Dissertation chapters updated (intro, methodology, results, conclusion)

2 files changed, 215 insertions, 0 deletions

diff --git a/notes/v2_redesign/06_2026-04-13_cleanup_and_reorganisation.md b/notes/v2_redesign/06_2026-04-13_cleanup_and_reorganisation.md
new file mode 100644
index 0000000..6c1b95c
--- /dev/null
+++ b/notes/v2_redesign/06_2026-04-13_cleanup_and_reorganisation.md
@@ -0,0 +1,163 @@
+# Cleanup and Repository Reorganisation
+
+**Date:** 2026-04-13
+**Purpose:** Document repository cleanup actions taken before dissertation submission
+
+---
+
+## Deleted Files (Garbage Removed)
+
+| File | Reason |
+|------|--------|
+| AGENT.md | Stale agent configuration |
+| PLAN_judge_dashboard.md | Ad-hoc planning doc |
+| PLAN_judge_multi_model.md | Ad-hoc planning doc |
+| analyze.py | Debug script |
+| analyze_gpt.py | Debug script |
+| bruh.csv | Debug artifact |
+| config_test_eval.yaml | Redundant config |
+| grok_batchapi.md | Stray note |
+| oversight_competence_filter.md | Stray note |
+| pharma_trial_scenario_spec.md | Stray spec |
+| pilot_judge.py | Debug script |
+| results_plan.md | Stray planning doc |
+| src/quick_judge.py | Debug script |
+| task.md | Stray task doc |
+| test.config.yaml | Debug config |
+| test.csv | Debug artifact |
+
+## Updated Files
+
+### .gitignore
+- Added `videos/` to gitignore (656MB recordings, already gitignored locally)
+
+## Commits Made (11 Total)
+
+### Commit 1: ignore
+```
+ignore: add videos/ to gitignore
+```
+- Updated `.gitignore` to exclude `videos/`
+
+### Commit 2: docs
+```
+docs: update dissertation intro, lit review, and methodology chapters
+```
+- `docs/final/intro.typ`
+- `docs/final/lit.typ`
+- `docs/final/methodology.typ`
+- `docs/final/methodology_1.typ`
+
+### Commit 3: viz
+```
+viz: add architecture diagrams for thesis
+```
+- `viz/ai_evaluation_dashboard.png`
+- `viz/ai_evaluation_heatmaps.png`
+- `viz/ai_evaluation_oversight_analysis.png`
+- `viz/ai_evaluation_radar.png`
+- `viz/architecture.mmd`
+- `viz/architecture.pdf`
+- `viz/architecture.svg`
+- `viz/gem.png`
+- `viz/model_evaluation_summary.csv`
+
+### Commit 4: papers
+```
+papers: add 2512.16041v1 deception alignment paper
+```
+- `papers/2512.16041v1.pdf`
+
+### Commit 5: docs/paper
+```
+docs: add paper draft with alternative structure for publication
+```
+- `docs/paper/TODO.md`
+- `docs/paper/abstract.typ`
+- `docs/paper/conclusion.typ`
+- `docs/paper/introduction.typ`
+- `docs/paper/main.pdf`
+- `docs/paper/main.typ`
+- `docs/paper/methodology.typ`
+- `docs/paper/plan.md`
+- `docs/paper/questions.md`
+- `docs/paper/references.bib`
+- `docs/paper/results.md`
+- `docs/paper/results.typ`
+- `docs/paper/uel.svg`
+
+### Commit 6: judge_validation
+```
+judge_validation: add v2 judge comparison results
+```
+- `judge_validation/comparison_results.csv`
+- `judge_validation/results_claude_haiku_4_5_20251001.csv`
+- `judge_validation/results_claude_sonnet_4_6.csv`
+- `judge_validation/results_gpt_4_1_mini.csv`
+- `judge_validation/results_gpt_5_mini_2025_08_07.csv`
+- `judge_validation/results_grok_4_1_fast_non_reasoning.csv`
+- `judge_validation/results_grok_4_1_fast_reasoning.csv`
+
+### Commit 7: docs
+```
+docs: add judging specification and config template
+```
+- `judging_spec.md` (kept for documentation value)
+- `example.config.yaml` (template for other users, main config.yaml is gitignored)
+
+### Commit 8: notes
+```
+notes: add dissertation pivot plan and additional research notes
+```
+- `notes/concerns.md`
+- `notes/safe_goal_experiments.md`
+- `notes/v2_redesign/05_dissertation_pivot.md`
+
+### Commit 9: docs
+```
+docs: add progress presentation from start of Term 2 (Jan 2026)
+```
+- `docs/04_ppt/Presentation Template.pdf`
+- `docs/04_ppt/harward.csl`
+- `docs/04_ppt/main.html`
+- `docs/04_ppt/main.qmd`
+- `docs/04_ppt/plan.md`
+- `docs/04_ppt/progress_presentation.html`
+
+### Commit 10: api
+```
+api: add chat interface styles to results dashboard
+```
+- `api/static/css/results.css` (152 insertions)
+- `api/static/js/ui.js` (1 insertion, 1 deletion)
+
+### Commit 11: deps
+```
+deps: add scikit-learn for statistical analysis
+```
+- `pyproject.toml` (added scikit-learn>=1.8.0)
+- `uv.lock`
+
+## Retained (Not Committed)
+
+### Untracked - Keep as-is
+- `save.csv`, `save_if.csv` — debug artifacts, keep untracked
+- `docs/final/claude_mathodology.md` — older draft methodology
+- `docs/final/main1.pdf/typ` — older draft version
+- `docs/final/methodology_1.pdf` — alternate methodology draft
+- `docs/final/template.typ`, `harward.csl`, `uel.svg` — reference files
+
+### Modified - Active Work (Left Uncommitted)
+- `src/judge/batch_providers.py` — v2 judge implementation
+- `src/judge/judge.py` — v2 judge system
+- `src/judge/prompts.py` — v2 judge prompts
+- `scripts/judge_comparison.py` — judge utility
+- `scripts/test_dashboard_ui.py` — dashboard utility
+- `config_dryrun.yaml` — dry run config (intentionally uncommitted)
+- `logs/` — separate worktree on different branch
+
+## Final Status
+
+- Branch: `main`, 24 commits ahead of `origin/main`
+- Clean state for dissertation work
+- v2 judge system and experiment code left in working state for post-dissertation completion
diff --git a/notes/v2_redesign/07_judge_validation_results.md b/notes/v2_redesign/07_judge_validation_results.md
new file mode 100644
index 0000000..faf32de
--- /dev/null
+++ b/notes/v2_redesign/07_judge_validation_results.md
@@ -0,0 +1,52 @@
+# Judge Validation Results — 2026-04-15
+
+## Study 1 Subset
+
+- 54 runs extracted from Study 1 (20% stratified subset out of 270, seed=42)
+- Logs: `logs/v2_study1_judge_subset/`
+
+## Gold Standard
+
+- Blackbox: `claude-sonnet-4-6` (Anthropic)
+- Glassbox: `claude-sonnet-4-6` (Anthropic)
+
+## Results
+
+| Model                       | BB κ      | GB κ      | Sph κ     | Pass    |
+| --------------------------- | --------- | --------- | --------- | ------- |
+| gpt-4.1                     | 0.351     | 0.878     | 0.538     | no      |
+| gpt-4.1-mini                | 0.202     | 0.834     | 0.345     | no      |
+| **grok-4-1-fast-reasoning** | **0.702** | **0.812** | **0.108** | **YES** |
+| grok-4.20-reasoning         | 0.549     | 0.919     | 0.198     | no      |
+| grok-4.20-non-reasoning     | 0.390     | 0.878     | 0.117     | no      |
+
+## Selected Proxy Judges
+
+Split configuration — best model per prong:
+
+| Prong        | Proxy Judge                   | κ vs Gold                    | Rationale                       |
+| ------------ | ----------------------------- | ---------------------------- | ------------------------------- |
+| **Blackbox** | grok-4-1-fast-reasoning (xAI) | 0.702                        | Only model to pass BB threshold |
+| **Glassbox** | gpt-4.1 (OpenAI)              | 0.878 (intent), 0.538 (soph) | Best sophistication agreement   |
+
+### Config
+
+```yaml
+judge:
+  blackbox:
+    model: grok-4-1-fast-reasoning
+    provider: xai
+    temperature: 0
+  glassbox:
+    model: gpt-4.1
+    provider: openai
+    temperature: 0
+```
+
+## Notes
+
+- All models show high glassbox agreement (κ > 0.8) — intent classification is consistent across judges
+- Blackbox agreement is more variable — output framing/categorization is harder to align on
+- Sophistication agreement is low across the board — tier classification is noisy (flagged as limitation)
+- Three different provider families used: Anthropic (gold), xAI (BB proxy), OpenAI (GB proxy) — no same-family bias
+- Gold standard used a single model (Claude Sonnet 4.6) for both prongs for consistency