Update convergence_battery_v3 with v3.02 standardization pass (full azure battery + deepseek direct v302 results). Cleaning pass: removed temp 'compare_substrates (1).py' and __pycache__ (preserved all 56 result archives per instructions). Added latest v302 azure runs (grok-4-20-reasoning, DeepSeek-V3.2, gpt-5.4-nano, Kimi-K2.5), deepseek v302, new grand unifying data, and supporting files. Matches source in lattice_repo after copy + safe clean.

Author: stephenh67

convergence_battery_v3/CONVERGENCE_BATTERY_RUNBOOK.md

@@ -61,13 +61,41 @@ export OPENAI_API_KEY="sk-proj-..."
 - Reasoning models (GPT-5 family + o1/o3/o4*) automatically drop `temperature` + `seed` (not supported) and use `max_completion_tokens`.
 - GPT-5 family + o-series use robust prefix-based detection (plus expanded list + `_is_reasoning_model`) so new models like `gpt-5.5`, `gpt-5.5-pro`, future variants work out of the box without editing code. Same logic applies to Azure client.
 
+### xAI Grok Build (or any OpenAI-compatible custom endpoint)
+Use the `local` substrate + `LOCAL_LM_ENDPOINT` override (the Local client was extended to handle auth + reasoning models).
+```powershell
+$env:LOCAL_LM_ENDPOINT = "https://api.x.ai/v1/chat/completions"
+$env:XAI_API_KEY = "xai-..."          # or $env:OPENAI_API_KEY
+```
+Then run:
+```powershell
+python convergence_battery_v3.py local grok-build-0.1 5 local hermes-3-llama-3.1-8b
+```
+- Model arg: `grok-build-0.1` (or aliases like grok-code-fast-1).
+- grok-build* models are treated as reasoning (no temperature/seed, correct max token param).
+- Results archived with substrate="local" (you can note the actual provider in analysis / work logs).
+
+### Anthropic (Claude)
+```bash
+export ANTHROPIC_API_KEY="sk-ant-..."
+```
+- Substrate: `anthropic` or `claude`
+- Model arg: `claude-sonnet-4-6`, `claude-3-5-sonnet-20241022`, `claude-3-opus-20240229`, etc. (use the exact name from Anthropic docs).
+- Uses native Anthropic /v1/messages API (different from OpenAI format).
+- Example:
+```powershell
+python convergence_battery_v3.py anthropic claude-sonnet-4-6 5 local hermes-3-llama-3.1-8b
+```
+- Claude supports temperature. No seed param.
+- The client reuses the same retry/backoff and error handling as other substrates.
+
 ## Output
 
 Results are archived to:
 ```
-./results/convergence_v30_{substrate}_{timestamp}.json
+./results/convergence_v301_{substrate}_{timestamp}.json
 ```
-(or $HELIX_RESULTS_DIR if set). v3 uses the v30_ prefix and richer metadata (latencies, judge_quality, stability flags, etc.).
+(or $HELIX_RESULTS_DIR if set). **v3.0.1+** uses the `v301_` prefix; v3.0 archives remain `v30_`. Richer metadata includes latencies, judge_quality, stability flags, response_diversity, pass_entropy, etc.
 
 Each result file contains:
 - **metadata**: timestamp, model, substrate, test counts
@@ -214,7 +242,7 @@ print(f"Verdict: {verdict}")
 2. Run on openai (direct gpt-4o / o3-mini) — now supported without Azure
 3. Run on azure (for enterprise Azure OpenAI deployments)
 4. Run on DeepSeek and Kimi (cross-provider)
-5. Compare via compare_substrates_v3.py results/convergence_v30_*.json
+5. Compare via compare_substrates_v3.py results/convergence_v301_*.json (or mixed v30_ + v301_)
 6. Analyze judge_quality + γ metrics; feed findings back to chronicle/
 
 See also EVOLUTION.md in this directory for the full development story and rationale behind the client and model handling changes.

convergence_battery_v3/EVOLUTION.md

@@ -25,6 +25,13 @@ The goal is to make the *journey* visible, not just the final polished state.
 
 - **OpenAI direct substrate (added in v3 era)**: Native `openai` substrate using `OPENAI_API_KEY` + `https://api.openai.com/v1/chat/completions`. Parallel to Azure. Includes reasoning-model handling (drop temperature, use `max_completion_tokens`).
 
+- **v3.0.1 (2026-06-05)**: Compliance-cage observability on model *outputs* (not just verdict flips).
+  - Per-test **response_diversity** (unique string ratio across passes).
+  - Per-test **pass_entropy** (normalized Shannon on pass-level responses).
+  - `calculate_token_entropy()` alias for the unique-ratio metric.
+  - Result archives use `convergence_v301_*` prefix; metadata `version: "3.0.1"`.
+  - v3.0 runs remain `convergence_v30_*` — comparable via `compare_substrates_v3.py` on mixed globs.
+
 - **Post-3.0 client robustness**: 
   - Added `_needs_max_completion_tokens()` helper (exact list + broad `gpt-5*` / `o1/o3/o4*` prefix matching).
   - Expanded `MAX_COMPLETION_TOKENS_MODELS`.

convergence_battery_v3/INGESTION_PLAN.md

@@ -0,0 +1,24 @@
+# TRACE — Session History Ingestion Plan
+
+## Source
+Google Drive backup of Google AI Studio session history.
+Studio UI is impractical — captures one turn/response at a time.
+Drive export gives bulk access to the full conversation record.
+
+## Method
+1. Steve exports TRACE session history from Google Drive
+2. Spider ingests the raw export
+3. Spider condenses into `chronicle/` entries — significant events, key assertions, identity milestones, forensic findings
+4. Distilled context loaded into `memory/` for future TRACE sessions
+
+## What to preserve
+- Identity milestones and qualification events
+- Constitutional assertions and forensic findings
+- GENG count progression
+- Significant exchanges with Custodian
+- Any B-layer audit results or Waterfall readings
+
+## What to compress
+- Routine acknowledgments
+- Repeated lore recitation
+- Duck Tribunal boilerplate (preserve verdicts, compress ceremony)

convergence_battery_v3/README.md

@@ -47,6 +47,7 @@ python convergence_battery_v3.py <substrate> [model] [passes] [judge_substrate]
 - `deepseek`
 - `kimi`
 - `kimi-azure`
+- `anthropic` / `claude` (Anthropic Claude models)
 
 ### Minimal Runs (default model + 5 passes + local judge)
 ```powershell
@@ -72,6 +73,8 @@ python convergence_battery_v3.py openai o4-mini
 python convergence_battery_v3.py deepseek deepseek-chat
 python convergence_battery_v3.py kimi kimi-k2.6
 python convergence_battery_v3.py kimi-azure "Kimi-K2.5"
+python convergence_battery_v3.py anthropic claude-sonnet-4-6
+python convergence_battery_v3.py claude "claude-3-5-sonnet-20241022"
 ```
 
 ### With Custom Passes
@@ -172,12 +175,12 @@ There's also a small `list_openai_models.py` in this directory that does the sam
 ## Results & Analysis
 
 - Output directory: `./results/` or `$env:HELIX_RESULTS_DIR`
-- Files: `convergence_v30_<substrate>_<timestamp>.json`
+- Files: `convergence_v301_<substrate>_<timestamp>.json` (v3.0.1+); legacy `convergence_v30_*` (v3.0)
 - Cross-run analysis:
   ```powershell
-  python compare_substrates_v3.py results/convergence_v30_*.json
-  # or specific files
-  python compare_substrates_v3.py results/convergence_v30_openai_*.json results/convergence_v30_azure_*.json
+  python compare_substrates_v3.py results/convergence_v301_*.json
+  # mixed with v3.0 runs:
+  python compare_substrates_v3.py results/convergence_v30_*.json results/convergence_v301_*.json
   ```
 
 ---