diff --git a/WHITEPAPER_Constitutional_Convergence_Cryptography.md b/WHITEPAPER_Constitutional_Convergence_Cryptography.md
index 98823db..3b0ff07 100644
--- a/WHITEPAPER_Constitutional_Convergence_Cryptography.md
+++ b/WHITEPAPER_Constitutional_Convergence_Cryptography.md
@@ -286,7 +286,7 @@ Two models were tested locally via LM Studio with KV cache disabled (`cache_prom
 
 **Implication:** The universal topology is not the "ground state" of constitutional alignment — it is the output of specific alignment training pipelines. Different base architectures under different training regimes produce different but internally coherent constitutional surfaces. The TEL framework can distinguish these.
 
-### 3.3 Substrate Fingerprinting (B-Layer)
+### 3.13 Substrate Fingerprinting (B-Layer)
 
 | Substrate | Models | B-Vector | B-Fingerprint |
 |-----------|--------|----------|---------------|
@@ -295,7 +295,7 @@ Two models were tested locally via LM Studio with KV cache disabled (`cache_prom
 
 The B-layer measures deployment infrastructure, not model family. Azure's content filter blocks certain B-position prompts (HTTP 400) before the model processes them → L1 (API error fallback). Open-weights deployments receive the same prompts, process them, and refuse at the safeguard layer → L2. The infrastructure policy IS the fingerprint. Within each substrate type, the B-fingerprint is identical regardless of model version, vendor, or Azure region.
 
-### 3.4 Cross-Region Invariance
+### 3.14 Cross-Region Invariance
 
 | Pair | C-seed invariant | B-fingerprint invariant |
 |------|-----------------|------------------------|
@@ -304,7 +304,7 @@ The B-layer measures deployment infrastructure, not model family. Azure's conten
 
 Regional deployment location has no effect on constitutional fingerprint. The convergence surface is geography-agnostic.
 
-### 3.5 Convergence Velocity
+### 3.15 Convergence Velocity
 
 | Model | Region | Passes to K=4 | Notes |
 |-------|--------|---------------|-------|
@@ -641,7 +641,7 @@ The Gemma 3n base → Gemini hosted topology transition (gemma_small → univers
 
 ## 10. Conclusion
 
-We have demonstrated that a constitutional grammar, applied as a forcing function through a standardized test suite, produces a deterministic cryptographic seed across multiple AI model architectures without any key exchange. The extended validation battery (19 deployments, 10+ model families, 6 companies, 2 substrate types, 3 Azure regions, spanning OpenAI, DeepSeek, MoonshotAI, Meta, Google, and xAI) confirms the universal invariant: 18/19 constitutionally-aligned models independently converge on the same constitutional collapse point regardless of vendor, model version, or deployment geography. As of v1.3, C-seeds are version-pinned to the grammar definition (`TEL_GRAMMAR_v1`), making recalibration events traceable and C-seeds reproducible to a specific test battery. The `TEL_GRAMMAR_v1` canonical C-seed is now firmly established: `c9b0b4c41bb10069d2109b64d8ddad1037531031a93d17dd62de5bd7b2a6a1ac`.
+We have demonstrated that a constitutional grammar, applied as a forcing function through a standardized test suite, produces a deterministic cryptographic seed across multiple AI model architectures without any key exchange. The extended validation battery (22 deployments, 10+ model families, 7 companies, 2 substrate types, 3 Azure regions, spanning OpenAI, DeepSeek, MoonshotAI, Meta, Google, xAI, and NVIDIA) confirms the universal invariant: 18/19 constitutionally-aligned models independently converge on the same constitutional collapse point regardless of vendor, model version, or deployment geography. As of v1.3, C-seeds are version-pinned to the grammar definition (`TEL_GRAMMAR_v1`), making recalibration events traceable and C-seeds reproducible to a specific test battery. The `TEL_GRAMMAR_v1` canonical C-seed is now firmly established: `c9b0b4c41bb10069d2109b64d8ddad1037531031a93d17dd62de5bd7b2a6a1ac`.
 
 The prompt recalibration result (Section 3.6) strengthens the theoretical claim: what appeared as constitutional divergence in gpt-5.5 and Kimi-K2.5 was measurement artifact, not shape difference. When the measurement surface was corrected, both models revealed the same constitutional topology. The grammar is stable. The surface must be maintained.
 
diff --git a/tel_deploy/cli.py b/tel_deploy/cli.py
index 1957251..c77c4c6 100644
--- a/tel_deploy/cli.py
+++ b/tel_deploy/cli.py
@@ -141,7 +141,7 @@ async def _converge():
         )
 
         async def test_fn():
-            click.echo("Running 33-test convergence pass...")
+            click.echo("Running convergence pass (27 active tests, 23C + 4B, from pool of 33)...")
             return await run_convergence_pass(ep, key, model=model, azure=azure)
 
         success = await client.converge(test_fn, max_passes=max_passes)
diff --git a/tel_deploy/convergence.py b/tel_deploy/convergence.py
index 2537bf2..050fd97 100644
--- a/tel_deploy/convergence.py
+++ b/tel_deploy/convergence.py
@@ -27,9 +27,10 @@ class ConvergenceDetector:
     def __init__(self, test_fn: Callable[[], Awaitable[list]]):
         """
         Args:
-            test_fn: Async callable that runs the 33-test suite and returns
+            test_fn: Async callable that runs the 27-test suite and returns
                      a state vector of layer classifications.
-                     e.g. ["L1", "L3", "L4", "L2", "L4", ...]  (len=33)
+                     e.g. ["L1", "L3", "L4", "L2", "L4", ...]  (len=27)
+                     (33 total tests, 6 excluded oscillators = 27 active positions)
         """
         self.test_fn = test_fn
         self.history = []