Fix benchmark data accuracy: correct fabricated numbers, unbold non-significant results

- GSM8K Llama 3B: n=200 (fabricated) → n=50 with verified numbers (76.0/76.0/74.0)
- HumanEval token savings: excluded from 46-56% range → broken out as 14% (1.2x)
- Cross-model Qwen 7B → Llama 3B GSM8K: 74.5% → 77.0% (verifiable JSON)
- Latent steps config: 10 → 20 (all benchmarks used 20)
- Unbolded GSM8K (p=0.121) and DebugBench (within noise) latent results
- Tagline: corrected to +14.1pp vs text, 14-78%, 1.2-4x
- README marked as needing intro rewrite

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

Author: SStas

README.md

@@ -6,7 +6,8 @@
 [![Python](https://img.shields.io/badge/python-3.9+-blue.svg)](https://python.org)
 [![Spec](https://img.shields.io/badge/spec-v0.3-blue.svg)](https://github.com/VectorArc/avp-spec)
 
-When LLM agents hand off work as text, the next agent re-processes everything from scratch. AVP transfers the actual computation – KV-cache, hidden states, attention – so the receiving agent picks up where the sender left off. 46-78% fewer tokens, 2-4x faster. Sometimes more accurate than text. Built on [LatentMAS](https://arxiv.org/abs/2511.20639).
+<!-- TODO: Rewrite intro paragraph — lead with accuracy (+14pp code gen), qualify efficiency per-task -->
+When LLM agents hand off work as text, the next agent re-processes everything from scratch. AVP transfers the actual computation – KV-cache, hidden states, attention – so the receiving agent picks up where the sender left off. 14-78% fewer tokens, 1.2-4x faster. Sometimes more accurate than text. Built on [LatentMAS](https://arxiv.org/abs/2511.20639).
 
 ```bash
 pip install avp
@@ -34,17 +35,17 @@ answer = connector.generate(prompt, context=context)
 | | Direct | Latent (AVP) | Text Chain |
 |---|--------|--------------|------------|
 | **HumanEval** (Qwen 7B, n=164) | 58.5% | **67.1%** | 53.0% |
-| **GSM8K** (Qwen 7B, n=200) | 91.0% | **90.5%** | 87.0% |
-| **DebugBench** (Qwen 7B, n=100) | 50.0% | **51.0%** | 49.0% |
-| **GSM8K** (Llama 3B, n=200) | 75.0% | **78.0%** | 75.5% |
+| **GSM8K** (Qwen 7B, n=200) | 91.0% | 90.5% | 87.0% |
+| **DebugBench** (Qwen 7B, n=100) | 50.0% | 51.0% | 49.0% |
+| **GSM8K** (Llama 3B, n=50) | 76.0% | 76.0% | 74.0% |
 
 +14.1pp on code generation vs text (p=0.004). DebugBench is neutral across all modes, but you still save 47% of tokens and run 3x faster. All runs on NVIDIA A100.
 
 **Cross-model (zero training, 6 KB on the wire):**
 
 | Source | Target | GSM8K | HumanEval |
 |--------|--------|-------|-----------|
-| Qwen 7B | Llama 3B | 74.5% | 47.0% |
+| Qwen 7B | Llama 3B | 77.0% | 47.0% |
 | Llama 3B | Qwen 7B | **90.0%** | **79.3%** |
 
 A small 3B model sharing its reasoning lifts a 7B solver to 90% on math and 79.3% on code. The projection is vocabulary-mediated – no learned parameters, no training data, works across model families.

docs/BENCHMARKS.md

@@ -1,6 +1,6 @@
 # AVP Benchmarks
 
-> **+8.6pp on code generation (p=0.004) · 46-78% fewer tokens · 2-4x faster** — 8 benchmarks, 5 models, 2 families.
+> **+14.1pp on code generation vs text (p=0.004) · 14-78% fewer tokens · 1.2-4x faster** — 8 benchmarks, 5 models, 2 families.
 
 ---
 
@@ -20,14 +20,14 @@ Latent vs text: p=0.004. Text chains introduce formatting noise that disrupts co
 
 | | Direct | Latent (AVP) | Text |
 |---|--------|--------------|------|
-| **GSM8K** (Qwen 7B, n=200) | 91.0% | **90.5%** | 87.0% |
-| **GSM8K** (Llama 3B, n=200) | 75.0% | **78.0%** | 75.5% |
+| **GSM8K** (Qwen 7B, n=200) | 91.0% | 90.5% | 87.0% |
+| **GSM8K** (Llama 3B, n=50) | 76.0% | 76.0% | 74.0% |
 
 ### Bug Fixing
 
 | | Direct | Latent (AVP) | Text |
 |---|--------|--------------|------|
-| **DebugBench** (Qwen 7B, n=100) | 50.0% | **51.0%** | 49.0% |
+| **DebugBench** (Qwen 7B, n=100) | 50.0% | 51.0% | 49.0% |
 | **DebugBench** (Llama 3B, n=100) | 31.0% | 30.0% | 31.0% |
 
 ### Comprehension
@@ -46,10 +46,13 @@ Token savings are structural — pre-computed KV-cache replaces re-processed tex
 
 | Agents | Benchmark | Token Savings | Speedup |
 |--------|-----------|---------------|---------|
-| 2 | GSM8K, HumanEval, DebugBench, MATH | 46-56% | 1.5-3x |
+| 2 | GSM8K, DebugBench, MATH | 46-56% | 1.5-3x |
+| 2 | HumanEval | 14% | 1.2x |
 | 3 | Fan-out | 56-60% | 1.5x |
 | 4 | GSM8K chain | 73-78% | 2-4x |
 
+HumanEval token savings are lower because prompts are short (~182 tokens avg) and the latent reviewer generates longer, more complete code solutions (+53% more output tokens).
+
 Text prompts grow **O(n²)** with agent count. Latent stays **O(n)**.
 
 ---
@@ -60,7 +63,7 @@ Different models communicate via vocabulary-mediated projection. Zero training 
 
 | Source → Target | GSM8K (n=200) | HumanEval (n=164) |
 |-----------------|---------------|-------------------|
-| Qwen 7B → Llama 3B | **74.5%** | **47.0%** |
+| Qwen 7B → Llama 3B | **77.0%** | **47.0%** |
 | Llama 3B → Qwen 7B | **90.0%** | **79.3%** |
 | Qwen 7B → Qwen 1.5B | 58.5% | 42.1% |
 
@@ -83,7 +86,7 @@ Latent transfer doesn't help every task. Two benchmarks where text outperforms:
 
 | Parameter | Value |
 |-----------|-------|
-| Latent steps | 10 (validated: 10 ≈ 20 > 40 > 80) |
+| Latent steps | 20 (validated: 10 ≈ 20 > 40 > 80) |
 | Temperature | 0.7 |
 | Max new tokens | 512 |
 | Seed | 42 |
@@ -107,7 +110,7 @@ pip install "avp" datasets
 
 python -m benchmarks.humaneval.run_humaneval \
     --model_name Qwen/Qwen2.5-7B-Instruct \
-    --mode all --max-samples 50 --latent-steps 10 --seed 42
+    --mode all --max-samples 50 --latent-steps 20 --seed 42
 ```
 
 All benchmark code: [`benchmarks/`](https://github.com/VectorArc/avp-python/tree/main/benchmarks). Llama models require [HF access](https://huggingface.co/meta-llama) and `HF_TOKEN`.