added banner + updated README

.gitignore

@@ -0,0 +1,4 @@
+*.py
+logicmark_test.jsonl
+*.txt
+*.json

README.md

@@ -2,9 +2,13 @@
 license: apache-2.0
 ---
 
+![Axiomic Banner](AxiomicBanner.png)
+
 # LogicMark
 
-A procedurally generated benchmark for evaluating symbolic reasoning in language models. Each problem presents a set of variable equality/inequality premises and asks the model to identify which conclusion necessarily follows.
+A procedurally generated benchmark for evaluating symbolic logic in language models. Each problem presents a set of variable equality/inequality premises and asks the model to identify which conclusion necessarily follows.
+
+Unlike knowledge-based benchmarks, LogicMark contains no facts a model could have memorised from pretraining. Every problem is generated fresh from abstract variable names (`a`, `b`, `c`, ...), so a model cannot pattern-match to training data - it must actually reason. This makes LogicMark a direct probe of **intrinsic reasoning capability**: the logical structure that has been built into the model's weights through training, independent of world knowledge or surface-level heuristics.
 
 Evaluation is log-likelihood multiple-choice — no chain-of-thought, no prompting tricks. Models are scored purely on how well they assign probability to the correct completion.
 
@@ -152,14 +156,4 @@ Evaluated using average log-likelihood over ending tokens, normalised by length.
 | pythia-2.8B | 2.8B | 70.40% | 49.60% | 35.84% | 36.80% | 31.40% | 44.50% |
 | Qwen2.5-3B | 3.1B | 48.80% | 54.90% | 46.00% | 42.40% | 39.40% | 48.64% |
 
-### Key Findings
-
-**Qwen training recipe separates from all other models at 3+ hop.** At 2-hop most models cluster around 49-51%. By 3-hop, Qwen2.5 models sit at 43-47% while everything else falls to 34-39% — a gap that widens further at 4-5 hop.
-
-**Pythia scaling is flat.** pythia-2.8B (44.5%) barely outperforms pythia-14m (42.1%) despite 200x the parameters. The Pile training data has a hard ceiling on this benchmark regardless of scale.
-
-**GPT-X2 1-hop anomaly.** GPT-X2 at 125M scores 74.8% at 1-hop — matching LFM2-350M-Math and exceeding GPT-2-XL (60.4%) and SmolLM2-1.7B (65.0%). This is notable given GPT-X2's tokenizer splits `!=` into two tokens, putting it at a structural disadvantage for direct premise matching.
-
-**Math training ≠ symbolic reasoning.** LFM2-350M-Math spikes at 1-hop (74%) then collapses at 3-hop (38.8%), while Qwen2.5-Math-1.5B maintains strong performance across all hops (46.6% at 3-hop). The difference reflects training data quality rather than the "math" label.
 
-**Qwen's reasoning advantage requires scale.** Qwen2.5-0.5B (42.5%) performs no better than the non-Qwen baseline. The jump to Qwen2.5-1.5B (+8 points at 3-hop) is the largest single scaling gain in the table.