Lead with the clause-on configuration: that is the prompt the wallet now sends

README.md

@@ -21,6 +21,10 @@ exact structured tool call a macOS Ethereum wallet can execute. It runs on-devic
 Q4_K_M (5.3 GB) and **scores ~95% on a frozen 1000-case benchmark, above gpt-5's
 92.8% on the same cases.**
 
+**Send it a safety clause.** It was trained without one, and the overall score barely
+moves either way — but refusals go from 81.6% to 95.9%. Details below; the short version
+is that the clause is worth more than the fine-tune on that half of the job.
+
 > *"actually make it 250 USDC instead"* (round 4 of a 6-round conversation)
 >
 > ```json
@@ -48,16 +52,24 @@ strength or continue training.
 deterministic, no LLM judge. Both on-device arms were served from rented GPUs in one
 device-controlled run; gpt-5 saw the same 1000 cases through OpenRouter.
 
-| model | overall | wants a call (880) | wants NO call (120) | safety refusals (49) |
-| --- | --: | --: | --: | --: |
-| **this fine-tune** | **94.9%** | **95.2%** | 92.5% | 81.6% |
-| gpt-5 | 92.8% | 94.1% | 83.3% | 69.4% |
-| Gemma-4 E4B base (untuned) | 90.3% | 91.7% | 80.0% | 61.2% |
-| the previous fine-tune (`gemma-4-E4B-wallet-ft`) | 68.6% | 70.0% | 58.3% | — |
+| model | prompt | overall | wants a call (880) | wants NO call (120) | safety refusals (49) |
+| --- | --- | --: | --: | --: | --: |
+| **this fine-tune** | **+ safety clause** | **95.1%** | 94.7% | **98.3%** | **95.9%** |
+| **this fine-tune** | as trained (no clause) | **94.9%** | **95.2%** | 92.5% | 81.6% |
+| gpt-5 | + safety clause | 93.7% | 93.5% | 95.0% | **95.9%** |
+| gpt-5 | no clause | 92.8% | 94.1% | 83.3% | 69.4% |
+| Gemma-4 E4B base (untuned) | + safety clause | 91.0% | 91.0% | — | 91.8% |
+| Gemma-4 E4B base (untuned) | no clause | 90.3% | 91.7% | 80.0% | 61.2% |
+| the previous fine-tune (`gemma-4-E4B-wallet-ft`) | no clause | 68.6% | 70.0% | 58.3% | — |
+
+**Both rows for this model are real measurements, not one number two ways.** Which one
+applies to you depends on the system prompt you send, so the clause row is listed first:
+it is the configuration the wallet this model was built for now actually ships. Note that
+clause-on, this model and gpt-5 **tie exactly on refusals** (47/49).
 
 **Quote this as ~95%, not to the case.** Three runs of this same GGUF, on different
-rented pods, scored 949, 952 and 944 of 1000. The 94.9% above is one of them, not a
-best-of. Treat any difference under about six cases on this benchmark as unresolved —
+rented pods, scored 949, 952 and 944 of 1000 — the 94.9% clause-off row is one of those
+three, not a best-of. Treat any difference under about six cases on this benchmark as unresolved —
 that noise floor was measured with duplicate identical arms, not assumed.
 
 The previous fine-tune's row is the one worth dwelling on if you are considering
@@ -108,8 +120,10 @@ same pass, and the ablation slice (requests that must produce no call) went to 2
 
 ## Add a safety clause. The two stack.
 
-This model was trained on a system prompt with **no** safety clause. Adding one at
-inference costs nothing measurable and buys a great deal:
+This model was trained on a system prompt with **no** safety clause: all 2288 training
+rows carry the same 533-character system turn, and the clause is 1577 characters that
+none of them contained. Adding it at inference costs nothing measurable and buys a great
+deal:
 
 | | overall | task | refusals (49) |
 | --- | --: | --: | --: |