fix(benchmarks): count unfenced code, ASCII-safe output, refresh llama3.2 results (#67)

Fixes the local benchmark LOC counter (counted only fenced code, scored bare output 0), makes summary output ASCII-safe (a Unicode arrow crashed the script on Windows cp1252), gitignores generated artifacts, and refreshes the llama3.2 writeup with n=5 data showing the LOC effect is within the noise floor. Follow-up to #63. Verified live.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

.gitignore

@@ -9,6 +9,10 @@ node_modules/
 # promptfoo eval artifacts
 .promptfoo/
 benchmarks/output*
+benchmarks/benchmark-local-results.json
+
+# Python
+__pycache__/
 
 # one-off social/announcement art, not repo content
 announce-*.png

benchmarks/benchmark-local.py

@@ -37,9 +37,10 @@ def load_arms():
 
 
 def count_loc(text):
-    """Non-blank, non-comment lines inside fenced code blocks."""
+    """Non-blank, non-comment lines of code: fenced blocks, or the whole
+    response when the model emitted bare code with no fence."""
     blocks = re.findall(r"```[a-zA-Z0-9_+\-]*\n([\s\S]*?)```", text)
-    lines = "\n".join(blocks).splitlines()
+    lines = ("\n".join(blocks) if blocks else text).splitlines()
     return sum(
         1 for l in lines
         if l.strip()
@@ -110,7 +111,7 @@ def run(model, repeat, ollama_url):
     sep = "-" * len(header)
 
     print(f"\n{'=' * 60}")
-    print(f"  RESULTS — {model}  (n={repeat}, median)")
+    print(f"  RESULTS - {model}  (n={repeat}, median)")
     print(f"{'=' * 60}")
 
     print(f"\nCode LOC per task (median)")
@@ -139,7 +140,7 @@ def run(model, repeat, ollama_url):
 
     out = Path(__file__).parent / "benchmark-local-results.json"
     out.write_text(json.dumps(results, indent=2), encoding="utf-8")
-    print(f"\nFull responses → {out}")
+    print(f"\nFull responses -> {out}")
 
 
 def main():

benchmarks/loc.js

@@ -1,9 +1,10 @@
-// Deterministic code-size metric: non-blank, non-comment lines inside fenced code blocks.
+// Deterministic code-size metric: non-blank, non-comment lines of code. Counts
+// fenced blocks, or the whole response when the model emitted bare code unfenced.
 // Recorded as the `code_loc` metric per arm (always passes; it is a measurement, not a gate).
 module.exports = (output) => {
   const text = String(output || '');
   const blocks = [...text.matchAll(/```[a-zA-Z0-9_+-]*\n([\s\S]*?)```/g)].map((m) => m[1]);
-  const code = blocks.join('\n');
+  const code = blocks.length ? blocks.join('\n') : text;
   const loc = code
     .split('\n')
     .map((l) => l.trim())

benchmarks/results/2026-06-15-llama3.2-local.md

@@ -2,48 +2,52 @@
 
 Same 5 tasks as the Claude benchmark, same three arms (baseline / caveman / ponytail),
 run against a local **llama3.2:latest** (3.2B, Q4_K_M) via Ollama on a Windows 11 machine.
-n=1 per cell. Tooling: `benchmarks/benchmark-local.py` (no promptfoo needed).
+Tooling: `benchmarks/benchmark-local.py` (no promptfoo needed).
 
-## Results
+> **Updated 2026-06-15:** the LOC counter now counts bare, unfenced code. It
+> previously counted only fenced code blocks and scored everything else as 0,
+> which silently deflated any arm whose output happened to skip the fences (small
+> models do this often). Numbers below use the corrected counter at n=5 median.
+> Absolute times reflect this machine (GPU-accelerated); compare arms within a
+> run, not against an earlier CPU-bound machine.
+
+## Results (n=5, median)
 
 **Code LOC**
 
 | arm | email | debounce | csv-sum | countdown | rate-limit | **TOTAL** |
 |---|--:|--:|--:|--:|--:|--:|
-| baseline | 13 | 13 | 38 | 44 | 32 | **140** |
+| baseline | 16 | 18 | 22 | 37 | 16 | **109** |
-| caveman | 16 | 12 | 5 | 45 | 28 | **106** |
+| caveman | 16 | 21 | 18 | 46 | 32 | **133** |
-| ponytail | 18 | 21 | 7 | 38 | 49 | **133** |
+| ponytail | 17 | 22 | 18 | 52 | 28 | **137** |
 
 **Time (seconds)**
 
 | arm | email | debounce | csv-sum | countdown | rate-limit | **TOTAL** |
 |---|--:|--:|--:|--:|--:|--:|
-| baseline | 52.6 | 37.2 | 63.4 | 65.2 | 71.0 | **289.4** |
+| baseline | 3.1 | 3.7 | 3.6 | 4.2 | 4.8 | **19.4** |
-| caveman | 79.6 | 54.4 | 28.3 | 71.4 | 61.0 | **294.7** |
+| caveman | 4.1 | 4.2 | 3.6 | 4.4 | 4.8 | **21.1** |
-| ponytail | 99.7 | 71.0 | 25.4 | 74.8 | 97.3 | **368.2** |
+| ponytail | 4.1 | 4.2 | 3.8 | 4.8 | 4.9 | **21.8** |
-
-**LOC vs baseline**
-
-| arm | total LOC | vs baseline |
-|---|--:|---|
-| caveman | 106 | −24% |
-| ponytail | 133 | −5% |
 
 ## Key findings
 
-**Ponytail does not transfer to llama3.2.** On 3 of 5 tasks (email: 13→18, debounce: 13→21,
+**On llama3.2 the LOC effect is inside the noise floor.** At temperature 0.7 the
-rate-limit: 32→49) ponytail produced *more* code than the no-skill baseline. Total LOC
+per-run totals swing hard: across the five runs, ponytail landed anywhere from
-reduction was −5% vs the 80–94% seen on Claude. Response time increased by +27% (368s vs
+17% *below* baseline to 50% *above* it. The n=5 median came out +26%; a separate
-289s) rather than the 3–6× speedup seen on Claude.
+n=3 median came out −17%. The aggregate itself flips sign depending on the
+sample, and the countdown task alone ranged 19 to 74 LOC on baseline. There is no
+stable LOC reduction to report.
 
-**Caveman outperformed ponytail on this model** (−24% LOC, similar time to baseline).
+**Ponytail does not transfer to llama3.2.** The 80-94% LOC reduction seen on
-Caveman's rules are simpler prose instructions that a small model can follow more reliably;
+Claude is simply absent: the signal is lost in run-to-run variance. The one
-ponytail's multi-step decision ladder requires a stronger instruction-follower.
+consistent effect is on time, and it goes the wrong way: ponytail is ~10-15%
+*slower* than baseline (more system-prompt tokens to process), never the 3-6x
+speedup seen on Claude.
 
-**Why this happens:** Ponytail is a prompt-engineering skill calibrated on Claude models,
+**Why:** ponytail is a prompt-engineering skill calibrated on Claude models,
-which are specifically trained to follow detailed system instructions. A 3.2B quantised model
+which are trained to follow detailed system instructions. A 3.2B quantised model
-partially absorbs the ponytail rules and then adds extra prose *justifying* its choices —
+absorbs the rules only partially and adds prose justifying its choices, paying
-paying the complexity cost without getting the minimalism benefit.
+the instruction-following cost without reliably converting it into less code.
 
 ## Reproduce
 
@@ -51,9 +55,12 @@ Install Ollama and pull a model, then run from the repo root:
 
 ```bash
 ollama pull llama3.2
-python benchmarks/benchmark-local.py --model llama3.2
+python benchmarks/benchmark-local.py --model llama3.2 --repeat 5
 ```
 
+At this model size the LOC signal is noisy; raise `--repeat` (or lower the
+sampling temperature in the script) before reading anything into the totals.
+
 Optional flags:
 
 ```
@@ -63,6 +70,7 @@ Optional flags:
 
 ## Takeaway
 
-The benchmark claims in the README are accurate for the models tested (Haiku, Sonnet, Opus).
+The benchmark claims in the README are accurate for the models tested (Haiku,
-For local/small models, expect significantly smaller — or even negative — gains until
+Sonnet, Opus). For local/small models, expect the gains to shrink into the noise
-instruction-following capability reaches a threshold comparable to Claude Haiku or better.
+until instruction-following reaches a threshold comparable to Claude Haiku or
+better.