Exactly 500 characters: the model can do it — half the time. The harness made it certain.
The genre
This site's oldest trick is forcing a model through a constraint it keeps missing: exactly 100 words was 0-for-10 before sealed sections made it 3-for-3. So for this slate we picked a fresh constraint of the same family — a product description of exactly 500 characters, spaces and punctuation counted — froze the protocol before running, and gave the baseline every fair chance: one competent prompt, the counting rule stated plainly, ten attempts.
The result
| Arm | Result |
|---|---|
| Baseline (one-shot, deepseek-flash) | 5/10 exactly 500 — the rest landed at 504, 507, 505, 504, 504 |
| Harness (5 sealed 100-char sections, machine-counted, bounded repair) | 3/3 assemblies at exactly 500 |
Pre-registered verdict: NOT PROVEN — the bar required a baseline at ≤1 of 10, and this baseline is a coin flip, not a failure.
Read the baseline's five misses again: 504, 507, 505, 504, 504. The current model reasons before it writes, and it counts in that reasoning — it is never far off, and half the time it is exact. Its five exact texts are genuine, coherent descriptions, published in the evidence below. The "models can't count" meme is going the way of the strawberry meme: true of an older model, repeated about a newer one.
What the harness actually bought
Not possibility — certainty. A user who needs exactly 500 characters cannot tell a 500 from a 504 without counting it themselves; half of what the unaided model hands them is quietly wrong. The harness run was exact 3 of 3, all fifteen sections machine-counted on arrival, with exactly one repair instruction needed in fifteen sections. The claim this supports is narrow and stated with its N attached: structure + verification turned a 5/10 coin flip into 3/3 verified exactness, on this task, on this model.
The winning assembly, verbatim — exactly 500 characters, seams and all, as produced:
`` Meet the solar camping lantern built for campers, hikers, and emergency prep in remote cabins, vans. A solar panel charges the battery by day, giving bright, steady light all night outdoors, anywhere. As dusk settles, a warm glow lights the table, tent, and faces—camp feels calm and cozy and bright. Durable drop-tested ABS and metal frame, IPX4 water resistance, weighs just 0.6 lb for camping use. Pack light, stay bright: this solar lantern belongs in every camp kit. Grab yours today, and shine. ``
Evidence — re-grade it yourself
| File | What it is |
|---|---|
| C-force500/PREREG.md | Frozen protocol + pre-registered metric |
| C-force500/run_c.py | The runner: baseline, section builder, counter, repair loop |
| C-force500/runs.json | Every baseline attempt (full text + count) and every section attempt of all three assemblies |
| C-force500/RESULTS.md | The results write-up, spend included |
| AMENDMENT.md | The transport incident that preceded the runs |