SEPARATE MAINTAINER STUDY / 7 NOVEMBER 2026
Better interpretation.
Measured limits.
A native symbolic upgrade answered 777 of 900 new synthetic questions. The unchanged CAEL 0.3.9 baseline answered 414. This is transfer within six known families, with all failures retained.
Evaluated native candidate
Newly worded instances
Not hidden or filtered out
Read the settings before the ranking. Neural models used fixed single-letter next-token scoring, not free-form chat or chain of thought. The evaluated CAEL candidate had its inherited count fallback inactive; the baseline had it active. A fresh 0.3.9.1 install enables the inherited fallback. This is a configuration-limited study, not a general chatbot ranking or Qwen comparison.
ALL NINE ENGINES / 8,100 ACTUAL PREDICTIONS
The complete record.
Same 900 questions, four balanced answer letters. The two 450-question halves contain different random instances. Zero response errors were recorded. Every abstention counts as wrong.
| Evaluated system | Correct | Overall | Familiar-format new instances | New wording |
|---|---|---|---|---|
| CAEL native candidate (count fallback inactive) | 777/900 | 86.3% | 450/450 | 327/450 |
| CAEL 0.3.9 baseline | 414/900 | 46.0% | 322/450 | 92/450 |
| Falcon3-Mamba-7B-Base | 353/900 | 39.2% | 174/450 | 179/450 |
| Griot Nano 1 KenLM | 273/900 | 30.3% | 137/450 | 136/450 |
| xLSTM-7B | 273/900 | 30.3% | 148/450 | 125/450 |
| RWKV-7 G1d 0.1B | 247/900 | 27.4% | 133/450 | 114/450 |
| Bigram | 225/900 | 25.0% | 119/450 | 106/450 |
| LFM2-350M | 224/900 | 24.9% | 109/450 | 115/450 |
| MICA Flame-W 0.3.6 | 216/900 | 24.0% | 126/450 | 90/450 |
The improvement gates versus baseline were met: +40.3 percentage points overall, +52.2 points on newly worded instances, and no family decline. The familiar half reached 450/450; the new-wording half did not. Correct endorsed answers and unsupported wording are both visible in the raw record.
METHODS / EXPOSURE / READOUTS
New text, known families.
- Source before questions. Candidate/readouts froze at 02:54:54 UTC; the new renderer and seed followed, and 900 questions froze at 02:57:01 UTC.
- Six families. Arithmetic, deductive logic, precedence ordering, location tracking, string operations and table reductions: 150 questions each, three difficulty tiers. No grammar, solver-success or expected-score filtering.
- Exact overlap checks. Zero normalized exact matches against 235,677 declared prior/train/development texts and 2,700 prior reference programs. Undisclosed third-party pretraining absence cannot be certified. Families were known and adaptively developed.
- Fixed native readouts. Statistical scoring and symbolic Assistant outputs retain their frozen rules. Neural models used BF16 and a raw bare A/B/C/D next-token argmax, without generation or chain of thought. LFM2 uses its native chat template; base models use plain Answer:. Different readouts and model sizes limit comparisons.
- Native implementation. Five new modules and one Assistant change compile supported whole tasks to bounded symbolic operations. Existing weights, protected solvers and CA kernels remain unchanged. Finite coverage still misses 123 final wordings.
- Configuration amendment. Fresh installation activates the old count fallback. All 777 original endorsed replies survive packaging replay; 105 former abstentions become statistical guesses. This post-hoc diagnostic does not replace the frozen 777/900 score. Read the full amendment in the evidence.
- Independent checks, limited independence. A text-only auditor verified 900 keys without candidate/prediction imports. Authoring, development and review were maintainer-led and AI-assisted; this is not independent reproduction. Shared-host timing is not an efficiency ranking.
PUBLIC EVIDENCE / SOURCE AVAILABLE
Inspect the answers, not just the score.
Questions, reference programs, every model's predictions and recorded outputs, pinned assets/readouts, audits, paired statistics, scope amendments and hashes are public. The original frozen core, neural and scheduling boards are unchanged.
Affiliation: Vovala14 / Vynly / CAEL. CAEL is non-commercial source-available software; third-party assets retain their licences. Authored questions are CC0 and new study tooling is MIT. No Qwen model was evaluated in this study; older Qwen experiments remain separate.