A Canadian-developed geometric language model called an "FMM" (Fractal Morphological Machine) scored a perfect 100% across 70 adversarial spelling and grammar torture tests, leaving conventional cloud AI and rule engines choking in the dust, according to internal benchmark data released Wednesday.
The technical evaluation, published Sept. 2 by Bad Character Scanner Labs, pitted their product called "BCorrect™ API" against established industry alternatives LanguageTool, PySpell NLP, and TextGears.
The verdict was an absolute blowout.
While competing systems stumbled over space-shifted compound words, extreme phonetic letter scrambling, and STEM mathematical syntax, BCorrect’s hyper-dimensional Fractal Morphological Machine (FMM) resolved all 70 challenge cases without a single hallucination.
Internal Benchmark & Best Estimates
| Engine Architecture |
Adversarial Score |
Corrected Targets |
Response Latency |
Hardware Required |
| BCorrect™ FMM (Native) |
100.0% |
70 / 70 |
~14.2 ms |
1x Standard CPU |
| LanguageTool (Rule Engine) |
55.7% |
39 / 70 |
~350 ms |
Dedicated JVM |
| PySpell (Statistical NLP) |
42.9% |
30 / 70 |
~180 ms |
Local Python |
| TextGears (Cloud REST) |
11.4% |
8 / 70 |
~820 ms |
Cloud Server |
[NOTE ON METHODOLOGY] These figures reflect our preliminary internal test harness and exploratory engineering estimates. They have not been audited by an independent third party and are not presented as absolute scientific truth. We are actively seeking third-party academic or industry partners to independently verify and audit these results.
Source: Internal exploratory evaluation TR-2026-FMM-04, Sept. 2, 2026 (70 internal adversarial targets).
Where Rivals Faceplanted
Legacy checkers crashed on three critical stress tests:
- Space-Shifted Words: Missed "andre painted" → "and repainted".
- Phonetic Scrambling: Choked on "teh purpl elefant" → "the purple elephant".
- Formulas & Brands: Mangled LaTeX (
$\Delta T = \frac{Q}{mc_p}$) and downcased trademarks.
"You don't need an eight-card H100 furnace to fix a sentence," said J. Shoy.
Sips Power: Engineering Estimates
Our internal calculations on the physical efficiency of FMMs (preliminary engineering estimates):
- Sub-15ms Latency: Native CPU binary execution eliminates cloud queue lag and token streaming stutter in internal runs.
- Estimated 40,000x Lower Energy Draw: Our calculated estimate is ~0.00012 kWh per million words on a commodity CPU, compared to ~4.80 kWh on an eight-card H100 cloud cluster (based on single-core 22W execution vs 10 kW GPU cluster runtimes; independent power metering audit welcomed).
- Zero Data Retention: Sentences are processed in volatile memory and discarded. No prompt harvesting, no retention logs, no training on customer text.
Availability & Verification
We are actively seeking third-party researchers and testing labs to verify these findings. Academic NLP groups and independent evaluators can contact our team for free verification API tokens to run independent audits.
The BCorrect API is live in production today starting at $5 for 2 million characters (~360,000 words)—roughly 94 per cent cheaper than token rates on frontier LLMs.
Developers can test the engine live on the BCorrect Pro Web Studio or deploy API keys directly through the API Documentation Hub.
[Complete Editorial Disclaimer & Legal Notice]
The preceding article represents an editorial news analysis and analytical commentary based on internal laboratory benchmark evaluations conducted by the Bad Character Scanner research group. All performance metrics, adversarial test indications, latency figures, and energy estimates reported herein reflect internal test harnesses and specific exploratory evaluation sets (Technical Report TR-2026-FMM-04). Testing methodologies, proprietary evaluation corpora, mathematical parameters, scoring heuristics, and the exact mathematical character and topological dimensionality of the H-Manifold™ architecture are classified trade secrets and are not publicly disclosed. Third-party brand names, models, and trademarks (including OpenAI, LanguageTool, PySpell, and TextGears) are the property of their respective owners and are referenced strictly for comparative and nominative reporting under Canadian Press journalistic fair-use principles. Independent external audits have not been performed, and individual real-world production performance may vary.
