Parameterless stability strongest leg
Where bare softmax needs external stabilizers and still spikes at high learning rates, rational stays calm — bounded by construction. This is our strongest, most-measured result.
Model X-Ray · VG1 deployed: internal end-to-end operational acceptance completed on 15 September 2026; external customer acceptance is still pending. The free first-beta scope covers eligible public models up to and including 7B, within the account allowance. Earlier withdrawn claims remain withdrawn.
Scope and service status →
Registration page →
Stealth · deep-tech AI · evidence-first
Attention, reimagined for stability.
A drop-in attention operator that stays stable where bare softmax needs external crutches — at zero extra parameters and parity compute cost. Not a quality miracle: a quieter, cheaper, more predictable way to train.
for humans, like humans.
We're raising compute & the right partner for the first properly-budgeted 3B validation — our runs so far (601M–7B) were diagnostic-scale. Fund the experiment that de-risks the thesis, not a claim that it's already proven.
training-gradient stabilitystabilizer-free · high LR
↑ gradient norm · → training step · ┄ excursion threshold
bare softmax — spikes & drifts (stabilizer-free, collapses at high LR)
Tetracta rational — bounded, calm
excursion threshold
softmax Cpk 0.30 · 29 excursions | rational Cpk ~1.59 · 0 — illustrative shape, real measured stats
Cpk = process-capability index (Six-Sigma): how consistently the process stays inside its limits — ≥1.33 is “capable”, <1 is out of spec
Evidence boundary: earlier Model X-Ray examples and interpretations are withdrawn and are not current evidence. Read the correction
0extra parameters
≤ 1%measured compute tax
drop-inop-behaviour preserved (bf16 cos ≈ 0.99999)
withdrawnlegacy Model X-Ray evidence — current status
Pre-registered
Deterministic recompute
md5-sealed evidence
gradient-norm SPC · Six-Sigma
Mechanism not disclosed — evidence is
4-engineer team · ~20 yr
Z-Next · our own architecture
Z-Next: no attention, no KV cache — a session state that never grows
A 947M-parameter language model we trained from scratch (10B tokens, one continuous run). Everything it knows about a conversation lives in a fixed-size state — 5.227 MiB, constant, enforced in code. Not a benchmark play: an economics play.
5.227 MiB
per-session state — constant at any context length
zero bytes
state growth when we fed it 31,926 tokens instead of 952
727 ⇄ 14
parked sessions on one 16 GB GPU vs our same-size twin (one config; twin not KV-optimised — caveats published)
45.24 @ 10B
macro at ~1,800× less training data than the open baseline scoring 66.36
Published with the results, not buried: long-context recall is weak beyond a few thousand tokens; inference is slower per forward pass than a comparable transformer; the mechanism is not disclosed. The full technical report keeps the limitations section longer than the results.
Chat with it live — next to a transformer, memory meters on screen Read the technical report
Two identities, one team
Revenue · today
An applied industrial-AI studio
Custom models, RAG & SFT, data→signal→model and digital-twin consulting, and quantitative research infrastructure — grounded in ~20 years of multi-sector industrial & analytical engineering. Real delivery, not slideware.
What we build
Reputation · IP
A frontier efficiency-research lab
Two lines of original work: Z-Next, our constant-state architecture with a session state that never grows, and a new attention foundation — measured training stability, drop-in / zero-parameter, capability-per-dollar. Both carried on a provenance-verified evidence chain (sha256-sealed artifacts, result-level checks).
The research
Technology & Research
Tetracta Rational Attention mechanism not disclosed
A drop-in replacement for softmax attention. Unlike softmax, its weights need not sum to one — so it can represent "spread little, commit to none" as well as "commit", inspired by the way biological neurons compete and self-limit. We describe what it does, and the effects we measured. We never disclose the recipe.
Drop-in · zero extra params
Identical parameter count and measured bf16 output similarity (cosine ≈ 0.99999) in the reported implementation. It replaced the softmax op in the named test configurations without changing model parameter count. This does not establish compatibility with every architecture, kernel or training pipeline.
Capability per dollar
Hardware-tuned kernels keep MFU at parity with vanilla; the measured compute tax is ≤ 1%. Higher learning-rate tolerance converts into steadier convergence on the same hardware — fewer failed runs, less babysitting, lower training risk.
A modest, honest edge
A small, direction-consistent loss improvement at 3B and 7B — pre-registered and md5-sealed. We sell direction, not a law, and best attribute it to learning-rate headroom. No quality-leap claim.
Model X-Ray evidence status validation pending
Earlier public examples and interpretations are withdrawn. They are not evidence of a validated diagnostic, model-quality, safety, performance or deployment claim. Read the correction
Methodology as a discipline
Pre-registration, deterministic recompute, gradient-norm SPC / Six-Sigma gates and evidence chains are used in the studies that declare them. These controls do not establish any withdrawn Model X-Ray claim.
Abstain research hypothesis
Because the weights need not sum to one, the operator can express "I'd rather not answer" instead of always spreading full attention. Whether this reduces hallucination is not yet proven — a controlled test was an honest null. It is exactly what we are raising compute to test at 3B behavioural scale.
Engineering & compatibility
Compatibility tested in named configurations
Rational replaces one operator. We exercised it inside the transformer configurations and components named below; those tests do not establish compatibility with every implementation, architecture or training pipeline.
The reported experiments used prototype GPU kernels on H100 / H200 and tested multi-GPU (DDP) workflows with resume, checkpointing and gradient-norm SPC in the stated configurations. A rational-specific quantizer remains in development. The measured efficiency numbers and their scope are below.
- qk-norm
- z-loss
- µP
- Muon + AdamW
- Mixture-of-Experts
- GQA
- RoPE
- bf16
- gradient checkpointing
- fused / Flash-style kernels
- multi-GPU (DDP)
Positioned honestly next to prior work — Apple's FlashSigmoid, NVIDIA's nGPT, attention-sink / softmax1 (Miller, 2023), PolaFormer, Microsoft's Differential Transformer. Each ingredient exists somewhere in the literature; our angle is the intersection — a bounded, drop-in, zero-param operator with a measured parameterless-stability effect. We cite, we don't overclaim.
Results & Evidence
What we measured — and you can verify
Revision · 4 Sep 2026 — Earlier Model X-Ray evidence, including the lesion-response result formerly shown here, has been withdrawn. Read the correction and current boundary.
What survived our own attempts to kill it — and what didn't
Selected public claims relevant to this page and their current disposition. For Model X-Ray, the dated correction controls.
| Claim | Status | Where |
|---|---|---|
| Stabilizer-free training stability: 0 spikes vs 29; peak grad-norm 6.6 vs 42.9; Cpk ~1.59 vs 0.30 | SURVIVED | results |
| Compute overhead of the operator ≤1% | SURVIVED | results |
| Kernel equivalence, bf16 cosine ≈0.99999 | SURVIVED | results |
| Z-Next per-session state 5.227 MiB, constant in context length — enforced in code | SURVIVED | Z-Next |
| Token tax: 204 languages × 9 tokenizers, reproducible | SURVIVED | note |
| Loss edge at scale (−0.0146 bpb @3B, −0.0172 @7B) | SURVIVED — attributed to LR headroom, not architectural superiority | results |
| Legacy Model X-Ray findings, figures and interpretations | WITHDRAWN — validation pending | correction |
Diagnostic-scale (sub-Chinchilla, ≤236M tokens) — provisional, pre-registered, reproducible from sealed result files. Headline effects below; the downloadable bundle lets you check the numbers yourself. No formula is disclosed.
Training-process defect rate — gradient-norm excursions (↓ better)stabilizer-free · high LR
Process capability — Cpk (↑ better)stabilizer-free · high LR
Compute efficiency — MFU (parity is the result)tax ≤ 1%
Best-vs-best held-out loss edge — direction across scale (lower = rational ahead)3B · 7B
Read this honestly. Two width points establish a direction, not a scaling law. The edge (~1.5% perplexity at 7B) is best attributed to learning-rate headroom, not a separate nonlinearity advantage — at equal LR with both arms stabilized, softmax comes out a hair ahead (+0.0153 bpb, the same order as the edge we report), which is why we call it parity.
Stability, dramatic
Stabilizer-free, bare softmax shows ~6.5× larger peak gradient norm (≈42.9 vs ≈6.6) and 29 spikes vs. 0 for rational — same body, optimizer, seed and data.
Scale tested
≈601M · 1B · 3B · 7B on H100 / H200. The clean comparison is 3B→7B; the decisive next step is a properly-budgeted 3B run where we expect rational to separate from vanilla.
The ask
Compute + capital + the right partner to convert provisional numbers into production-grade evidence, and to test downstream + abstain for the first time.
Model X-Ray · VG1 deployed · external customer acceptance pending
A checkpoint comparison you can keep.
Compare supported public checkpoints or inspect an in-memory quantization simulation. The versioned HTML report records the observation and its limits; it is not a model-quality or safety grade. Earlier withdrawn findings remain withdrawn.
Tetracta-evaluated research note · merge study
Chimera-1: a training-free merge study
The merge measurements evaluated by Tetracta separately from Model X-Ray remain available; this is not an external independent review. Any former claim that Model X-Ray guidance predicted or improved merge outcomes is withdrawn and is not part of the retained result.
Read the study →
Independent research note · tokenizer measurement
The LLM Token Tax
This public-tokenizer comparison is retained as an independent measurement. It is not a Model X-Ray finding or evidence for Model X-Ray.
Read the tokenizer study →
Engineering & efficiency · measured
Efficiency & ecosystem fit — engineered, not assumed
The new weighting is a custom op, so we built the systems around it — hardware-tuned kernels, multi-GPU training, a deployment quantizer — and proved it composes with the modern stack. Numbers below are measured on H100 / H200 across four model scales. Provisional, diagnostic-scale — and no kernel internals are disclosed.
Scales actually run — 601M · 1B · 3B · 7B
Each isolated, pre-registered and reproducible on H100 / H200. A working pipeline at billions of parameters — not a notebook demo.
Kernel + memory engineering (1B)
Hand-tuning the kernel, memory and optimizer paths more than doubled our own implementation's training MFU (≈5.5% → ≈12%) and lifted its throughput ≈2.7× — with model quality preserved.
Multi-GPU, end-to-end (3B)
A 3B model trained on 2×H200 via DDP at ≈17% MFU and ≈25k tokens/s, checkpoint pulled back locally. Sustained ≈17–18% MFU holds at 7B.
Parity vs softmax — the honest result
Against a tuned softmax baseline the operator runs at parity MFU for a measured ≤1% compute tax — not faster than softmax, but costing essentially nothing extra. Provisionally ~1.2× capability per dollar.
Full ecosystem fit
Trained inside a modern stack — Muon + AdamW, µP, qk-norm, z-loss, MoE, GQA, RoPE, bf16, gradient checkpointing, fused/Flash-style kernels, multi-GPU DDP. One operator changes; nothing else has to.
Deployment quantizer (in progress)
Off-the-shelf tools assume a standard attention op, so we are building a rational-specific quantizer (bf16 → int8 → 4-bit) that shrinks the model while preserving the operator's stability.
Every number measured
Deterministic recompute, md5-sealed evidence, gradient-norm SPC — the same disciplined harness behind the loss and stability results. No hand-waving.
Downloads · sanitized evidence
Don't trust us — verify
The full data — not summaries: every experiment, every training step, every stability statistic. The only things withheld are the recipe (formula, kernel internals, exact learning rates, schedule, SPC thresholds) and quant performance figures — those are the IP. Checksums let you verify each file.
★ Full results report
The complete write-up — isolation design, all results with numbers, every table, methodology, provenance and honest limits. Start here.
Download .mdExperiment ledger — all 17 runs
Per-run BPB, gradient-norm median/peak, excursions, Cpk, defect rate, MFU. Every experiment, nothing cherry-picked.
Download .csv3B convergence — full 36-step curve
Held-out BPB at every step for rational vs. tuned softmax, plus the per-step delta. Rational leads at every converged step.
Download .csvStability pairs — gradient-norm SPC
Stabilizer-free softmax vs. rational at the same LR: gn median/peak, excursions, Cpk, defect rate. The collapse, in numbers.
Download .csvKernel equivalence & efficiency
Forward/backward cosine ≈ 0.99999, VRAM −86% vs. our naive reference, plus a corrupted control that's correctly rejected.
Download .csvScale summary & 7B edge
3B / 7B best-vs-best edge (−0.0146 → −0.0172) + the 7B converged-phase deltas. Direction, not a law.
Download .md.csvStability benchmark — summary
The headline stability table (Cpk, defect, peak gn, MFU parity, drop-in equivalence) in prose + csv.
Download .md.csvEvaluation methodology — one-pager
Pre-registration, matched-step, deterministic recompute, gradient-norm SPC, md5 chain.
Download .mdQuant research — methodology card
Walk-forward, multi-horizon, leak-hardened, placebo-tested. Methodology only — no performance figures.
Download .mdReproducibility manifest
md5 of every public artifact above — verify with md5sum -c. Internal raw-log chain available under NDA.
Download .txt
Want the full, mechanism-level technical brief, the internal raw-log evidence chain, and team
intros? Those are shared under NDA with qualified investors and partners —
request it here.
Capabilities · portfolio
What the team can build
Architectural range that signals one thing to investors and partners: this team ships, it doesn't just talk. ML engineering is one capability among several — the rarer edge is analytical engineering beyond ML.
Quantitative research engine ● live
A production-grade, walk-forward, multi-horizon cross-sectional pipeline — data → signal → model → decision — producing models today, with leak-hardened, placebo-tested evaluation and in-production SPC monitoring. (B2B, licensed institutions only; not investment advice.)
Industrial data & digital twins
Turning raw industrial/market data into decisions: predictive maintenance, time-series, FFT / spectral signal analysis — and digital-twin data strategy (what to collect, how, the real ROI) before a single model is built.
Brain-inspired architecture R&D
Research proof-of-concepts outside the softmax ecosystem (concept binding, spreading activation). Early-stage, framed honestly — evidence of deep architectural flexibility, not a shipped product.
Applied LLM systems
On-prem / data-sovereign assistants: Turkish-fluent SFT, tool-routing, live-web grounding with sources, data-cleaning & QA pipelines, and a rigorous training-QC discipline.
Services · for partners
How we engage
Ordered by how quickly and reliably they create value. We work with partners worldwide; Türkiye is our R&D base, not a ceiling. Under NDA we share enough to evaluate seriously.
Custom model + SFT + RAGcore
A capable assistant on your data & tools — vanilla or rational (a stability / cost option, not a quality claim).
Training-data pipelinesdata
You bring raw text; we return a clean, deduplicated, domain-mixed, packed & tokenized training-ready stream — leak-checked, with a quality report. Your corpus, made model-ready by automated pipelines.
Data→model & digital-twin consultingwedge
Low-commitment diagnostic report → full pipeline. Predictive maintenance, FFT / DSP, regime detection.
Strategic & analytical engineering advisoryadvisory
For companies and individuals: where the value is, which technical-business move to make. Results-oriented, realistic.
Quant research & signal infrastructureB2B
Engine licence / signal-as-research for licensed institutions only. Technology & rigour — not investment advice.
"Türkiye arm" / local presencepartnership
A TR R&D & delivery arm for a foreign firm or lab — including a frontier-lab office option. Our IP stays ours.
Training-stability / QC auditmethodology
We make training failures visible: pre-registered, reproducible, SPC-gated. Live-monitor your own runs.
Flexible / hourly micro-servicesfractional
Editorial, data & evaluation services for teams building Turkish models. Hourly / fractional.
Data-sovereign / on-prem assistantvertical
For regulated SMBs that can't use public APIs (KVKK / data residency): a grounded assistant behind your firewall.
IP · protected by design
Protected, narrowly and honestly
We show the effects; we never publish the mechanism. Where the defensibility actually lives: a measurement infrastructure that survives adversarial review — pre-registration, seed-null controls, signed receipts; initialization & stability discipline learned across four scales of from-scratch training (601M–7B); kernel engineering; and a publication record of killing our own findings, which cannot be copied retroactively. The method stays undisclosed; the discipline is the brand.
Effects, measured and sealed
Every public claim is a specific composition of measured technical effects (training stability + learning-rate headroom) with sealed, recomputable evidence — never a broad formula claim.
Prior art, cited not hidden
We position openly against attention-sink / softmax1, FlashSigmoid, nGPT, PolaFormer and Differential Transformer. Hiding precedent fails diligence; citing it is the defensible position.
Recipe — never disclosed
The formula, kernel internals, exact learning rates, schedule, and SPC thresholds are never disclosed. Effects are public; "how" is not.
Disclosure policy
This site carries effects only and is publication-safe. The mechanism is shared under NDA only; we announce nothing about legal filings in public copy, by policy.
Public disclosure follows IP protection — by design. The moment our priority filing is locked in, we'll open the full results and a technical paper to the world. Until then the complete picture — methodology, raw-log evidence, and the team behind it — is already on the table under NDA for serious investors and partners. The evidence is ready; the public unveiling is just a matter of timing.
About
Stealth by design
We are a four-engineer team with ~20 years of industrial experience on average — top-tier expertise across logistics, industry and software, plus advanced analytical engineering beyond ML: systems, optimization, numerical methods, signal processing, economics. ML engineering is one capability; our rarer edge is the combination. We are results-oriented and realistic — engineering that works and can be measured, not academic abstraction.
We are in stealth to protect our IP. Full technical disclosure and team introductions are available under NDA to qualified investors and partners. Don't trust us — run it: our headline numbers regenerate from sealed, pre-registered result files.
Our ambition is global and our mindset is borderless — we're open to the right strategic partners, investors and support wherever in the world they are. Türkiye is our R&D base and cost-talent root, not a ceiling. We are investment-ready and can incorporate cleanly the moment the right partner and terms appear.
Investors · partners
Fund the experiment that de-risks the thesis
The thesis, one sentence: a measurement-discipline lab — we publish diligence-grade, reproducible evidence, including against ourselves, and we build a constant-state serving architecture cheap enough to test that discipline at scale.
Five numbers: 5.227 MiB constant session state (enforced in code) · Cpk 0.30 → ~1.59 training stability, stabilizer-free · ≤1% compute overhead · 45.24 macro at 10B tokens (~1,800× less data than the open baseline's 66.36; our instruction-tuned line reaches 45.75, statistically equal with our previous generation — supporting records are owner-controlled) · 204 languages × 9 tokenizers, fully reproducible.
The ask, honestly split: the compute for Stage 1 (a Chinchilla-sufficient 3B validation, with pre-registered pass/stop thresholds — if it fails, we stop and publish the failure) is a compute-grant-sized ask and we route it through grant channels. Capital attaches to the revenue side — future validation-stage measurement services and the industrial-AI studio — not to the research bet; a clean entity is formed at term-sheet time.
We measured rational's parameterless-stability, drop-in / zero-param, capability-per-dollar and modest edge at a 200M-token diagnostic scale. The support we're seeking — compute, capital, the right partner — lets us validate at 3B with a Chinchilla-sufficient budget. We expect that run to prove our difference from vanilla more clearly and bring our capabilities into the open:
- we expect the edge over vanilla to hold — and we'll learn whether it comes from the nonlinearity or from learning-rate headroom;
- the first real test of our downstream-capability and calibration / abstain (anti-hallucination) hypotheses — behaviour we hypothesise (not yet demonstrated — a controlled test was an honest null) might set rational apart, which we simply haven't been able to measure yet;
- it turns provisional signal into production-grade evidence.
A virtual data-room — full technical brief, internal raw-log evidence chain, team intros — opens under NDA.
Request the technical brief
NDA-gated. Email us and we’ll reply from a Tetracta address to arrange access.
Email to request accessNo mechanism is shared before an NDA. Effects & methodology are public above.
Contact
Let's talk
Anonymous in public, but ready for online meetings with qualified counterparts. Tell us who you are and we'll share what's appropriate — under NDA where it matters.
- Investors & partners: the 3B-validation brief, results and team intros — under NDA.
- Customers: book a pilot (custom model, RAG / SFT, data & digital-twin, quant).
- Researchers: methodology and an honest discussion of what's proven and what isn't.
Or email [email protected]
Tell us who you are
Prefer e-mail? [email protected] — we never share your details.