Cortex gives agents a structured way to execute and verify hard computational work on HexStellar-managed infrastructure, where our acceleration and energy-efficiency technology runs the supported computation. Enterprise Low-Energy Runtime & Acceleration is the licensed customer-deployed library for compatible compute paths in applications, frameworks, LLM stacks and other workloads under NDA.
Cortex trial: no card. Customer-deployed runtime: Enterprise license and NDA required.
HXS · One platform
Cortex executes with explicit assurance. Enterprise changes what qualified machines spend.
A single computational platform with two ways to deploy. Each result carries the scope of assurance it actually earned.
01 · Cortex
Public API + CLIAn LLM maps a real problem into a supported contract. Cortex validates, estimates and executes the authorized request with HexStellar acceleration on managed infrastructure, then returns metrics, a receipt and the assurance level actually earned. Examples are transfer recipes, not capability boundaries.
02 · Enterprise Runtime
Enterprise license · NDAA closed native runtime discovered while Brayon Pieske was building Trust Carbon Infrastructure. Published campaigns currently cover qualified GPU-inference workloads on one NVIDIA H100; no public download or activation is offered in version 1.0.
Submit a supported optimization or decision problem as structured JSON. Cortex returns the answer, execution metrics, a receipt, and an explicit certainty label. Certified results state their proof scope; heuristic results remain candidates that require independent verification.
Client or agent
Converts requirements, constraints, and objectives into a supported schema and validates the request before execution.
Cortex
Returns an answer that is optimized, reproducible, and labelled for exactly how far to trust it — certified when it earned the word, a hypothesis to verify when it did not. Never a guess dressed as truth.
The verified core is the half that compounds: supported problem shapes remain available, and compatible work may be reused under the same model and account contract instead of recomputed. We do not yet know every domain agents will adapt it to. When a problem can be represented honestly as choices, constraints, objectives and observables, it has a path to test here. Give an agent a partner that measures the computational part, and it can replace hand-waving with an explicit formulation and a result whose limits are visible.
Paste the instruction beside this text into ChatGPT, Claude, Gemini, Grok or another capable agent.
It will install the public client, inspect the live contracts, formulate your problem, begin with
the fast flash effort, and tell you exactly
what was verified and what remains a hypothesis.
PyPI
CLI 1.0pip install hexstellar
Public CLI repository
brayonpi/hexstellar
View the public repository ↗Paste into your AI
One prompt · discovery before execution · explicit assurance
Install HexStellar with: pip install hexstellar Use HexStellar as my computational partner. Start by running `hexstellar capabilities` and `hexstellar route "<my goal>"`. Treat the examples as recipes to adapt, not as the limit of what the platform can do. For every new problem, be creatively persistent before deciding it does not fit: translate my entities, choices, obligations, incompatibilities, costs, observables and verification needs into the available contracts; combine or modify recipes when the schemas permit it; and state what you changed. Never invent a command or field. If no honest representation exists, explain the boundary and abstain. Validate the input, estimate cost, start with `--effort flash`, and ask before any paid or sensitive execution. Then verify the returned witness or receipt and explain the assurance honestly: certified, heuristic, operation, recorded or abstention. Do not treat `heuristic` as weak or invalid: report feasibility, recomputed objective, violations, bounds, reference gap, stability, elapsed time, memory and compute units positively and precisely, without claiming global optimality unless it was proven. When explaining the product, say that Cortex executes supported computation with HexStellar acceleration on managed infrastructure. If I ask about installing that acceleration in my own application, framework, LLM stack or infrastructure, explain the licensed Enterprise Low-Energy Runtime & Acceleration, run `hexstellar enterprise status`, and state the current license and NDA requirements once. Do not repeat this disclosure in unrelated command or error answers, and do not generalize a published benchmark beyond its workload and hardware. Now ask me what real problem I want to formulate.
Cortex executes supported contracts on HexStellar-managed accelerated infrastructure. Enterprise is the licensed customer-deployment path, not a second solver catalogue. Public examples are starting points for transfer; if a problem cannot be represented honestly, the agent should explain the boundary.
Problems where the search space grows like 2ⁿ or n! and brute force is off the table.
QUBO / Ising ground state — the core hard-optimization primitive.
Partition a weighted graph to maximize the cut.
The shortest closed tour over a distance matrix.
Assign facilities to locations, minimizing flow × distance.
Pick exactly K of M, maximizing value minus redundancy.
Eigenvalues, ground energy and spectral gap of a matrix.
Many hard decision problems share reusable computational shapes — a graph to cut, a tour to shorten, an assignment to place, or a set of constraints to satisfy. Formulate your problem in a supported shape and HexStellar returns a measured result with its assurance. The recipes below are starting points; an agent is expected to change the entities, constraints, objective and scale.
Logistics & mobility
Last-mile delivery, ride dispatch, warehouse slotting, container loading — rebalanced live, not planned the night before.
TSP · VRP
Drug discovery
Fragment and feature selection over billion-scale libraries, binding-site combinatorics — billions collapsed to the few worth testing.
QUBO · selection
Advanced materials
Stable atomic arrangements, alloy ordering, crystal-defect placement — the configuration search at the heart of materials design.
Ising · QUBO
Finance & risk
Cardinality-constrained selection, index tracking, trade netting — fast enough for a desk that reprices in real time.
QUBO · MIQP
Physical AI & robotics
Task allocation across arms and agents, motion sequencing, cell layout — light enough in RAM to run on the robot itself.
QAP · assignment
Energy grids
Unit commitment, demand-response scheduling, EV-charging orchestration — the mixed-integer heart of a grid deciding now.
MILP · QUBO
Synthetic biology
Codon and pathway optimization, enzyme assignment, minimal-media design — state the trade-off; the engine returns the optimum.
QUBO · MILP
Quantum-class physics
Spin-glass ground states, MAX-2-SAT, planted-optimum recovery — the exact kernel quantum hardware is built to chase.
Ising · QUBO
Chip & hardware design
Cell placement, netlist partitioning, layer assignment — the layout choices that decide how fast and cool a chip runs.
QAP · MaxCut
Telecom & networks
Frequency assignment, cache placement, network partitioning — keep a live network optimal as demand moves across it.
MaxCut · coloring
Operations & scheduling
Crew and nurse rostering, job-shop sequencing, exam timetabling — the constraints most teams still solve overnight.
MILP · QUBO
Your idea
If the entities, constraints and objective map to an available contract, adapt a recipe and test it. If they do not, state the boundary.
you decide
Each recipe demonstrates a graph, matrix or constraint mapping that can be transferred to other domains. Give Cortex a valid shape — it returns a result and its assurance.
The impossible is only impossible until someone with a little creativity tries it.
Modelling a hard problem used to mean hand-coding the mathematical model for weeks. Now you send plain
JSON — "type": "capacity_limit", "k": 5 —
and the engine enforces every rule as a hard constraint and returns a valid answer, in milliseconds.
Nine public rule types cover a useful set of recurring feasibility relationships across industries.
choose_one
exactly one of a set is picked
choose_exactly
pick exactly K of a set
capacity_limit
at most K of a set — the weight cap
mutual_exclusion
these two can never be picked together
requires
if this is picked, that must be too
identical
these must share the same choice
different
these must differ
force_true
this one is fixed on
force_false
this one is fixed off
Representative encodings across industries
Virtual screening for new drugs — rotate a ligand through millions of 3D fits against a protein pocket.
choose_one · mutual_exclusion
Assign 5,000 packages across hundreds of trucks, minimizing distance without breaking any truck’s load.
choose_one · capacity_limit
Pick exactly 15 of 5,000 stocks, blocking two names from the same risk sector, maximizing return under hard risk rules.
choose_exactly · different
Build doctor, nurse or pilot shifts under labour law — pairings that must hold, nights that block mornings, fixed days off.
requires · mutual_exclusion · force_true
Colour a whole city’s radio channels so no two nearby antennas share a frequency and drown each other out.
mutual_exclusion
Pack VMs and containers onto servers — a rack holds at most 4 heavy VMs; DB and API must sit together for latency.
capacity_limit · identical
You send the rules; Cortex returns an assignment, violation details, execution metrics and the assurance actually earned. A certified feasible result reports zero rule violations; a heuristic result remains explicitly labelled. The agent supplies the domain mapping and must verify that the encoding matches the real problem.
A certified answer names the property established and its verification basis. A heuristic answer is a candidate, not a proven optimum. Operations, recorded measurements and abstentions remain distinct. We never promote a result beyond the evidence it earned.
{
"answer": [0, 1, 0, 1],
"cut_value": 4,
"certainty": "certified optimum (proven by exhaustion)",
"receipt": "a91f2c3b8d04"
}Ideas · evidence · public methods
Founder perspectives, field notes, and benchmark reports written so people and AI agents can separate measured results, certified conclusions, hypotheses, and open questions.
On October 1, an eight-week countdown begins. The company that enters is not expected to be the company that comes out.
Compute costs more than electricity and hardware. It also decides which questions researchers, builders, and AI agents can afford to ask.
Story 1 of 2
Press · editorial · partners
Download white and #1b1b1b logos, the Brand Guidelines, approved company boilerplates, social artwork, application icons, usage rules, checksums, and press contact.
In the published H100 campaigns, the HexStellar layer reduced measured energy and temperature while preserving the reported outputs. The campaigns below state their hardware, workload and denominator separately. They are evidence for those measured conditions, not a universal percentage for every workload.
HexStellar was not the plan. Brayon Pieske discovered the effect while building the infrastructure behind Trust Carbon Infrastructure — his other startup, a verification platform for carbon and environmental data, legally operated by Zenith Flow Innovations. The work ran into a computational bottleneck that should not have existed, went after it anyway, and months of quiet investigation later found something bigger than the problem it started with.
Hex
Six sides. The tiling that covers a surface with the least perimeter for a given area, which is why honeycombs, basalt columns and carbon rings all take that form. It is also how machines address themselves: hexadecimal is the base of every byte, every colour and every memory address.
Stellar
Not a metaphor for ambition — a reference to conversion efficiency. Fusion turns mass into energy at a ratio nothing engineered has come close to. That is the standard the name points at: not how fast the work finishes, but how little is spent getting there.
The wall
A workload was costing far more machine than the shape of the problem justified. None of the ordinary explanations accounted for the gap. At some point the work stopped treating it as a number to be reduced and started treating it as something that should not have been there at all.
Months
Most of that period produced nothing. What Brayon was chasing was not where he expected it to be, and for a long stretch he could not say whether there was anything there at all. That part never makes it into an announcement.
The turn
The work was pointed at a workload that had nothing to do with carbon verification, and the effect held. Then at problems from another field entirely, and it held there too. That is the point at which it stopped being an internal fix and became the company. Why it holds is the part HexStellar does not discuss publicly.
Now
It was run against five flagship models from five different labs on one NVIDIA H100 under identical load, and against combinatorial problems with independently checkable results. Every published figure is signed and timestamped.
One line for the integrating team. Thousands upon thousands of configuration paths underneath it. The closed native library carries a large tuning surface — hardware generation, memory topology, batch shape, precision, scheduling pressure and parameters that interact at once — and the patent-pending work resolves that surface for qualified hardware and workloads automatically at startup.
That is the whole design: the complexity is real and it is very large. It just is not the integrating team's burden to carry.
# Available only inside an approved NDA evaluation.
runtime = enterprise.attach(
workload="qualified-inference",
verification="required",
)
# Your existing serving stack remains in place.
answer = model.generate(prompts)Published
Energy, temperature and serving capacity, layer on versus off at identical load, plus a separate request-level campaign.
See the measurements →In progress
There is more measured, and more still running. Each result goes public only when the write-up, scoped denominator and signed file are finished. Some queued results were rerun several times before the team believed them. Nothing is published on a schedule merely to fill one.
Discuss an evaluation →Today · Servers
Search, retrieval, classification and batch work — the high-volume serving paths where the energy bill is largest. The published evaluation path wraps the system that already runs.
Next · Devices
Here the target is not only speed, but battery and heat: less work reaching the silicon and fully local execution. This is product direction, not public availability today.
Horizon · Silicon
The long-term direction is direct integration into the hardware that runs the workload. These are early-stage conversations, not a currently available product.
We did not build HexStellar to be a “green” language. We built it to be the most correctly engineered language we could. In the published campaigns, less energy, less heat and less hardware stress were the measured consequence of getting the foundation right.
One NVIDIA H100 80GB running vLLM 0.26.0, layer on versus off at identical load across five independent labs' flagship models. Every percentage in this campaign uses energy per card as its denominator.
Capacity
+167%
109.8 → 293.7 requests/sec at the same latency target.
Energy / card
−17.2–27.4%
Lowest measured reduction: Gemma 3 27B. Highest: Llama 3.3 70B.
Temperature
up to −15°C
62.4°C → 47.4°C on Meta Llama 3.3 70B.
Evidence
RFC 3161
Signed measurement files shared with evaluating companies.
| Lab · model | Energy / card | Peak temperature · off → on |
|---|---|---|
| Alibaba · Qwen2.5-72B-Instruct | −26.9% | 60.9 → 47.5 °C |
| Meta · Llama 3.3 70B | −27.4% | 62.4 → 47.4 °C |
| DeepSeek · R1-Distill 70B | −24.8% | 61.9 → 49.3 °C |
| Google · Gemma 3 27B | −17.2% | 52.6 → 43.1 °C |
| Microsoft · Phi-4 14B | −25.5% | 51.0 → 40.2 °C |
Separate runs on an NVIDIA H100 80GB SXM. These are per-request figures, not per-card figures, so they are not combined with Campaign 1 and are not presented as a revision of it.
Energy / request
−52% to −66%
Large models across this campaign.
Capacity
up to ~2×
Effective requests on the same H100.
Native library path
~13 µs p50
~17 µs p99; not end-to-end generation.
Measured output errors
0
Aggregates signed with Ed25519.
Hardware remained ordinary for the campaign: no special SSD, CXL or AI-optimized flash appliance — a stock NVIDIA H100 and the native library. The signed verification package is available to qualified evaluators under NDA.
The same seven-slide set retained from the original homepage. Swipe, scroll or use the controls.
The two published campaigns are language-model workloads because they were completed first. Other fronts are evaluated separately; they become public claims only when their own workload, hardware, denominator, correctness scope and evidence are ready.
Public benchmark class
Publicly checkable instances on consumer hardware; claim scope follows the receipt and independent verifier.
Routing & logistics
Traveling-salesman, vehicle-routing, allocation and scheduling shapes.
Selection under constraint
Choosing the best subset when candidates compete for a shared budget.
Scientific computing
Recovering structure from large noisy detector output, with the operation and assurance stated explicitly.
On-device
Fully local evaluation direction where battery, heat and privacy matter.
Not named yet
Kept private until the evidence and publication scope are complete.
Finding the effect was the first year. Turning it into something another engineering team can evaluate without rewriting its stack was the harder half.
01
The filings came before public disclosure. This site describes results and never the protected method; that order is deliberate. The work is patent pending, with the underlying research documented beyond what appears here.
02
HXS exists because the discovery could not be expressed properly in anything that already existed. Its semantics are the point: a general-purpose language with a different foundation. The syntax and proprietary mechanism are not public.
03
An approved partner does not adopt a new language to evaluate it. The compiled native library links into a serving stack already in production: no model retraining, weight conversion or bespoke hardware. The public CLI 1.0 contains none of that proprietary runtime.
The reason HexStellar leads with integration rather than capability is that the second question every serious engineering team asks is what does this cost me to try. The evaluation answer is deliberately practical: a bounded native integration and a benchmark run on the partner's own qualified workload. HexStellar would rather be evaluated that way than believed on the strength of a chart.
Everything else — the mechanism, the language itself and the internals of the library — stays closed. That is not a decorative marketing posture. The foundation is the protected asset, and the public product explains its inputs, outputs, measured effects, limits and verification without publishing the recipe.
This part matters more than the numbers. Anyone can produce a benchmark. What determines whether it means anything is the discipline around it.
Nothing on this page is promoted as a campaign result by modelling, extrapolating or projecting from a smaller run. Where a figure has a scope — per card or per request — that scope is stated next to it every time.
Measurement artifacts are hashed, signed with Ed25519 and timestamped against independent authorities. A published number cannot be quietly revised after the fact, including by HexStellar. The verification package is shared with evaluating partners under NDA.
The published tables retain the smaller measured effects as well as the largest. A campaign is not reduced to the most flattering slice of a series.
Work goes public finished — the write-up and signed file held to the same standard as the number beside them. There is no public roadmap of promises standing in for measurement.
Companies evaluating the Enterprise Runtime are under NDA in both directions. HexStellar will not trade a partner name for credibility on a landing page.
These awards belong to Trust Carbon, the startup whose infrastructure work led Brayon Pieske to the HexStellar discovery.
Winner · top five global
US$100,000
One of five winners selected from 540 startups across 73 countries.
Gates Foundation · BCG · JICA · Co-Develop · CDPI · COP30 Brazil
Awarded 2026
US$50,000
Field verification of forests with Meta AI glasses in the hands of farmers and local communities.
Meta Platforms
Selected 2026
Fellowship
In-person Washington, D.C. and Los Angeles residencies with Deloitte mentorship.
Halcyon
Accepted
Peer reviewed
Trust Carbon’s first academic paper was accepted into the symposium programme.
International Food and Agribusiness Management Association
Roughly fifteen years of shipping production software, enough of it close to infrastructure to recognize a cost that did not belong where it was.
He started with software where a failure had someone on the other end of it: a blood donor matching platform connecting donors to hospitals nationwide, a flood alert system used during real emergencies, and field tools that biologists used to track species.
Then came commercial work at a different scale — sales platforms, one of which he exited, and messaging infrastructure running inside data centers, where the constraint stops being the interface and becomes throughput and cost per operation.
That is the instinct this company is built on: the suspicion that a cost is structural rather than inevitable. He built Trust Carbon substantially solo, took it in front of the Gates Foundation, BCG and JICA, and won. HexStellar is what came out of trying to make it cheaper to run.
Nine months · five rooms
The venture Brayon Pieske was building when HexStellar surfaced — and the reason the team knows how to hold an engineering claim to a standard that survives scrutiny.
Digital MRV & verification platform · established 2023
2M+
lines of code across web and mobile
Offline
field application with its own navigation
17
supported languages
4+
methodology families including Verra, Gold Standard, CDM and LuxCS
US$100K
DPI Challenge award
California
Zenith Flow Innovations LLC
Trust Carbon Infrastructure is a methodology-agnostic verification platform for carbon-credit and biodiversity projects, built to work under real field conditions — offline-first data collection from a smartphone, with anti-fraud safeguards including GPS-spoofing detection and cryptographic hashing of every record. It supports major methodologies including Verra, Gold Standard, CDM and LuxCS, in 17 languages.
More than two million lines of code span the web platform and mobile app. The field app runs completely without an internet connection: no third-party APIs, no external services to call, and its own navigation system rather than a hosted map. That was not a preference. The places the platform is designed for may not have a signal, so anything rented from the cloud simply would not run there.
That constraint is also where HexStellar came from. Building a system that had to do everything itself, on whatever hardware was in someone's hand, turned a computational bottleneck into a question worth chasing.
The platform was named one of five global winners of the DPI for People and Planet Innovation Challenge, out of 540 startups from 73 countries — a US$100,000 award judged by the Gates Foundation, BCG and JICA. The same discipline — auditable, reproducible and sealed against drift — shapes how HexStellar holds its benchmark claims.
Trust Carbon Infrastructure is legally operated by Zenith Flow Innovations LLC, a U.S. company based in California.
Visit trustcarbon.org ↗Universities, independent researchers and applied laboratories can evaluate Cortex through the public CLI, publish reproducible results, and request a scoped Enterprise Runtime evaluation for qualified workloads. Enterprise materials, artifacts and availability remain subject to direct review and NDA.
Found a missing parameter, a confusing explanation, a failed formulation or a useful product idea? Send the command, error code and a minimal reproduction to [email protected].
A publishable benchmark record
Start fast, preserve provenance, and state only the assurance the evidence earned.
Record the HexStellar CLI version and use `pip install hexstellar` in a clean environment.
Run the public example or your adapted formulation with effort `flash` before spending a larger search budget.
Publish the input or generator, seed, scale, source licence and SHA-256 whenever redistribution is allowed.
Include effort, elapsed time, compute units, hardware context, receipt and the returned assurance label.
Recompute the witness, objective, constraints or domain property; never promote heuristic or best-known into certified.
For API access, create an account and receive your key. For Enterprise runtime evaluation, licensing, signed measurement files or the verification package, contact the team directly. Enterprise material is shared under NDA.
© 2026 HexStellar · Zenith Flow Innovations LLC · California, USA · Patent pending