BUSINESS AI & LLM WORKFLOWS
Give your AI workflow a verified local execution path. Complete supported tasks with fewer model-provider calls, compact requests and checked results—then measure the saving in your own business.
Plan your evaluation ↗See how the 50% savings share works →MEASURED PERFORMANCE · GROWING EVIDENCE
Retaining the original achievement and adding a larger, preregistered evaluation of the current customer package.
ORIGINAL SPEED STUDY · 24 SEPTEMBER 2026
90.02×32/32 exact results in both routes. Two provider calls per task became zero. This measured the earlier candidate against the selected-context API workflow.
See the original trial →SEPARATE REQUEST-REPRESENTATION STUDY
95.05%59,064 → 2,926 tokens across 18 requests. A measurement of request size; the customer’s provider bill is evaluated separately.
See the token measurement →From an initial speed result to a larger test of exactness, safe boundaries and local execution. The studies use different candidates, workflows and timing boundaries; their speed ratios are not a before-and-after regression measurement.
FOLLOW-UP STUDY · 20 OCTOBER 2026
300 supported requests and 100 boundary controls, with source and learned state frozen. These are new measurements with their own comparison boundary.
Every supported request returned exact parameters and the complete numerical result in this structured patch/statistics operation family.
See the workload →Twenty classes of unsupported, ambiguous, stale or unauthorised inputs. Zero false releases. These controls are separate from completed tasks.
See the control results →All 300 exact answers came from local execution. The workers had network access denied throughout the test.
See all three routes →Ratio of median times on 237 mutually correct requests, including local startup and exit. The API route completed 237/300 supported requests exactly.
See timing and denominators →The conventional local route also passed 300/300 answers and 100/100 controls with zero API calls. It was faster in this test, with lighter release checks. Read the full comparison and restart checks →
PRODUCT PRESENTATION · 21 OCTOBER 2026
Twelve slides covering TokenHoloFold, the latest recorded studies and the proposed evaluation and savings-share model.
THE TOKENHOLOFOLD FILM · UPDATED 22 OCTOBER 2026 · 2:20
Aeternum presents the latest 300-task qualification, the 95.05% representation-token reduction, and the separate 10.61× and historical 90.02× comparisons, with their measurement boundaries.
English captions available. Read the transcript · Download captions · Explore the result context
WHAT THIS ENABLES
TokenHoloFold interprets supported requests, checks their context, performs the operation locally and verifies the result before release. That gives an AI application another way to get useful work done.
Complete eligible operations without waiting for a remote model round trip. The measured latency result shows the opportunity; evaluation establishes the effect on your complete application.
Avoid model-provider work where verified local execution meets the task. Count local compute, integration, verification and fallback before calculating the money saved.
A local route can process supported requests without sending them to a model provider. Your integration determines data flows, logging and access controls.
With the required data and dependencies available locally, supported operations could continue without a model-provider connection. Offline operation is qualified for the chosen deployment.
Use a qualified runtime and adapter to connect matching operations to an AI workflow or other business software. You can be a customer using an LLM; you do not need to own one.
Check the supported operation and its result before release. Keep quality, cost and completion visible, so the commercial value rests on successful work.
These are deployment opportunities arising from the supported local route. The evaluation binds the operations, platform and dependencies before customer performance or availability is promised.
PROPOSED PRODUCTION LICENCE
BBF receives 50% of verified net eligible savings. You keep the other half. Agreed running, integration and fallback costs are deducted before the split.
The proposal aligns the production licence fee with measured financial benefit. Where verified net eligible savings are zero, the savings-based licence fee is zero.
Speed, compact representation and provider-call avoidance strengthen the same offer. We count each eligible financial saving once.
Plan your evaluation ↗ILLUSTRATIVE EXAMPLE · THE SAME COMPLETED WORK
For illustration only, suppose actual new costs total £20; £80 is then available to share. This is not a 20% installation charge or a forecast.
Your total spend becomes £60. You retain £40 compared with the £100 baseline.
The £20 represents agreed customer operating costs. BBF’s licence fee in this example is £40.
Explore the illustrative savings workspace →Compare equivalent completed work at actual avoidable costs against the best feasible alternative, including conventional local execution where applicable. Preserve the agreed result quality.
Agree eligible workloads, cost allocation, measurement period, fee cap and reconciliation before production licensing. These are proposed terms, confirmed in a customer agreement.
The evaluation has a separately quoted, capped fee and agreed deliverables. It establishes fit and savings before a production offer; agreed amortisation is counted once in the cost boundary.
CONSISTENCY & COMPACT REPRESENTATION
The latest trial extends the exactness evidence. Representation-token reduction comes from a separate study.
Separate checks confirmed exact results after restarting the local process.
26.834 ms median warm response. These diagnostics are separate from the 300 fresh requests and primary complete-process timing.
59,064 → 2,926 across 18 request representations, counted with o200k_base. This was not retested in the 400-case campaign and is not a total invoice reduction. Read the study context →
For the supported operations in the comparison, TokenHoloFold supplied the interpretation and checked execution locally. The application did not need a remote model answer for those tasks. This is specialised local capability; the call count alone does not measure a change in an LLM’s general intelligence. Work outside the supported scope uses the application’s agreed routing and fallback.
That is the purpose of a scoped integration. Controlled evaluation ZIPs exist for Apple Silicon macOS, Linux ARM64, Linux x64 and Windows x64, with the recorded environments and remaining acceptance steps shown below. Offline operation and customer-specific adapters are qualified for the intended deployment. TokenHoloFold can make supported operations more independent of a provider; it does not transfer the provider’s model into your device.
UK patent application GB2623308.0 was received on 4 October 2026. The application is pending; a patent has not been granted.
Customer integration is prepared through a scoped business evaluation. Public SDK distribution and live billing are separate release steps.
FROM FIT TO DEPLOYMENT
Define the integration and compare complete, equivalent work. The evaluation report informs the installation and commercial offer.
Describe the repeated operation, deployment platform and current implementation. We agree the technical scope, delivery and evaluation fee before work begins.
Build your brief →Keep token usage, runtime cost, quality and fallback visible. Explore how a customer savings workspace could report the agreed comparison.
Explore workspace preview →Qualify the exact runtime and adapter for the customer's supported workflow before software delivery. General-purpose LLM compatibility is assessed individually.
Download evaluation outline ↓EXTEND THE WORKFLOW
TokenHoloFold can be evaluated on its own. Add GLYX for exact retained records, and explore Seed State for selective state reconstruction. Each performs a complementary job.
Licensing is non-exclusive. TokenHoloFold's 50% share applies to verified net eligible savings; a combined agreement counts each saving once. Installation costs are actual agreed costs, with no automatic installation surcharge.
Explore the product combinations →CONTROLLED EVALUATION PACKAGES
Your team receives a versioned ZIP, installation instructions, integrity checks and a self-test for the agreed workflow. There is no automatic installation surcharge.
| Platform | Package and recorded checks | Customer acceptance |
|---|---|---|
| Apple Silicon Mac | Installer, preflight, self-test and uninstaller. Package-relative crypto dependency. 161/161 checks on the owner Mac. | macOS 26.0+ and Python 3.10+. Independent Mac acceptance and signing/notarisation remain to be qualified. |
| Linux ARM64 | Evaluation ZIP; 163/163 checks in a local ARM64 Linux virtual machine. | Debian 12 / glibc 2.36+ starting point. Test on the customer's target host. |
| Linux x64 | Evaluation ZIP; 163/163 checks under x64 emulation. | Native customer-host acceptance is the next platform check. |
| Windows x64 | Evaluation ZIP and laptop test bundle; 160/160 compatibility checks under Wine. | Actual Windows 11 Intel/AMD laptop testing is outstanding. |
The frozen evaluation supports a defined structured patch/statistics operation through CLI, JSON Lines or Python adapters. A customer's operations and AI application are qualified before delivery. General download distribution is separate from a controlled evaluation.
Define your platform and workflow →