Enterprise corpora can exceed model context windows by orders of magnitude. Sending, chunking or repeatedly summarising everything adds cost, latency and information risk before reasoning begins.
The QCE — Quantum Context Engine applies a proprietary quantum-inspired selection framework to configured record signals and a fixed token budget. It returns full selected records—not generated summaries—for the customer's downstream model.
Six public semantic domains, 898 independently labelled query units, 21,130 recorded arm-level outcomes, 2,574 model/planner requests and physical testing through 5M estimated tokens. Failures remain reviewable.
The latest result is a measurable infrastructure wedge: QCE processed a 5M-token physical corpus that the tested full-model arms could not accept, retained 91.5% of labelled evidence, and passed only 1M tokens onward. Partner-owned replication is the next value inflection.
Arm-level outcome counts include ablations, scale points and model comparisons; they are not independent samples. Physical 5M is distinct from represented 500M. The latest research candidate is not automatically routed in production.
Supply authorised records, messages or document chunks with stable IDs, source/speaker metadata and token costs. Very large archives use bounded shards rather than one unsafe request.
Our algorithm scores every record across multiple mathematical dimensions simultaneously using our proprietary quantum-inspired framework.
We return the actual records — unmodified, unparaphrased. Your LLM gets the truth, not an interpretation of it.
The latest scale ladder reported 48.8% reduction with 96.5% recall at 500K, and 80.0% reduction with 91.5% recall at physical 5M. At represented 500M→5M, reduction reached 99.0% but recall fell to 62.5%—the next quality frontier.
| Input → budget | Recall | Precision | MAP / MRR | Reduction |
|---|---|---|---|---|
| 500K → 256K | 96.5% | 1.39% | .5387 / .6880 | 48.8% |
| Physical 5M → 1M | 91.5% | .80% | .3996 / .5599 | 80.0% |
| Represented 500M → 5M | 62.5% | 12.48% | .3918 / .5576 | 99.0% |
Precision uses incomplete public qrels: unjudged selected records count as nonrelevant. Represented 500M reweights a bounded physical 5M corpus; it is not physical ingestion of 500M tokens.
In the locked no-assistance study, full physical-5M inputs were context-ineligible for Sonnet 5, Gemini 3.8 Flash and Opus 5. Ineligibility is an operational outcome—not a model-quality score.
Customer keeps model choice. QCE returned a deterministic 1M-token evidence budget from 5M physical tokens. Gemini 3.8 accepted eight of ten reduced semantic prompts; Sonnet 5 and Opus 5 remained ineligible after declared input plus overhead.
500K: all three latest model families were eligible for an attempt; each completed 8/10 semantic calls. 5M: no full-model arm was eligible. Represented 500M: QCE reduced declared input by 99%, but the approximately 5M-token output remained beyond every tested context.
At matched 500K valid pairs, QCE reduced model tokens by a median 53.17%. Recall improved for Sonnet and Opus, and fell 1.25 points for Gemini. Pair counts were 6–8, so this is diligence evidence rather than formal superiority.
Future path: Weavi trains and operates domain-specific LLMs (legal, medical, financial) optimised to work natively with the QCE. Full-stack AI service. Higher margin. Deeper customer lock-in.
These are not forecasts. These are verified figures already published in 2026. The AI token economy is here — and it is enormous.
7× year-on-year growth. 375+ Google Cloud customers each processed over 1 trillion tokens in the preceding 12 months. 19 billion tokens per minute via Google model APIs. Source: Google I/O keynote, 19 May 2026.
Rising to $3.49tn in 2027. AI model spending alone: $32.6bn in 2026 → $59.2bn in 2027 (+81%). By end of 2026, up to 40% of enterprise applications will include AI agents. Source: Gartner, May 2026.
28 U.S. federal agencies hold AI contracts. DoD alone: $90.7B potential value, 1,319 contracts. Alphabet Q1 2026: $35.7B capex, majority for AI infrastructure. Source: Brookings, May 2026; Alphabet IR.
When QCE reduces model input by 50–99%, the financial saving scales with each organisation's actual AI spend. These are verified published figures, not projections.
Across loan processing, fraud detection, compliance review, trade surveillance and customer intelligence — all text-heavy, query-intensive workloads.
Source: JPMorgan Chase 2025 Annual Report; $17.6B total tech budget with $4-5B AI-explicitly stated.
Patient records, clinical notes, imaging reports and diagnostic history — massive, distributed, multi-source corpora with strict zero-retention requirements.
Source: NHS England Long Term Workforce Plan 2025; NHSE Digital AI Strategy 2026.
28 agencies with AI contracts. DoD: $90.7B potential AI contract value. Immigration, benefits, tax compliance, intelligence — all text-heavy LLM inference workloads.
Source: OMB FY2026 Budget Request; Brookings Institution AI contracts analysis, May 2026.
Research, M&A analysis, risk modelling, trade surveillance. Up to $750M–$1.45B illustrative.
Source: GS Technology Strategy 2025 Investor Day.
AML/KYC, credit risk, regulatory reporting. Up to €2.7B–€5.2B illustrative.
Source: ECB Financial Stability Review, AI in Banking, 2025.
At our observed 50–97% test range, up to $16B–$31.6B illustrative addressable saving globally in 2026.
Source: Gartner Worldwide AI Market Forecast, May 2026.
The opportunity framing: QCE is not competing for a share of the AI budget — it is positioned to reduce the largest single line item in that budget (inference token cost) for every organisation that runs LLMs on governed enterprise data. At scale, the total addressable saving exceeds the entire global AI market spend.
The QCE applies mathematical ideas developed through a long-running physics programme to budget-aware information selection, informed by decades of work in security and regulated environments.
The potential moat is the combination of a differentiated mathematical framework, a growing algorithm portfolio, implementation knowledge and domain configuration. Patents, independent benchmarks and customer deployment will determine how defensible that moat becomes.
The convergence of a physicist with 20 years of enterprise security experience, a novel mathematical framework, and the exact moment when enterprise AI costs become existential — this is not a coincidence. This is the right person, with the right technology, at the right time.
Enterprise corpora frequently exceed a model's usable context. The current evidence ladder spans public six-domain labels, real SEC material, 500K and 5M physical-token studies, and bounded represented pressure through 500M logical records.
QCE retained 91.5% of labelled evidence with an 80% declared reduction. Physical padding volume and independent semantic diversity are reported separately.
Sonnet 5, Gemini 3.8 Flash and Opus 5 advertised 1.0–1.048M contexts. The locked full 5M arms were ineligible before a provider call.
QCE retained 62.5% of labelled evidence at 500M→5M. This exposes both the scalable mechanics and the open extreme-compression quality frontier.
Hospitals, law firms, financial institutions and governments hold corpora far beyond a single model call. A reliable pre-inference selection layer could make more of that evidence economically usable.
Latest matched results are now available: 96.5% recall at 500K, 91.5% at physical 5M, and 62.5% at represented 500M. Performance is budget-dependent rather than assumed scale-invariant.
If quality is preserved, avoided model input grows with excluded tokens. Partner query volume, model pricing and realised retention will determine the actual economic value.
Physical 500K: 1,753 complete records; 500K estimated tokens; 256K budget. Physical 5M: 5,000 records; exactly 20M content characters; 1M budget. Represented 500M: the bounded physical 5M corpus with 100× declared record weights and a 5M budget.
Represented scale validates ranking and budget pressure; it does not claim physical ingestion of 500M distinct tokens. Genuine customer archives require the implemented asynchronous manifest/shard direction plus a measured sharded-versus-unsharded equivalence study.
Organisations — government, healthcare, finance, legal, defence — hold vast data repositories. A model should receive a traceable evidence budget rather than an unaffordable haystack. Weavi is building and measuring that control layer.
The LLM is not a search engine. It is an intelligence engine. Weavi-QCE gives it intelligence — not a haystack.
Send bounded records, documents or messages from one or many sources. A first-stage index remains necessary for genuinely unbounded archives.
Our algorithm scores every record across multiple mathematical dimensions simultaneously using our proprietary quantum-inspired framework.
We return the actual records — unmodified, unparaphrased. Your LLM gets the truth, not an interpretation of it.
Latest observed reductions span 48.8% at 500K, 80.0% at physical 5M and 99.0% under represented-500M pressure—with recall reported beside each.
Today: the Weavi QCE (Quantum Context Engine) is the intelligence layer — the most defensible position in the AI stack. Tomorrow: with a trained domain LLM, Weavi becomes the complete AI infrastructure for regulated industries.
Customer brings their own LLM. Weavi-QCE filters the context. Revenue: per-token API pricing. Gross margin: 85%+. No LLM cost to us.
Weavi trains domain-specific LLMs (legal, medical, financial, defence) optimised to work natively with the QCE. Customer gets a complete, pre-integrated AI stack.
Weavi is the default AI infrastructure for regulated industries. Data in → intelligence out. No third-party LLM dependency. Highest margin. Deepest moat.
Each enterprise deployment generates performance data that improves QCE routing. A domain LLM trained on QCE-selected, high-signal data is inherently better than one trained on raw data. The more customers, the better the models. The better the models, the more customers. This is a compounding data flywheel.
The reference application queries public OpenFDA adverse-event reports, runs a selected LLM over a baseline cohort and over QCE-selected records, then displays both responses and token usage side by side. It demonstrates a working integration—not clinical validation.
The endpoint returns the observed comparison for the chosen cohort, model and budget.
Same question and model, different supplied context. No claim of clinical correctness is implied.
The next step is blinded scoring against a prespecified factual and citation rubric.
The real SEC medium/large economics each had only one valid full-model pair. In the 90-query independent study, savings traded away recall. Latest 500K paired valid-call recall improved for Sonnet and Opus and fell 1.25 points for Gemini; pair counts were only 6–8.
The Evidence Lab uses real SEC EDGAR 10-K records; the OpenFDA reference app uses live public adverse-event cohorts. Paired retrieval, ranking, resampling, uncertainty and multiplicity diagnostics are retained where mathematically applicable. Test count is not itself proof.
This is not a claim. It is a reproducible, auditable, verifiable fact. We operate on a zero-knowledge proof principle: you don't have to take our word for it. We will show you the test, run it live, and you can verify every result yourself.
You give us your data. We run the algorithm. You see the token reduction, the quality scores, the statistical outputs — all calculated in front of you, with code you can inspect. You don't need to trust us. The mathematics speaks for itself.
In large-corpus SEC filing scenarios: significant token reduction (typically 70–90%+ on dense financial documents). Cost/doc falls materially at Claude Sonnet 5 rates. Speed and accuracy both improve — exact results vary by dataset and query type.
Large patient-record datasets are a high-value validation target. Deployment would require customer-side recall testing, privacy review, contractual controls and applicable healthcare compliance—not only an internal benchmark.
Multi-source claim datasets (police reports, medical records, witness statements, policy docs) see significant token reduction. Cross-referencing across sources improves signal quality — fraud indicators that would be missed in single-source review are surfaced. Adjuster throughput increases materially.
Large e-discovery corpora (1M+ emails/documents) benefit most — the QCE filters algorithmically before any LLM cost is incurred. Full corpus coverage replaces expensive sampling. Processing time reduces from weeks to hours. Results vary by corpus density and query specificity.
High-volume loan application processing benefits significantly from token reduction on dense regulatory documents. At scale (200K+ applications/year), savings compound materially at published LLM rates. Compliance review timelines reduce. Exact savings depend on LLM pricing, document density, and query volume.
Long conversation histories (thousands of messages) are where the QCE excels — selecting the relevant recent and historical context rather than sending everything. At high ticket volumes, per-ticket LLM cost falls dramatically. Exact savings depend on history length, LLM pricing, and query volume.
* Use-case financial figures are scenarios, not forecasts. Latest matched technical evidence observed 48.8% reduction with 96.5% recall at 500K and 80.0% reduction with 91.5% recall at physical 5M. Actual savings depend on model pricing, corpus density, query volume, selected budget and customer-side quality requirements.
25M+ token maintenance manuals are an ideal QCE use case — the corpus is vast, queries are specific, and the relevant answer is a tiny fraction of the total. Technician lookup time reduces dramatically. Engineer hours saved at scale. Safety risk reduced by surfacing the right procedure reliably.
150M+ tax returns represent a massive corpus where the QCE's scale advantage is most pronounced. Algorithmic pre-filtering at zero LLM cost, followed by semantic ranking, dramatically improves audit targeting precision. Hit rate improvement and cost reduction both compound at this scale.
A mid-tier bank processing 50,000 mortgage applications/year has 100M+ tokens of applications, valuations, credit reports, legal covenants and compliance documentation. QCE selects only the evidence relevant to the underwriting query — fraud signals, income verification, risk indicators — before any model call. Compliance review and credit-decision quality both improve.
27M+ tokens of pre-launch data (telemetry, inspection reports, historical launches, weather) is a high-value QCE use case. The relevant safety signals are a small fraction of the total corpus. Analysis time and engineering resource requirements both reduce materially. Data coverage improves.
Average enterprise user: 300M tokens (emails + documents). Standard AI context window covers <1%. With Weavi-QCE: full archive fits. "Only AI that can search your entire work history." 4x adoption vs standard tier.
Large case law corpora (500+ documents, 12M+ tokens) see the highest QCE savings — the relevant legal reasoning is a tiny fraction of the total. Single LLM pass replaces expensive multi-pass chunking. At scale across a legal practice, per-query cost falls dramatically. Results vary by corpus size and query specificity.
AI and data platforms: one integration could expose context selection across many customer workloads, subject to independent benchmark and partner economics.
Enterprise users: banks, healthcare networks, law firms, insurers and government teams with large, governed evidence corpora.
AI product companies: teams with recurring long-context inference can test whether selection improves gross margin without reducing answer quality.
The timing is attractive, but the strategy should be disciplined: secure partner-owned evidence, prove one high-value workflow, protect the IP and let measured customer economics determine the pace and valuation.
We are not relying on a top-down percentage of global AI spend. The first market is a focused set of regulated, document-heavy workflows where token cost and retrieval quality can be measured against an existing baseline.
A defensible bottom-up market model will follow paid pilots: verified annual query volume × realised saving × value-based share. Until those inputs exist, large top-down revenue claims are scenarios, not forecasts.
| Vertical | Why the QCE fits |
|---|---|
| Financial services | Long filings, auditability and high query volume |
| Healthcare | Longitudinal records and strict data handling |
| Legal & insurance | Large evidence corpora with measurable retrieval quality |
| Government & defence | Multi-source data, private deployment and corpus scale |
| Data & AI platforms | Distribution leverage across many enterprise workloads |
The relevant comparison is not a marketing percentage; it is recall, ranking quality, cost and latency under the same corpus, labels and token budget. Strong BM25, dense, hybrid and reranker baselines remain part of the confirmatory roadmap.
Introduces hallucination risk. Destroys original data fidelity. Not suitable for regulated industries. Loses the needle.
A comparable layer requires algorithms, bounded execution, provider integrations, labelled evaluation and accumulated failure learning. Customer-specific build cost requires diligence.
Continue paying to send all context where models can accept it—and build ad hoc truncation where they cannot. The avoidable amount depends on each workload.
The adversarial needle test is designed to make retrieval as hard as possible — needles buried deep, surrounded by highly similar decoy content. We ran this against every major alternative approach. The results are not close.
Keyword frequency models. They find documents that mention the right words — not documents that contain the answer. On adversarial tests with decoy content, they collapse. No semantic understanding. No relevance scoring. No quality awareness.
Assumes the answer is near the query in sequence. In real enterprise data — emails, records, filings — the needle can be anywhere. A fixed window misses it entirely on hard cases. No intelligence. Just position.
Recency + length + keyword match. Works on simple cases. Fails when the needle is old, short, or uses different vocabulary. The QCE scores across multiple quantum mechanical dimensions simultaneously — heuristics score one dimension at a time.
Historical exploratory comparison retained for context; it is not part of the current authoritative evidence set and should not support a superiority claim. Current diligence should use the frozen six-domain and three-scale artifacts.
We ran two distinct importance-needle scenarios across 500K, physical 5M and represented 500M corpora — using real multi-domain records, 20 planted targets at varying importance tiers, and adversarial hard negatives at maximum importance.
All 20 targets at importance = 1.0 (maximum). Hard negatives also at 1.0 but content actively denies the signal. When your metadata says it is critical — do we find it?
Same 20 targets but spread across tiers: 1.0, 0.75, 0.5 and 0.25. Hard negatives stay at 1.0. Do we find useful evidence even when it sits at a lower-priority tier?
Al = aligned · St = stratified. Each cell is one synthetic control case per scale. Latency: QCE at physical 500K–5M. QCE–QCE at physical 5M.
When upstream metadata is trustworthy, QCE finds every max-importance target in 3–9 ms at physical scales to 5M. Token savings remain above 98%. Ideal for critical-alert feeds, triage queues and any corpus where "importance" is genuinely signal-rich.
QCE through QCE returned 100% recall on both aligned and stratified at all three scales. They find evidence wherever it sits in the importance hierarchy — critical where relevance and priority metadata are not perfectly correlated.
Customers with reliable importance metadata can deploy the ultra-fast importance-first family. Customers needing evidence at any priority level use the query-aware family. Routing can blend both in a single request.
The QCE solves a universal problem: given a large set of items and a query, find the most relevant ones. This is not just an AI cost problem. It is the fundamental problem of every recommendation, discovery, and search system on the planet.
Each industry has different relevance signals. The current family spans v1–QCE; QCE adds query-aware lexical and corpus signals. Domain-specific weighting, semantic matching and held-out vertical validation are next-stage opportunities rather than completed claims.
The QCE selects the signal. A domain-trained LLM reasons over it. Together: a complete, private, auditable AI stack for regulated industries. No public LLM dependency. No data leaving the secure perimeter. Full-stack AI as a service.
The same mathematics that finds the right medical record finds the right film, the right song, the right product. Netflix recommendation. Spotify discovery. E-commerce search. Any system that needs to find the most relevant item in a large set.
Netflix has 15,000+ titles. A user query ("something like Inception but lighter") is a needle-in-haystack problem. The QCE scores every title across multiple relevance dimensions simultaneously — genre, tone, pacing, cast, viewer history — and returns the top candidates. The same quantum mechanical scoring framework. Different domain. Same mathematics.
Spotify has 100M+ tracks. "Something like this but more upbeat for a morning run" is a multi-dimensional relevance problem. BM25 and TF-IDF fail here — they're keyword models. The QCE scores across audio features, mood, tempo, listener context simultaneously. This is the same algorithm. Different data. Massive market.
Every major AI platform faces the same challenge: enterprise customers are hitting cost walls. Whoever acquires Weavi-QCE gains an immediate, validated competitive weapon.
The test programme strongly supports technical efficacy. It does not, by itself, establish product-market fit or a financing valuation. We separate a conventional venture valuation from the strategic value the technology could create for a large platform.
Working API, real-data testing, reproducible methodology and a broad algorithm portfolio reduce core technical risk.
Indicative pre-money range for a pre-revenue deep-tech financing; final pricing depends on team, patents, pilots and round terms.
Possible only where independently reproduced savings create exceptional acquirer-specific value or competitive tension.
Patent filings, customer-side validation, paid pilots and credible LOIs are the fastest route to a materially stronger position.
| Company | Acquirer | Price | Context | Year | vs Weavi-QCE |
|---|---|---|---|---|---|
| Inflection AI | Microsoft | $650M | Talent + IP licence. Team of ~70. Consumer AI product (Pi) with users. Structured as hiring + licence, not traditional M&A. | 2024 | Much larger team · consumer product · different deal structure |
| Adept AI | Amazon | $450M | Talent acquisition. Enterprise automation demos, early pilots. Team of ~100. | 2024 | Larger team · earlier-stage validation |
| DeepMind | £400M (~$650M) | 75 world-class researchers. Peer-reviewed publications. Demonstrated superhuman Atari performance. | 2014 | World-leading research team · published peer review | |
| Perceptio | Apple | ~$200M | On-device AI inference. Small team. No revenue. | 2015 | Comparable team size · less statistical validation |
| Weavi-QCE | TBD | Not yet priced | Working API. Real-data validated. Reproducible results. Pre-revenue, pre-patent. | 2026 | Technical evidence strong · commercial evidence next |
These are not projections. These are verified 2024–2026 published findings from the IEA, Goldman Sachs, Microsoft, Google and the EPA. The scale is staggering — and it is accelerating.
IEA Electricity 2024 Report; IEA AI & Energy Supplement 2024; Google I/O keynote May 2026.
Goldman Sachs "AI's Growing Water Footprint" 2025; Microsoft Sustainability Report 2024; Google Environmental Report 2023.
IEA 2024; DESNZ 2024 grid intensity; EPA eGRID 2023; EU CSRD regulation 2023/2775.
…is compute not run → energy not consumed → water not evaporated → carbon not emitted. QCE's token reduction maps directly to proportional compute avoidance on inference infrastructure.
3.2 × 10¹⁵ tokens/month × 50% QCE median reduction × 0.001 kWh/1K tokens (IEA 2024 inference estimate) = ~1.6 TWh saved per month at one company. Proportional water and carbon follow.
Energy reduction requires measured deployment telemetry (actual compute, cooling PUE, grid mix). The token saving is measured; the proportional resource saving requires a partner pilot with instrumented infrastructure to confirm.
At 50–80% typical input reduction, the proportional resource impact at enterprise scale is material. Below are illustrative estimates from published energy, water and carbon intensities — pending verified deployment telemetry.
At a large bank running 1B tokens/day with QCE 80% reduction and IEA's inference estimate of 0.001 kWh/1K tokens.
IEA AI Energy Supplement 2024 · pending deployment telemetry to confirm.
1.8L/kWh cooling ratio × energy saved from token reduction = proportional water avoidance.
Goldman Sachs AI Water Footprint 2025 · IEA cooling intensity estimate.
Token reduction per query is the auditable input metric. With deployment telemetry, it chains to energy and carbon for EU CSRD, UK TCFD and US SEC climate disclosures.
EU CSRD 2023/2775 · UK TCFD mandatory 2026 · SEC Climate Rule 2024.
QCE is not an ESG story with a technology attachment. It is a technology that inherently improves ESG metrics by solving the root problem: AI processes far more data than it needs. The resource savings are a direct consequence of the economic efficiency — not an overlay or offset scheme. That makes them genuine, auditable and repeatable.
Token and provider-cost reductions are measured. Energy and carbon are not assumed proportional: hardware, batching, caching, utilisation and grid mix require customer/provider telemetry.
Latest matched 500K Sonnet/Gemini/Opus calls. Pair counts 6–8.
With 91.5% labelled algorithm recall in the latest matched diagnostic.
Requires measured provider/customer compute and matched-quality completion.
A customer pilot should record successful-query quality, model/provider tokens, QCE compute, accelerator utilisation, energy and grid intensity. Only then should avoided-energy, carbon or CSRD reporting claims be calculated and independently reviewed.
Sustainability reporting increases demand for measured infrastructure efficiency. Weavi-QCE already records token economics; reportable energy or carbon reduction requires deployment telemetry, an agreed boundary and independent review.
Token savings do not automatically scale linearly into energy or carbon savings. A customer/provider pilot must establish the conversion and reporting boundary.
Think of a large insurance claim. Multiple organisations hold relevant data. Weavi-QCE ingests all of it, finds the relevant signal across every source, and delivers only that to the LLM.
Across all 5 sources combined — results vary by dataset
Cross-source signal quality improves significantly
More claims reviewed per day with better context
Gov, mil, edu, health, legal, finance
The LLM is not a database index. Weavi-QCE performs final question-aware evidence selection across the authorised candidate records it receives, so the LLM can focus on reasoning. Enterprise-wide discovery remains a connector/indexing layer.
Inflection AI and Adept AI were primarily talent and IP transactions involving much larger teams and materially different circumstances. They illustrate strategic appetite for AI capability, but they are not reliable valuation comparables for Weavi-QCE. Our value must be established through IP diligence, customer-side validation and partner-specific economics.
At pre-revenue stage, precise five-year ARR and valuation claims create false confidence. The operating plan is milestone-based; a forecast follows once pricing, sales cycle and expansion behaviour are observed.
| Period | Commercial proof | Product proof | Decision unlocked |
|---|---|---|---|
| 0–3 months | 3 design partners | Partner-owned corpus replication | Beachhead and success criteria |
| 3–6 months | First paid pilot | Production reliability and security scope | Initial pricing and deployment model |
| 6–12 months | 3–5 paid deployments | Repeatability across customers | Bottom-up forecast and seed/Series A plan |
| 12–24 months | Expansion revenue | Vertical tuning and platform integration | Scale direct sales or channel distribution |
Compute cost, support load, realised customer saving, gross margin and willingness to pay will be measured in pilots rather than assumed.
Customer engineering, production hardening, patents, security assurance and focused enterprise development — sequenced against evidence.
Publish a base, upside and downside model after the first paid deployments establish price, usage, sales cycle and retention inputs.
Sixteen registered variants from v1 through QCE cover different budget and selection behaviours. The unregistered QCE research candidate leads the latest 500K and physical-5M semantic tests. The moat is measured quality, execution evidence and accumulated failure knowledge—not variant count.
The portfolio reflects years of framework development and multiple algorithm variants. The duration and commercial value of that lead must be established through prior-art work, strong comparator benchmarks and deployment learning.
| Method | Status | Purpose |
|---|---|---|
| Full-context LLM | Implemented | Current end-to-end baseline where context permits |
| Importance controls | Implemented | Aligned and stratified importance-needle controls |
| BM25 / TF-IDF | Planned | Classical sparse-retrieval controls |
| Dense / hybrid RAG | Planned | Modern semantic production baseline |
| QCE ablations | Implemented | Preserve accepted and rejected candidate hypotheses |
With appropriate customer permission and privacy controls, deployment telemetry can improve routing and domain configuration. That learning loop is a potential future moat—not current traction.
Regulated industries — healthcare, finance, defence, government — require compliance before they can deploy. Weavi-QCE is designed for this from the ground up.
The bounded API is designed not to persist or content-log synchronous customer records. Data is transmitted to the selected deployment endpoint; customer-controlled/VPC execution is a future deployment option requiring validation.
Data minimisation and bounded processing support future healthcare deployment. HIPAA applicability, BAAs, risk analysis, operational controls and customer validation are required before any compliant-production claim.
Data-minimising processing can support purpose limitation. Controller/processor roles, lawful basis, DPAs, residency, deletion and deployment-specific assessment remain necessary.
Security, availability and confidentiality controls are being built toward an assurance scope. No SOC 2 Type II report exists yet; readiness and audit timing require an independent assessor.
Token reduction is measurable, but energy and carbon are not necessarily proportional because provider hardware, batching, caching, utilisation and grid mix vary. Customer/provider telemetry is required before any CSRD or avoided-emissions claim.
The evidence suite includes paired tests, intervals and multiplicity correction relevant to model-risk work. It has not been approved by a regulator or demonstrated as sufficient for Basel III/Solvency II compliance.
For classified environments (defence, nuclear, government): algorithm runs on customer's own infrastructure. No cloud dependency. No data leaves the secure perimeter. TS/SCI clearance pathway available.
The strongest next proof is customer-side validation. We are seeking strategic partners, design customers and investors who can help turn reproducible technical evidence into commercial evidence.
Investors respect honesty. Here are the real risks — and our mitigation for each.
Likelihood: Medium. OpenAI/Anthropic could attempt to build context optimisation. Mitigation: 18–24 month technical lead. Our validation methodology is peer-reviewable — theirs would not be. First-mover enterprise contracts create switching costs. And: if they're building it, they should buy it instead.
Likelihood: Certain — but irrelevant. Even Gemini's 2M token window covers only 0.6% of a typical enterprise user's data. Google processes 3.2 quadrillion tokens/month — and still wastes 78% of them. Larger windows = larger waste = larger savings from Weavi-QCE.
Likelihood: Current reality. We have £0 revenue and 0 paying customers. Mitigation: The API is live and the test results are reproducible and reviewable. The next milestone is independent customer-side validation; financing and strategic value should be discussed in that context.
Likelihood: High for VC path. Enterprise deals take 6–18 months. Mitigation: Platform partnerships (Databricks, Anthropic) bypass direct enterprise sales — one deal unlocks thousands of customers. Acquisition path avoids this entirely.
Likelihood: Current reality. No patents filed yet. Mitigation: Provisional patents can be filed within 30 days for approximately £8–25K. The algorithms are novel and patent-worthy. Trade secret protection applies in the interim. Filing immediately materially strengthens the IP position and valuation.
Likelihood: Current reality. Early-stage team. Mitigation: The technology is built and proven. Series A funding adds enterprise sales, legal, and engineering. Acquisition path integrates into acquirer's team immediately. Perceptio (acquired by Apple, ~$200M, 2015) was a similarly small team with novel IP and no revenue.
Every result has a create-once hash-locked artifact. Scale-specific caveats and negative outcomes remain part of diligence.
| Test scenario | n | Reduction | Recall | MAP | MRR |
|---|---|---|---|---|---|
| QCE · physical 500K→256K | 10 | 48.8% | 96.5% | .5387 | .6880 |
| QCE · physical 5M→1M | 10 | 80.0% | 91.5% | .3996 | .5599 |
| QCE · represented 500M→5M | 10 | 99.0% | 62.5% | .3918 | .5576 |
| QCE · six-domain locked validation | 210 | 52.9% | 87.1% | .6952 | .8069 |
| Gemini-facet QCE · discovery | 328 | 51.9% | 89.4% | .7263 | .8341 |
| GPT-facet QCE · discovery | 328 | 51.9% | 90.0% | .7053 | .8020 |
Arm-level outcomes include ablations and scale points; they are not independent samples. Precision uses incomplete public qrels. Partner-owned labels, strong external retrieval comparators and independent replication remain decisive next standards.
Weavi-QCE is a working, pre-revenue infrastructure asset with differentiated IP, a live API and promising internal evidence. The opportunity is meaningful; the remaining work—independent benchmarks, patents and customer proof—is explicit and measurable.
We are looking for a thoughtful partner who can bring a representative corpus, help define success criteria and evaluate the technology seriously. Strategic investment, licensing, a venture round or acquisition can follow the evidence rather than precede it.
© 2026 Weavi. Commercial in Confidence. E & OE. Internal benchmark observations are subject to diligence and independent replication. This document does not constitute a public offer of securities.