An engine for the number, not for the picture.
Vemissure is being built as a quantification-and-uncertainty engine that sits above any satellite feed. It does not detect, and it does not image. It takes retrievals and produces a site-level constraint that carries a calibrated interval, a traceable uncertainty budget, and a manifest that lets anyone re-derive it.
Design principle
Sensor neutrality is a structural choice.
Vemissure owns no satellites — and will not.
Hardware ownership would destroy sensor neutrality and disqualify Vemissure from the verifier channel that carries roughly a third of the revenue plan. A company that owns an instrument has a commercial interest in that instrument's numbers looking good; a measurement layer that verifiers are expected to rely on cannot carry that conflict.
It also means the platform improves as the sensing market improves. New instruments, better retrievals and denser revisit all become inputs rather than threats.
Designed to sit above
Public feeds — Sentinel-2, Sentinel-5P, Landsat 8/9, EMIT — and commercial feeds bought per scene. These describe the classes of input the architecture targets. No existing integration, licence, endorsement or relationship is claimed.
Scope limit, stated before you find it
Vemissure does not produce a site's primary inventory.
Satellite detection floors sit well above most individual site emissions. Any material that implies otherwise would be wrong, and would deserve to be rejected.
What Vemissure will not claim
It will not claim to see everything a site emits, and it will not claim to replace an operator's bottom-up inventory. Overstating this is the single most likely cause of technical rejection, so the limit is published rather than buried.
Platform architecture
Six decisions. Two of them are the company.
Borrow, build-and-claim-nothing, build-proprietary, or buy. Every layer gets one of those four answers, and the discipline is to own only what is defensible.
The economics of not building things. Every commodity layer built in-house consumes capital without adding defensibility. Capital concentrated on layers 04 and 05 buys the only differentiation that compounds.
01 — Satellite feeds
Sentinel-2, Sentinel-5P, Landsat 8/9 and EMIT are free at the point of use; commercial feeds are bought per scene. Sensor-neutral by design — no feed is privileged.
- Public retrievals as the default working set
- Commercial feeds where revisit or sensitivity justifies the cost
- Operator site data and bottom-up inventory as the reconciliation counterpart
- Controlled-release and operational observations feeding the calibration corpus
02 — Retrieval, masking, flux estimation
Implemented from published methods. Vemissure claims no novelty here, and says so: plume detection and quantification have been open-sourced on free public feeds, and presenting them as proprietary would fail an endorsing body's first filter.[9]
- Scene acquisition, geolocation, background and surface-artefact masking
- Plume mass-balance flux estimation from published approaches
- An independent production implementation — open-source models used only as a benchmark, under their non-commercial licence terms
- Provenance captured from ingest: the first entry in the reproducibility ledger
The rate is the input to the Vemissure contribution. On its own it is not the product.
03 — Wind fields
ERA5 and ECMWF ensemble members. Producing meteorology is not the novelty — propagating its spread into the emission interval is.
- Ensemble members treated as a distribution, not a single chosen field
- Meteorological spread carried through to the reported interval (I-7)
- Wind treated as an explicit contribution in the uncertainty budget, not an absorbed residual
04 — Calibration & uncertainty
This is the company. Innovation register entries I-1, I-2 and I-3.
- I-1 Coverage-calibrated uncertainty — distribution-free (conformal) interval calibration against a controlled-release corpus, with scene-conditional covariate adjustment, so a stated 95% interval contains the truth 95% of the time
- I-2 Metrological traceability — a GUM-form Level-4 uncertainty budget with SI traceability via a national measurement institute[19]
- I-3 Cross-instrument transfer standard — harmonises response across sensors so multi-year trends are real rather than sensor artefacts
- Supporting: wind ensemble propagation (I-7), below-detection-limit inference (I-8), intermittency handling (I-9)
Method basis: conformal prediction for finite-sample coverage[20] applied on top of the physical quantification, not in place of it.
05 — Reconciliation & evidence
This is the sellable artefact. Innovation register entries I-4 and I-10.
- I-4 Bayesian reconciliation — a posterior combining top-down and bottom-up, reported with a residual and a credible interval
- I-10 Reproducibility ledger — any published figure re-derivable bit-identically from its manifest, years later
- The Quantification & Reconciliation Evidence Pack as the delivered artefact
- A Verifier Edition designed to sit inside an accredited body's own assurance product (I-6)
06 — Identity, billing, infrastructure
No differentiation is available here, so none is attempted. Bought off the shelf and kept boring.
- Identity and access management
- Billing and subscription management
- Cloud infrastructure, with UK/EEA data residency as a customer-driven requirement
Method
What "calibrated uncertainty" actually means.
Three ideas, in plain terms first and technical terms second.
01Coverage, not confidence labels
In plain terms: if a system says "95% confident", then about 95 out of every 100 such intervals should actually contain the truth. That is a testable claim, and it is the claim Vemissure intends to be held to.
Technically: distribution-free conformal calibration against a controlled-release corpus, with scene-conditional covariate adjustment, targeting empirical coverage that matches nominal coverage across the operating envelope.[20] The pass criterion is explicit: ≥90% empirical coverage of nominal 95% intervals on held-out controlled-release data.
02Traceability in the metrological sense
In plain terms: you can follow the number backwards. Every step from raw observation to reported figure is recorded, and every source of error is written down rather than absorbed into a single opaque margin.
Technically: a GUM-form Level-4 uncertainty budget[19] enumerating each contribution — retrieval, background, wind ensemble, intermittency, spatial attribution, temporal aggregation — with SI traceability via a national measurement institute. Traceability is what converts a number into something a third party can audit rather than accept.
03Reproducibility as a deliverable
In plain terms: the same inputs produce the same answer, later, on request, by someone else. If the figure is challenged in five years, it can be regenerated exactly as it was produced.
Technically: a reproducibility ledger pinning input scene identifiers, retrieval versions, meteorological ensemble members, parameter sets, code versions and human decisions to a manifest that travels with the evidence pack — targeting bit-identical re-derivation.
Nominal vs. observed coverage
Blind controlled-release testing
Intervals reported at a nominal 95% confidence achieved roughly 70% empirical coverage in blind testing, falling to 52% in the second stage of the same campaign.[6][14] The R0 gate sets ≥90% as the pass criterion for Vemissure.
Honest limits
Calibration does not make an interval narrow. Sometimes an honest interval is wider than the one it replaces. That is the point: a wide interval that behaves correctly is more useful in a reconciliation than a narrow one that does not.
What Vemissure does not claim
No claim that Vemissure detects methane better than existing systems, that its quantification core is more accurate than published methods, or that any calibration result has yet been produced or validated. The R0 coverage benchmark exists to settle that question, and it either passes or it does not.
The one metric
Median interval width per site at fixed confidence. If the calibration corpus is working, it falls monotonically. That is the measurable proof that the moat compounds.
Innovation register
Eleven entries, three filters.
Each register item is tested against three conjunctive filters — F1 is it defensible? F2 is it load-bearing for revenue? F3 is it delivered in-house? The register is organised by role, not by marketing weight.
Core innovation
The calibrated number and its provenance. This is the company.
I-1Coverage-calibrated uncertainty
Distribution-free (conformal) interval calibration against a controlled-release corpus, with scene-conditional covariate adjustment, so a stated 95% interval actually contains the truth 95% of the time.
F1 ✓ F2 ✓ F3 ✓ · testable at gate R0
I-2Metrological traceability
A GUM-form Level-4 uncertainty budget with SI traceability via a national measurement institute.[19]
F1 ✓ F2 ✓ F3 ✓
I-10Reproducibility ledger
Any published figure re-derivable bit-identically from its manifest, years later. Reproducibility is what turns a measurement into evidence — and it is why historical continuity stays with Vemissure.
F1 ✓ F2 ✓ F3 ✓
Supporting technology
The components that make a calibrated interval achievable in real conditions.
I-3Cross-instrument transfer standard
Harmonises response across sensors so that multi-year trends are real, not sensor artefacts. Cross-instrument response slopes measured at a single controlled-release facility have spanned 0.09 to 1.13 — the same release, differing by more than a factor of ten.[15]
F1 ✓ F2 ✓ F3 ✓
I-4Bayesian reconciliation
A posterior combining top-down measurement and bottom-up inventory, reported with a residual and a credible interval — so the reconciliation conclusion carries its own uncertainty rather than hiding it.
F1 ✓ F2 ✓ F3 ✓
I-5Calibration corpus
The compounding data asset: a growing library of controlled-release and operational tuples against which intervals are calibrated. Every validated site tightens the interval for every subsequent site.
F1 ✓ F2 ✓ F3 ✓
I-7Wind ensemble propagation
Wind is among the largest error sources in satellite flux inference. ERA5 and ECMWF ensemble members are borrowed; the contribution is propagating their spread into the reported interval.
Supporting component of the interval
I-8Below-detection-limit inference
A non-detection carries information. Given detection floors well above typical site emissions, handling non-detection explicitly avoids the systematic bias of treating "not seen" as "not emitted".
Supporting component of the interval
I-9Intermittency handling
Emissions are episodic and overpasses are sparse. Inferring a reporting-period figure from intermittent observation needs its own treatment and its own contribution to the budget.
Supporting component of the interval
Commercial & workflow innovation
How a defensible number reaches the people who need it.
I-6Verifier-embedded licence
The engine licensed inside an accredited body's own assurance product, reaching many operators through one programme relationship rather than one sales process per operator.
F1 ✓ F2 ✓ F3 ✓
I-11Assurance SLAs
Contractual commitments on turnaround, reproducibility and defensibility support — built for assurance timetables, where an evidence pack must exist by a date and survive review.
Supporting
What was dropped, and why
An earlier version of the plan claimed novelty in plume-flux inversion, background-artefact discrimination and multi-source reconciliation. Detection and quantification have since been open-sourced on free public feeds,[9] and the third is closer to an open protocol than a proprietary method. Those claims were withdrawn rather than defended. What remains is narrower and stronger: the calibrated number and its provenance — not a model and not a feature.
Status. Every register entry describes planned capability. None is presented as operational, granted, tested or independently validated.
The product
Quantification & Reconciliation Evidence Pack.
The MVP thesis: the smallest thing a verifier will pay for is a per-site evidence pack for one reporting period. Not an API key. Not a dashboard. Not an alert. A signed-off-able artefact — one site, one year.
Planned contents
Planned MVP scope, targeted at R2 (Month 14). Subject to change as design-partner work proceeds.
The artefact, not the access
The product is not data. It is an assurance input — the artefact a verifier attaches to an Article 12(3) reconciliation and an Article 8 verification statement.[4] Everything in the plan follows from that distinction.
Evidence Pack — structure
Vemissure supplies a measurement basis and supporting evidence. It does not verify, certify or assure emissions.
A limitations section is included by design. A measurement product that cannot state where it does not apply is not a measurement product.
Technical questions
Straight answers for technical readers.
QIs Vemissure an AI company?
No. The work is measurement science: physical flow inference, uncertainty quantification and metrology. Statistical methods — conformal prediction among them — are tools where they earn their place, not the identity of the company.
QDo you own or plan to own satellites?
No. Hardware ownership would destroy sensor neutrality and disqualify Vemissure from the verifier channel. It is a design decision, not a limitation.
QDoes Vemissure verify or certify emissions?
No. The operator carries the Article 12(3) obligation; the accredited verifier carries the ISO/IEC 17029 accreditation.[22] Vemissure supplies an input to both and carries neither. It holds no licence, permit or accreditation of its own.
QIf detection is open-source, what stops anyone doing this?
In the short run, nothing but the work. In the medium run, the calibration corpus: a competitor can copy a method, but a library of validated controlled-release and operational tuples accumulates site by site. The honest read is that within roughly thirty-six months quantification will be a commodity and traceability will approach a public good. The defensible position exists in the window between those two facts — and the plan says so rather than dressing it up as permanence.
QHow would the calibration claim be tested?
Gate R0, at Month 4, on published controlled-release datasets: ≥90% empirical coverage of nominal 95% intervals. It costs founder time and no capital, and it either validates or kills the central claim before a facility is booked. If it fails, the documented response is a pivot to a reconciliation-only product — a smaller business, but a real one.
QWhat exists today?
Nothing operational. Vemissure is at pre-endorsement stage and every capability described on this site is planned. No customer, revenue, deployment, accreditation, partnership or granted patent is claimed.
Start a conversation
Build a defensible measurement basis for methane MRV.
Talk to us about operator workflows, verifier programmes, technical validation or strategic partnerships.