The problem
A recommendation you cannot audit is a liability.
- Article 50 in force
- 2 Aug 2026 EU AI Act Article 50 transparency obligations have applied since this date, and the Digital Omnibus did not move it. Disclosure is a legal duty, not a roadmap item.
- Annex III high-risk
- 2 Dec 2027 Annex III standalone high-risk obligations apply from this date — creditworthiness, insurance underwriting, access to essential services. Systems that cannot show their reasoning are not deployable in them.
- Signable black-box outputs
- 0 Black-box outputs a compliance officer can sign. An unexplained decision cannot be defended to a regulator, a board, or the person it was made about.
Explainability moved from procurement preference to filing requirement.
Mechanism
How Quorum decides
Three steps. Each one leaves a record.
-
01
Assemble the evidence first
Quorum pulls every source it will rely on — policy documents, ledger entries, prior cases, model outputs — and pins each one to a citation before a recommendation exists. Nothing is cited after the fact.
pinned sources · 14
-
02
Score the evidence and keep the dissent
Each source carries a weight and a confidence. Where two sources disagree, the disagreement is recorded and surfaced to the reviewer. Quorum does not average conflict away.
weighed sources · 14 conflicts · 2
-
03
Publish the decision and its conditions
The recommendation ships with a confidence figure, its three strongest supports, and a list of what would reverse it. A named human accepts, overrides or escalates, and the record stores which.
stated sources · 14 conflicts · 2 confidence · 0.86 would reverse · 3
The record is the product. The recommendation is what it explains.
Runtime
Ferro — the runtime that never leaves your building
Ferro is Quorum's on-device small-language-model runtime. Models execute inside your own infrastructure. No prompt, no document and no inference call crosses your network boundary. Privacy and speed by architecture, not by policy.
- Subprocessors in the decision path
- No egress Inference, indexing and logging all run inside the boundary. There is no subprocessor in the decision path to review.
- Median time to first token
- 240 ms Median time to first token on a single on-premise GPU node. No round trip to a hosted API, so latency is a hardware question you control.
- Reproducibility
- Version-pinned Every decision records the exact model build that produced it. An audit two years from now reproduces the same output from the same inputs.
Trust & compliance
The four questions that stop a purchase
-
Where does our data go?
Nowhere. Ferro runs the model inside your infrastructure, so documents and prompts stay behind your firewall. Quorum writes evidence pointers and decision records to your own database. There is no vendor-side inference, no training on your data, and no third-party subprocessor anywhere in the decision path. The deployment is yours to inspect.
artefact · network diagram + data-flow register
-
Can we show a regulator how a decision was made?
Yes. Every recommendation writes an immutable record: sources cited, weight given to each, confidence figure, model build, the reviewer who accepted or overrode it, and the timestamp. Export it as signed PDF or JSON. That record is the artefact an Article 50 disclosure request asks for, produced without a reconstruction exercise.
artefact · signed decision record (PDF · JSON)
-
What happens when Quorum is wrong?
It tells you first. Each recommendation lists the conditions that would reverse it, so a reviewer can test the weakest link before signing. Overrides are recorded with reasons and feed the next evidence review. Quorum never issues a decision. A named human does, and the record says who.
artefact · override log + reviewer attestation
-
When do the high-risk obligations actually bite?
Annex III standalone high-risk uses — creditworthiness assessment, insurance underwriting, access to essential services — carry their obligations from . Article 50 transparency has applied since 2 August 2026 and the Digital Omnibus did not move it. Product-embedded Annex I systems are dated 2 August 2028.
artefact · obligation timeline mapped to your deployment date
Proof
What changes once the reasoning is inspectable
- Meridian Mutual · consumer credit
- 9 days → 4 hours Adverse-action review cycle for declined applications. Same reviewers, same policy. The evidence chain arrived assembled instead of being rebuilt by hand.
- Kestrel Health · prior authorisation
- 41,000 decisions · 100% reproducible First twelve months on Quorum. Every recommendation carries a record that reproduces from its original inputs and model build.
- Northgate Legal · regulatory reporting
- 3 weeks removed Quarterly compliance report now assembles from decision records rather than interviews with the people who made the calls.
Illustrative outcomes for a fictional concept brand. Not real customers.
FAQ
Eight questions, answered plainly
No. Quorum assembles evidence, scores confidence and states what would change its mind. A named human accepts, overrides or escalates, and the record stores who did it, when, and against which model build. There is no confidence threshold above which Quorum acts alone, and no configuration that removes the human step.
Finance, insurance, healthcare and legal. Anywhere a decision has to be defended to a regulator, a board, a court, or the person it was made about. The reference deployments are adverse-action review, prior authorisation and regulatory reporting. It fits poorly where decisions are high-volume and never contested: the evidence chain costs more than the decision is worth.
Citations are attached after generation. Quorum pins its sources before a recommendation is formed, weights them, records disagreement between them, and publishes the conditions that would reverse the result. Order is the whole difference: a citation found afterwards proves a source exists, not that it was used. Conflicts are surfaced to the reviewer rather than averaged into one number.
No. Quorum can call a hosted model, and the evidence chain, confidence figure and decision record are identical either way. Ferro is for firms that cannot allow inference outside their boundary. Without it, your prompts and documents reach a third-party API, which becomes a subprocessor you have to disclose and review. Most regulated buyers choose Ferro for that reason.
The reference configuration is one GPU node per environment: 48 GB VRAM, 128 GB system memory, local NVMe for the evidence index. That configuration produces a 240 ms median time to first token. Ferro also runs CPU-only at roughly a fifth of that throughput, which suits batch review and does not suit interactive use.
No. Evidence indexes are built locally and stay local. No customer document, prompt or decision is used to train or fine-tune any shared model. Under Ferro there is no mechanism by which we could: the weights are version-pinned inside your infrastructure and nothing is transmitted outward. The contractual position and the architecture agree, which is the point.
Six to ten weeks to a first decision surface: two weeks of connector work, two weeks of evidence mapping, the remainder in reviewer calibration against decisions you have already made. The range is set by how many source systems the decision touches, and four or fewer lands near six weeks. Calibration needs roughly 200 historical decisions with their outcomes.
Article 50 transparency has applied since 2 August 2026; pre-existing generative systems have until 2 December 2026 to carry machine-readable marking. Annex III standalone high-risk obligations apply from 2 December 2027. Quorum supplies the logging, human-oversight and transparency evidence. Classifying your system and its conformity assessment remain yours; no vendor can certify you.
Question we have not answered? Put it in the briefing form below.
Next step
Request a briefing
Forty-five minutes. We take one of your live decision surfaces and map it to an evidence chain, on screen. No slideware.
- You name a decision your team has to defend.
- We map its evidence chain and its reversal conditions.
- You keep the map, whether or not you buy.