Exact Answering MachineSources & limits Live demo

Reference · not a product claim

Sources & limits.

Every number on the main page is listed here with the artefact it came from and the boundary it must not cross. If a figure cannot be traced to a primary source, it does not belong on the page.

Last verified 15 November 2026Single-page ledgerCorrections welcome

01 / The rule

Figures are governed, not copied.

The project keeps one fact register as its single source of truth. Every figure below is checked against a primary artefact — the thesis text, the demo bundle, the verifier output, or the measured build notes — rather than against an earlier summary. Where two values exist inside one document, the convention is stated explicitly.

This page exists because an exact-answer project should be exact about its own evidence. A number with no boundary is a claim, and a claim with no artefact is marketing.

02 / The ledger

What each figure establishes.

FigureValuePrimary sourceWhat it does and does not establish
Benchmark totals 1.50 · 3.47 · 5.00 KU Leuven thesis §7.5; fact register §2.2 On one dataset, changing only the access method, graph + structured query scored 5.00/5 against 3.47/5 for embedding retrieval and 1.50/5 for TF-IDF. It does not establish performance on customer data, a market result, or superiority on every question class.
Question-class scores 0.50→1.00 · 0.67→1.00 · 0.30→1.00 KU Leuven thesis §7.5; fact register §2.2 The three columns shown are numeric range (Q2), aggregation (Q4) and temporal ordering (Q5), embedding retrieval → graph. They show where retrieval loses the condition, not that retrieval is useless.
Thesis graph 7,831 triples Thesis results table; fact register §2.1 The page uses the thesis-table basis. The abstract, §6.5 and the conclusion carry 7,912. The two are never mixed inside one document. 7,831 is a graph of MLflow experiment metadata, not a customer graph.
SHACL validation 68 / 78 · 0 violations Thesis; fact register §2.1 68 NodeShapes and 78 property constraints, with zero violations on the thesis graph. This is validation against the project’s shapes, not a certification by an external body.
Query time < 100 ms Thesis; fact register §2.1 Standard aggregation queries on the thesis graph using Oxigraph. It is not a latency guarantee for other graphs, other hardware or production load.
Verifier 264 checks keel README and verifier output on the demo bundle An independent verifier re-derives 264 checks from the bundle: audit chain, record digests, reason paths including abstention evidence, rule outcomes and replay consistency. It is a local reference implementation, not a third-party audit or a certification.
Replayed cases 7 of 7 Demo bundle; keel README All seven synthetic demo cases re-execute against pinned inputs. It shows the record is reproducible, not that the business case works.
Breach record €14.16m / €12.00m / 118% Demo bundle, case C1 Harbourline Bank, synthetic: exposure €14,160,000 against a €12,000,000 limit, 118.00%, rule risk.exposure_limit v12. Not a customer decision and not a real bank.
Stale-data refusal 173 h / 24 h / −34% Demo bundle, case C5 limits.ref snapshot is 173 hours old against a 24-hour gate; row count is 34% below the trailing median. The engine produced no conclusion. Synthetic.
Missing-rule refusal 81.81% · €9.0m / €11.0m Demo bundle, case C6 Structured Credit is not enumerated by the published rule and is 81.81% of Kestrel Securities’ exposure. The missing edge is named; no default class is assumed. Synthetic.
Rust kernel 122 tests · 3,723 rec/s EAM Zero README, measured on an Apple M4, 10 cores, 16 GB 122 workspace tests; ingest rate at 2,000 records. 0.106 ms per proposal at 20,000 records. A deliberately model-free skeleton with dummy belief and policy implementations, not the product and not a trained model.
ProcedureGraph 3,156 LOC · 3,208 quads Fact register §6, re-measured 2026-10-02 TypeScript lines in src/ across 10 files; quads in the trade graph at that date. It grows as outcomes are recorded. Internal tooling, not a product.
Apache Ecosystem Atlas v0.1 · one case Apache Ecosystem Atlas README One Apache Jena issue (#4202) traced to its merged pull request and commit, with an RDF export and an RML mapping. A limited snapshot, not a complete project history and not affiliated with the Apache Software Foundation.
Pen holder 90 × 70 × 115 mm Specification modular_pen_holder_v1.md; FreeCAD model and exports A parametric desktop organiser with 2.4 mm walls, a 3.0 mm base and an R4 edge, designed for PLA/PETG FDM prototypes. It is not an injection-moulding production design and has not been factory-qualified.

The retail working package is a design and planning case. No store has run it, no customer has bought it, and no operating result is claimed. See the retail boundary below.

03 / The boundaries

What the page refuses to imply.

Six boundaries that do not move

  • No customer result. Zero customers, zero contracts, zero letters of intent, zero revenue and no validated price. Every operational example is synthetic.
  • No trained language model. The engine’s answers come from a constrained graph and written rules. The models elsewhere in the founder’s work are small task-specific models, not LLMs, and they are not in this answer path.
  • No legal requirement. The EU AI Act does not require knowledge graphs, ontologies, RDF or this architecture. The design is aligned with auditability; that is the whole claim.
  • No performance promise. No saving, ROI, uplift, accuracy rate or payback is claimed. The benchmark is a controlled result about one access method.
  • No compliance certificate. The verifier checks the record’s internal consistency and reproducibility. It is not a regulatory audit, a legal opinion or a certification.
  • No finished product. The engine is a working reference implementation. Packaging, onboarding and production multi-tenancy are build items, not shipped facts.

The retail case is a design, not a result

The retail working package sets out a pilot design and the gates that would have to pass before expansion. The engine has been exercised on synthetic data. No store has run the design, no customer has bought it, no supplier has signed it, and no revenue, footfall, margin or expansion result is claimed. Read it as a plan for producing evidence — not as evidence that the plan works.

One name, two spellings

The project appears in earlier documents as both Exact Answering Machine and Exact Answer Machine. They refer to the same thing. The longer form is used on this site because it is the founder’s current wording; neither form refers to a telephone.

Corrections

If a primary artefact changes, this ledger is updated before the main page. If a figure here is wrong, the correction should name the artefact that supersedes it, not merely replace the number. Contact jason.jia@thirdhour.eu.

04 / Go and check

The primary artefacts.

← Back to the main page