What you probably pictured
A telephone answering machine
Records a voice message when you are out, then plays it back. Its job is to remember what someone said.
Memory, not reasoning.
A working definition · Brugge, 2026
It answers questions. Not calls.
The Exact Answering Machine computes an answer from your own material — then hands over the exact derivation, the source behind every figure, and a re-run any third party can execute. When the answer is not there, it says so.
Decision recordSynthetic / C1
Is Harbourline Bank within its approved exposure limit?
Exposure€14.16m
Limit€12.00m
Of limit118%
exposure_minor_units ≤ limit_minor_units
A path, not a paraphrase
cp.harbourline has_position pos.harbourline.eugb31 of_instrument ins.eugb31 classified_under class.eu_government_bonds
record sha256:9efaed25…Replay ✓
Not a voicemail box.
Not a chatbot.
Not a chatbot.
01 / The name problem
Three machines. Only one shows its work.
The name is a coincidence, not a category. A telephone answering machine remembers a voice. A chatbot produces a plausible reply. This machine does a different job: it computes an answer that can be checked.
The current default
A chatbot
Generates a fluent reply from patterns in text. It is often useful. It cannot show which source forced the answer, and it rarely refuses.
Plausibility, not proof.
This site
The Exact Answering Machine
Computes the answer from your material. Shows the rows and the rule. Re-runs on demand. Refuses when the material has no answer.
Proof, or an honest refusal.
Working definition
An exact-answer engine returns the answer, the exact derivation that produced it, and a re-run any third party can execute — or it returns an explicit refusal that names what is missing.
02 / Why exactness matters
A wrong answer and a right answer look identical from the outside.
That is the silent failure. The answer is fluent, confident and wrong — and nothing in the interface says which. If a decision has to be defended afterwards, “it sounded right” is not a defence.
The same data. Three ways of reaching it. Only one can show the condition was applied.
In a controlled benchmark, the founder’s KU Leuven thesis queried one dataset three ways and changed nothing but the access method. Retrieval scored well enough to look credible. It failed exactly where the answer had to be exact: numeric thresholds, joins across records, and ordering in time.
Numeric range0.50 → 1.00Embedding retrieval → graph
Aggregation0.67 → 1.00Embedding retrieval → graph
Temporal order0.30 → 1.00Embedding retrieval → graph
One dataset, one changed variable. This is a result about the access method, not about a customer, a market or a product. Read the benchmark in the thesis, §7.5. Source and limits ↗
TF-IDF retrievalKeyword overlap
1.50/ 5
Embedding retrievalThe mainstream approach
3.47/ 5
Graph + structured queryThe approach EAM builds on
5.00/ 5
03 / The mechanism
Four steps. Then the step most systems skip.
The engine does not ask a language model what it thinks the answer is. It computes the answer from the material, records each step, and refuses when the material cannot support a conclusion.
-
01
Source kept
Your material goes in
Rate cards, contracts, price lists, stock records, internal rules — whatever the answer has to come from. Every item keeps its source, its date and its row identity.
-
02
Unknowns stay unknown
It becomes a structured map
Loose text becomes typed statements: this channel, this price, published by this source, in this year, on this unit. Fields that are absent stay absent; they are not filled with a guess.
-
03
No prompt in the answer path
The question becomes a condition
“Is Harbourline within its limit?” becomes an expression the material either satisfies or does not. There is no room in the middle for a plausible-sounding answer.
-
04
Replay, not explanation
The answer is computed, sealed and re-run
Every figure keeps a pointer to the row it came from. The record is hashed. An independent party can execute the same computation over the same pinned inputs and compare the result.
The fifth step
Refuse.
If a price is not published, the answer is “not published — this must be quoted”, not an estimate. If the reference data is stale, no conclusion is produced from it. If a rule does not cover the case, the engine names the missing rule and stops.
datasetlimits.ref
snapshot age173 hours
freshness limit24 hours
row-count drift−34%
No conclusion produced
04 / A record, opened
Read the answer. Then read why.
Three cases from the synthetic demo: one breach, one refusal because the data is stale, one refusal because the rule does not cover the case. None of them is a customer result.
Case C1 · fin.exposure_limit_checkSynthetic demo
Breach
Is Harbourline Bank within its approved exposure limit?
No. Exposure is €14,160,000 against a €12,000,000 limit — 118.00% of limit. The rule returns breach. A human approval is required before any action.
Rule applied
exposure_minor_units ≤ limit_minor_unitsrisk.exposure_limit v12 · observed 11,800 bp · threshold 10,000 bpDatasetpositions.eod · 12 rows
As of2026-10-01
Reason paths5 · all valid
Decision stateAwaiting approval
What the run recorded
- s1Fetch positions.eod as-of 2026-10-01 · 12 rows
snapshot b99268d1… - s2Materialise positions.eod into graph:ws_market_risk
- s3Aggregate exposure by counterparty · subject cp.harbourline → €14,160,000
- s4Evaluate risk.exposure_limit v12 → breach at 118.00%
- s5Compose the report the reader is looking at
No model is in this answer path. The figure comes from the rows; the verdict comes from the rule. The engine can show the exact edge chain that made the aggregation valid.
record_digest sha256:9efaed25…Replay: query-level · separate verifier re-derives 264 checks
Case C5 · fin.data_freshness_gateSynthetic demo
Blocked on data
Is the limits reference book fresh enough to use?
No conclusion was produced. The snapshot is 173 hours old against a 24-hour limit, and its row count is 34% below the trailing median. The engine stops rather than compute from stale reference data.
Gate applied
snapshot_age_hours ≤ 24 and abs(row_delta_pct) ≤ 5fin.data_freshness v7 · gate failed on freshness and row-count driftDatasetlimits.ref · 4 rows
As of2026-09-25
Retained evidences1, s2
ConclusionNone
What the run recorded
- s1Fetch limits.ref snapshots
- s2Gate freshness and row-count drift → fail
- —Stop. No rule is evaluated, no conclusion is written, no record is sealed.
A refusal is not an empty result. The failed gate, the numbers that failed it and the evidence already read all stay on the record. Nothing is quietly filled in to make the queue look complete.
record_digest not sealedFailure retained with evidence refs s1–s2
Case C6 · fin.exposure_limit_checkSynthetic demo
No applicable rule
Is Kestrel Securities within its approved exposure limit?
The engine cannot answer. Structured Credit is not enumerated by any published version of the rule, and it is 81.81% of this counterparty’s exposure — €9,000,000 of €11,000,000. No default class is assumed.
Rule that does not cover the case
exposure_minor_units ≤ limit_minor_unitsrisk.exposure_limit v12 · coverage incomplete · outcome not_applicableUncovered classclass.structured_credit
Uncovered share81.81%
Missing edgerule → applies_to → class
ConclusionNone
What the run recorded
- s1Fetch positions and reference data
- s2Materialise the graph
- s3Aggregate €11,000,000 · coverage incomplete
- s4Evaluate rule → not applicable; name the missing edge
- —Stop. No default class, no conclusion, no sealed record.
The missing edge is the finding. The engine does not silently classify Structured Credit as “other” and carry on. It reports the gap so a person can decide whether the rule should change.
record_digest not sealedFailure retained with evidence refs s1–s4
These are cases from the reference demo, not customer decisions. The figures are synthetic.
Open the full demo and run the replay yourself05 / Selected work
The work behind the claim.
Four pieces matter most: a running engine and its verifier, the graph work underneath it, a retail case designed around evidence gates, and a public research programme that uses the same discipline.
01Reference implementation
The engine, and its verifier
A deterministic core that seals a conclusion, replays it against pinned inputs, and lets a third party re-derive every check without trusting the author. It is a working reference implementation, not a packaged product.
- 264independent checks
- 7demo cases replayed
- 0models in the answer path
02Published
The graph underneath
KU Leuven master’s thesis: MLflow experiment metadata becomes an MLSO-aligned RDF graph, validated with SHACL and queried with SPARQL. The conversion package is published on PyPI.
- 7,831validated triples
- 68 / 78shapes / constraints
- 0SHACL violations
03Synthetic evidence
The retail case
A cultural-creative lifestyle retail pilot designed around operating gates: replenishment, product-file completeness, campaign readiness and the decision to expand. The mechanism has been exercised on synthetic data; no store has run it.
- 1design case
- 0stores run
- 0customers
04In formation
The research programme
Retail AI Research applies the same evidence discipline to public AI-in-retail studies: force a precise question, record what the system produced, state the limits, and return only what the evidence supports.
- Publicfield notes
- Noclients
- Openmethods
Also on the bench
Internal builds and side work
Kernel
Open source
EAM Zero
A Rust skeleton for decisions where the record may be wrong and execution may fail. Seven separate business objects, an append-only log.
Apache Ecosystem Atlas
One Apache Jena issue traced to its merged commit, with every graph edge pointing back to a public source.
Infrastructure
ProcedureGraph
An RDF/Oxigraph MCP service for recording which procedures work, under which conditions, and at what cost.
Product design
Modular desk pen holder
A parametric FreeCAD design, exported for suppliers and 3D printing.
06 / The honest status
What is built. What is measured. What is not.
A page about exactness should be exact about itself. This is the line between a working mechanism and a business.
Built
Can be run and inspected- Deterministic engine: canonical JSON, SHA-256 record digests, reason paths, replay and an independent verifier.
- A browser demo that runs the whole loop over synthetic data, including the refusals.
- A source-to-graph pipeline: RDF, MLSO, SHACL validation and SPARQL querying.
- Store-operations modules in the same core: stock, movements, campaign readiness and daily rule-based tasks.
Measured
Controlled or synthetic evidence- Thesis benchmark: 3.47 / 5 for embedding retrieval, 5.00 / 5 for graph + structured query, on one dataset.
- 7,831 triples, 68 NodeShapes / 78 property constraints, zero SHACL violations.
- 264 checks re-derived by the verifier; all seven demo cases replay; tampering is caught at the exact sequence.
- A Rust kernel with 122 tests and measured throughput of 3,723 records/s on this machine.
Not yet
Do not infer these from the page- A customer, a contract, a letter of intent, revenue or a validated price.
- Real operational data, or a store, buyer or auditor who has used it.
- A packaged product another person can install without the founder.
- Production multi-tenancy, identity integration or a run with real personal data.
What this page will not claim
No large language model was trained for this. The engine’s answers come from a constrained graph and written rules. The EU AI Act does not require knowledge graphs, ontologies or this architecture. The design is aligned with auditability; that is all. No saving, return or ROI is promised. The evidence on this page is about a mechanism, not a commercial outcome.
Start with one question
Bring the question your team answers by hand.
Not a slide deck. One question where an approximate answer would be expensive, and the material the answer has to come from. That is enough to test whether this is useful.
jason.jia@thirdhour.eu