Synapse Photonics Get in touch

Sophia Antipolis, France · Semantic encoding for safety-critical communication

Meaning that can be checked.

Synapse Photonics builds a semantic encoder for safety-critical communication. It reads two messages and tells you whether they mean the same thing — and if not, exactly which part disagreed. No training data. No model weights. Every answer traces to a rule you can read.

Standardised message formats Explainable by construction Runs inside your infrastructure
0
dangerous confusions across five standardised domains, one mechanism, no re-tuning between them

A single threshold separates safe paraphrases from operationally opposite instructions with no overlap. “Cleared for takeoff” and “cleared to land” are never confused. Neither are “alter course to port” and “alter course to starboard”, nor “legal advice” and “business advice” on a privilege log, nor 10 mg and 100 mg on a medication order.

Measured Five domain packs · ICAO · IMO SMCP · FRCP 26(b)(5) · ISO 15022 · HL7 FHIR
1.000AUC

Perfect separation in every one of five standardised domains, with no re-tuning between them.

Measured
100%

Routing accuracy with 30% of components dead. 1,000 independent trials.

Measured
86–142concepts

The size of a domain pack. The engine is fixed; the curated vocabulary is what we build.

Measured
5

Domain packs complete — aviation, maritime, legal privilege, corporate actions, clinical medication orders.

Built

The problem

Readback is checked by ear, or not at all.

In air traffic control and vessel traffic services, instructions are read back to confirm they were heard correctly. Readback error is a documented cause of accidents. There is no automated check, because both available options fail.

String Exact text matching is too strict. “Cleared for takeoff” and “cleared takeoff” are the same instruction and different strings.
Model General-purpose AI cannot be certified. Safety software must be explainable. A neural network is not.
Rule We compare meaning against a rule you can read. Two messages in; a similarity score and the specific disagreement out.

Where it works

Wherever a standardised message format already exists.

That is the whole qualification. Not “bounded domains” — the test is whether the utterances are standardised, not merely the subject matter.

DomainStandardAUCMargin
Air traffic phraseology ICAO 1.000 +0.063
Marine communication IMO SMCP 1.000 +0.101
Privilege log entries FRCP 26(b)(5) 1.000 clean
Corporate action messages ISO 15022 / 20022 1.000 +0.430
Medication orders Structured / HL7 FHIR 1.000 +0.120
The same content as free prose 0.04–0.875 overlap

The form is the requirement, and we state it plainly because it decides whether we can help you. We tested the obvious objection — that prose fails only because our dictionary doesn’t cover it. Closing the vocabulary until out-of-vocabulary words fell from 53% to 2% moved the score by 0.021. Coverage is not the constraint. The standardised form is.

AUC separates safe paraphrases from operationally opposite instructions; margin is the gap between the two classes, and positive means no overlap at any threshold. Maritime ground truth verified against COLREGs Rules 16 and 17; privilege against FRCP 26(b)(3) and Upjohn; corporate actions against published ISO 15022 event codes; medication error classes against ISMP and WHO patient-safety material. Not asserted by us.

How we test ourselves

We publish the predictions we got wrong.

The fourth domain was a prediction written down before the test was run: ISO-formatted corporate actions would separate cleanly, press releases would not. We then built the pack and measured it. Both halves held. The fifth domain failed that prediction on first run — and we publish that too.

3 Defects the medication-order pack exposed in our own encoder — including one that made two opposite altitude instructions score identical. All three are fixed.
5 / 5 Domains re-measured on the corrected encoder rather than carried forward. The figures above are the new ones.
Identical Fixes are not domain-specific patches. We change one domain’s vocabulary and re-run the others unchanged; they return bit-identical.

Fault tolerance

100% routing accuracy with 30% of components dead.

Across 1,000 trials. This is a property of how the encoding is built, not an error-correction layer bolted on. A missing element leaves a known gap rather than a wrong value, so losses are located rather than silent — which is the difference between a degraded system and an untrustworthy one.

Measured 1,000 independent trials · reproduced independently from source

How it fits your system

A component, not a platform.

Two messages in; a similarity score and a specific disagreement out. Integration is weeks, not years.

Input A Source message the instruction as issued, via your speech-to-text
Input B Readback the instruction as repeated, via your speech-to-text
SYNAPSE Semantic encoder · rule-traceable
Output Similarity score one threshold, no overlap between safe and opposite
Output The part that disagreed named explicitly, traced to a rule you can read

↓ into your existing alert display

Attach point

The structured artefact, not the text around it

We attach to the radio call, the log entry, the ISO message, the medication order — the part of your workflow that already has a standard form. Not to the free-text documents surrounding it.

Deployment

Embedded library, network appliance, or air-gapped module

Nothing leaves your infrastructure. The architecture is designed for on-device, isolated and air-gapped operation from the start, rather than as a deployment option added afterwards.

Output

A score, and the part that disagreed

Every answer traces to a rule you can read, so the result can be reviewed, argued with, and certified — which is what safety software requires and what a neural network cannot offer.

Domain packs

The engine is fixed. The vocabulary is the deliverable.

Five packs complete

Aviation, maritime, legal privilege, corporate actions and clinical medication orders — each a curated vocabulary of 86 to 142 concepts, validated against that industry’s own standard.

A pack is small on purpose

The corporate-actions pack is 86 concepts and took one working session, because the vocabulary was already written down by ISO. We did not invent it — we encoded it.

One mechanism, no re-tuning

The same encoder produced every result in the table above. Nothing is tuned per domain, and we verify that by re-running earlier domains unchanged after each new pack.

What qualifies next

Any workflow with an existing message standard: railway signalling, emergency dispatch codes, military voice procedure.

One question qualifies a prospect: does a standardised message format already exist in your workflow? If it does, that format is where we sit. If it does not, we will say so.

Photonic roadmap

Specified and simulated. Not built.

Design-point figures are derived from a real foundry silicon-nitride process design kit, not from assumptions — ring footprint, group index, propagation loss and minimum bend radius are all read from the PDK. The design is built to the foundry’s minimum bend radius rather than requiring the foundry to change it.

These are first-principles projections. No silicon has been fabricated and no photonic measurement has been taken. Validation against fabricated hardware is a future milestone.

80 µm Ring radius on a 300 nm silicon-nitride process, at Q ≈ 36,000.
19,200 Concept addresses across a 20 × 20 cell array and 8 wavelength channels.
643 mm² Die area at that design point. Projected

What we don’t claim

The limits, stated before you ask.

  • Free prose is outside our operating condition. If your workflow has no standardised artefact, we are the wrong supplier and we will tell you so.
  • An 8-channel chip covers European grammar. The six further channels needed for non-European languages cost die area.
  • A chip addressing a full general lexicon does not fit one reticle at this bend radius. Neither limit affects the bounded-domain product, which needs 86–142 concepts.
  • Test pairs are written by us, not yet by your domain experts. That is the first thing we would want to change in any engagement.
  • Low optical loss is load-bearing for the photonic roadmap, and is a process parameter we do not yet hold.

Why a French company matters

Built for deployments where data cannot leave the device.

Based in Sophia Antipolis. European sovereignty requirements favour European suppliers for defence and critical infrastructure, and our architecture is designed for on-device, isolated and air-gapped operation from the start rather than as a deployment option.

Synapse Photonics SAS 535 Route des Lucioles · 06560 Valbonne SIREN 106 579 493 R.C.S. Grasse Every figure on this page traces to a re-runnable script