Every passage, trip, and charge a machine makes in a regulated public space is a decision. Today nothing checks that authority before the action, and the operator pays for every wrong answer afterwards.
A free-flow gantry reads a plate, classifies the vehicle, applies a tariff, and posts a charge. A dispatch system releases a personal delivery device onto a sidewalk. In both cases the operator's system acts first and the question of whether the action was within the operator's authority is answered later, by a customer-service queue, an audit, or a court. Brazil now prices that gap in law: an improper toll charge is refunded at double within seven calendar days, by the concessionaire. US cities price it per violation. Insurers price it in the policy.
The problem is not "AI." It is delegation: a machine acting or paying under authority granted earlier. Autonomy raises the stakes, but the gap exists today with a human-driven car at a gantry.
What the operator pays for a wrong decision, today
- An improper free-flow charge is refunded at double within 7 calendar days by the concessionaire (Resolução ANTT 6.079/2026 art. 63-G; 19º Termo Aditivo to the RioSP concession, cl. 5.3, June 2026).
- An unpaid passage is 10% of the concessionaire's loss with both notifications on an auditable record, and 100% without it (19º TA cl. 5.2.1, cl. 5.3 I and III).
- As context: 1,324,797 free-flow non-payment fines cancelled between 29 Apr and 24 Aug 2026 after the underlying tariff was regularized, about 35% of the 3.8 million such fines issued nationally to 2 Sep 2026 (Ministério dos Transportes, 24 Aug 2026; Valor, 2 Sep 2026, citing the ministry).
Every point above is priced after the action; a determination in front of it holds or refuses the charge before the money moves. The Maryland duplicate-charge review (82,847 customers, 62 of 65 sampled confirmed) and the Pennsylvania device penalties are in the sources index, rows S4, S7 and S8.
What we do
In card payments, an authorization decision sits in front of every settlement, and the decision itself is priced: US networks charge about two cents per authorization request, approved or declined (Visa and Mastercard network fees as published in a US payment processor's fee schedule, retrieved September 2026). Nothing equivalent sits in front of a toll charge, a per-trip fee, a curb reservation, or a sidewalk dispatch.
Observer Protocol is that authorization step for machines acting under public rule. Before the operator's own system posts the charge or releases the move, it receives allow, refuse, or undetermined against the operator's delegated authority and the public floor, with the clauses checked.
We do not control the machine. We determine whether an action is within its authority, and the operator's own systems enforce that determination before the machine moves or pays.
Where the rule does not settle the answer, we say so instead of picking one, so the operator is never exposed on a reading the regulator did not make. Every determination can be re-derived by the driver, the concessionaire, the city, or the regulator on their own machine, without the operator and without us.
Where it sits
We do not touch the vehicle's control loop. Braking, steering, and obstacle handling are the operator's safety stack and the safety-case world, and we stay out of both. The layer sits at points the operator's own software already controls:
1. Before it pays. Toll charge, fare, curb fee, charging session. The gantry's read, the classification, the tariff, the exemption, and the account state are inputs; the determination sits between them and the charge, where authorization sits in card payments. In Brazil this is the point where a wrong charge becomes a double refund in seven days.
2. Before it moves. The dispatch layer releases each mission or route only with a determination against the permit and the public floor. This is the point where a wrong move becomes a violation, a suspended permit, or a smaller fleet.
3. Where it may go. The zones, hours, and speed limits operators already implement by hand from their permits are compiled from the same register, so every constraint traces to the rule that created it.
4. When it is contested. A disputed charge or a reported violation is judged against what the machine was authorized to do at that moment.
A determination has two errors and they are not symmetric. A wrong allow is the refund, the fine, or the suspension. A wrong refuse is a lost passage or trip. In payments, wrongful declines cost 1.51% of e-commerce sales, more than fraud (Datos Insights, May 2024, from a 2023 survey of US and UK merchants). We report both error rates on every corpus: the reason this has to run at determination time, and not be reconstructed in an audit.
Where the operator is exposed even when it does everything right
The clearest case is a runtime fact, not an ambiguity: under 19º TA cl. 5.2.1, an unpaid passage is 10% the concessionaire's loss if both notifications are on an auditable record at the time of notification, and 100% if they are not. On the BR-101 Rio-Santos corpus of 10,000 synthetic passages, 73 of 318 unpaid passages sit at 100% for want of that record, against 11 at 10%.
The rest is textual ambiguity, converted clause by clause from a national regulation, a concession contract, and two amendments:
| Open point | Where the texts disagree or fall silent | On the corpus |
|---|---|---|
| Whether the 1,66 weekend factor applies before or after the category multiplier | Contract cl. 19.3.6 against cl. 19.3.10 and 19.7.6; a category-4 truck on a Saturday is R$ 24,30 or R$ 24,40 | 936 of 10,000 passages held undetermined between two amounts |
| Amounts the contract's own arithmetic produces beyond two decimals | Anexo 12 exempts discounted amounts from rounding; no text settles the centavo | 4,938 of 8,163 passages allowed, with the open centavo registered against the clause that leaves it unsettled |
How it works: four steps, the same ones built for financial regulation
| Step | What it does | What the operator's system gets |
|---|---|---|
| Convert | Turn a written public rule (tolling regulation, concession contract, ordinance, permit, fare rule) into a versioned clause register. Each clause is determinate, undetermined (a standard the text does not settle), or ungrounded (a term nothing public defines). Ambiguities are registered, never quietly resolved. | The register with a source locator for every clause; a gaps file listing every document the rule depends on that is not public |
| Simulate | Run a synthetic population of passages, trips, or missions against the register and an operator's authority. See which actions can be decided, which cannot, and at what refusal rate. | The corpus, every figure with its denominator |
| Determine | In operation, evaluate each proposed charge or move against the operator's delegated authority and the public floor before it happens, for every act class the mandate names: the charge, the refund, the notification, the report. | One determination per action, in front of the charge or the dispatch, in time for the operator's system to act on it |
| Contest | When a charge or a move is disputed, any party re-derives the determination offline and checks it against the register. | The offline check a driver, concessionaire, city, regulator, or insurer runs on their own machine |
We do not claim our encoding is the law. We make the encoding explicit enough that the authority can disagree with it, clause by clause.
Where it applies
Same layer, next domains: free-flow tolling and mobility payments first; sidewalk delivery devices second; then curb, parking and charging fees, automated road vehicles and robotaxis, drones and BVLOS delivery, and ports, airports and industrial zones, each with one public rule, one operator's delegated authority under it, one machine action, and one determination in front of it.
Why now
The rules that price the wrong answer are being written on a calendar. ANTT's free-flow interoperability sandbox is in public hearing on 8 October 2026 with contributions open to 23 October; the draft names identification, integrity, and traceability, with record and evidence duties on participants: duties of the kind a determination produces as a by-product. Singapore introduced a bill on 8 September creating a five-year sandbox for autonomous vehicles on public paths, to 30 September 2031. Chicago's device pilot expires 31 May 2027 without Council action.
What runs today
Before it pays, on a federal concession (September 2026). On BR-101 Rio-Santos under CONTRAN 1.013/2024, ANTT 6.079/2026, and the RioSP concession contract as amended by the 19º and 20º Termos Aditivos, converted clause by clause into 93 register entries (81 determinate, 9 undetermined, 3 ungrounded). 10,000 passages determined before the charge: 8,163 allowed with an amount and the clauses checked, 803 refused, 1,034 undetermined. A subset of 502 posted charges that depart from the rule: 431 refunded at double under art. 63-G, 29 undercharges with no refund, 42 undetermined; every determination re-derives from the register with zero mismatches. On the 502 charges posted at the wrong amount: without the determination, 431 of 502 refunded at double, R$ 3.373,01; with it, 398 of 502 post at the determined amount, 60 never post, 44 are held.
Before it moves, in a city (September 2026). Under Chicago's personal delivery device ordinance and its permit section as written, 0 of 10,000 synthetic trips can be shown to be authorized; under an illustrative permit file and three readings the city could adopt, 2,503 of 10,000 are authorized and 7,497 are refused, each refusal citing the clause. The ordinance converts to 54 register entries (34 determinate, 15 undetermined, 5 ungrounded); eight of the 54 decide whether a single trip is authorized, and public text alone lets a trip be evaluated on four.
Before it pays, in finance (2025 to 2026). The same four steps run in regulated payments: a register of 104 clauses from a national fraud-loss framework, encoded with a bank design partner, and 42,188 synthetic determinations against it.
Stated limits
The live path in front of an operator's charge or dispatch is what the first engagement establishes. All three studies run in batch against synthetic populations. The Chicago permit file is illustrative and the refusal rate under it is a property of that file, not of any operator. The BR-101 corpus is calibrated to concession-level indicators from ANTT's own sandbox report, with its assumptions labeled; its counts describe the corpus, not traffic; the base tariff and index for the September 2026 revision are unpinned. The two mobility studies carry no signed output and are marked synthetic. A live determination needs operational inputs the public rule does not contain (the current permit, rate tables, exemptions, account state); in production the operator supplies them. Re-derivation proves the engine made this determination given this register; it does not prove the authority agrees with the encoding. Nothing here describes any operator's or concessionaire's operations, and nothing is a safety claim, a savings claim, or a proposal for new regulation. The refund, notification, and report determinations run on the same synthetic corpus; no notification was sent and no refund was paid. The mandate carrying the four act classes is specified and unsigned.
What we are looking for
1. A concessionaire proposing for the ANTT free-flow sandbox whose solution names a pre-charge determination as a component (the draft names integrated agents and bars only undue lock-in), or a payment firm joining as an integrated agent under its own admission (identification, integrity, and traceability; central-bank habilitation where required). Outcome: the four act classes run on one real class of charge, with the counts above produced from the concessionaire's own passages.
2. An operator willing to put one real permit or operating authorization through the layer. Outcome within one engagement: a determination before the machine moves on one real class of action; violations on that class measured against the permit.
3. A city or agency drafting rules for autonomous devices that wants the register, and the refusal rate it implies, before the vote.
4. An insurer or reinsurer underwriting autonomous operations that wants the determination, rather than the telemetry, as the underwriting input.