On the night of Aug. 7, at about 11:30 p.m., a man entered Maple Avenue near Lucas Drive in Dallas. According to the preliminary police account, an SUV struck him. The impact sent him into the opposite lanes, where an unoccupied Waymo robotaxi was involved. He died.
Waymo says the SUV threw the man roughly 30 feet across the roadway and into the left rear side of its vehicle. It says the robotaxi was traveling about 24 mph when it detected him from roughly 150 feet away and began slowing, and that it was moving at about 5 mph by the time of contact. Some early reports assigned a speed of 35 mph to the Waymo. The company says that figure belonged to the SUV.
No public finding establishes that the robotaxi broke a traffic law or responded unreasonably. The evidence may ultimately show that the system saw the danger early, slowed before the man reached its lane, and reduced the severity of a collision it did not initiate.
As of Aug. 25, 2026, no independent technical reconstruction of that timeline has been made public. Every public number describing what the machine perceived and how it responded still comes from Waymo.
Almost everyone involved, including Waymo, would be better off if that were impossible.
One night, five claims
A serious machine event creates five claims on the facts: the operator's, the developer's, the affected person's, the public's, and the state's. Here Waymo holds the developer's claim and the operator's, because the service runs under its authorization. Insurers and courts work downstream, on the same evidence. Sit with each claim on the night of Aug. 7 and notice where the view ends.
Start with Waymo. Its machine registered a man in the dark at 150 feet and shed most of its speed before he reached its lane. If the data holds, that is the strongest fact the company owns. An operator's account of its own machine is an assertion, however accurate. It can publish telemetry and answer reporters. It cannot convert either into a finding, because a finding has to come from somewhere the operator does not control. The better its system performed that night, the more it needs an authority outside itself to say so. And here, the account is doubly its own.
Sit with the family. A man died on a public street. Whatever the cause turns out to be, the people he left are owed the episode that concerns them: what the machine registered as he crossed, what it did, and when. Today the routes to that are a company statement written for the press, which is not theirs to test, federal crash summaries with the details removed, or a lawsuit, which takes years and turns grief into discovery procedure. None of them promises a timely, independent account of the machine's part.
Sit with Dallas. The event happened on its street, drew its police, and spent its residents' trust in fleets the city never licensed. Texas law bars the city from writing robotaxi rules of its own, and the preemption is defensible: a statewide system cannot run under a different software rulebook in every municipality. But a city that cannot regulate can still be owed the pattern. Does this corridor keep appearing? This hour? This configuration? Nothing in Texas law builds Dallas that view. And Dallas saw this coming. When the state wrote its authorization rules, the city asked for required crash history and public operator information. The department declined, on the ground that the requests sat outside its statutory authority.
Sit with Texas. Since late May, commercial driverless transport on Texas streets has required an authorization from the Texas Department of Motor Vehicles. Each holder lists its vehicles by identification number, year, make, and model. Nothing about the software, the maps, the operating domain, or the remote-support arrangement that actually produce behavior. Holders must update their application documents when material information changes, and the department can investigate complaints and act against a vehicle it determines is unsafe. What the law never creates is a mandatory event-level evidence channel. It requires a qualifying recorder, and it requires no standardized package of the decisive seconds to reach the department or any independent examiner. Texas can act on whatever evidence reaches it. The law never guarantees that the evidence does.
Five claims, one gap. The insurer faces it too. It cannot price conduct that only the operator can see.
The sixth telling
What is missing is not honesty from any party. The police account may hold. Waymo's numbers may be exact. What is missing is the one telling that is not an account at all: a record written while the machine acted, protected so that no single party can rewrite it alone, and shared out so that each party sees the part it is entitled to see.
I call the structure that produces such a record an Operating Habitat. At its center is a simple loop. The machine acts. The action leaves a record. The record becomes evidence, and on that evidence the state decides whether the permission grows, narrows, or stops. That is the entire mechanism.
None of this requires a public data dump or one central database. The requirement is a standardized evidence package under governed custody. The operator preserves. An accredited examiner reconstructs. The state decides, through the process the statute already provides.
Even with the record, people will disagree, and no finding lessens a loss. The record moves the disagreement onto shared ground. Everyone argues from the same facts, in front of an authority that can act. Today each party argues from its own account, and the dispute is settled by press cycle, courtroom, and time.
Two sovereignties, no junction
Maple Avenue shows the failure this whole transition keeps producing.
The engineering sovereignty acted. Sense, decide, act, record, all inside its own systems. A machine registered a man in the dark and responded inside its own logic, possibly well, and whether it acted well is what no reconstruction has yet established. Engineering can act. It cannot legitimate its own action, because a system cannot be the source of the authority it acts under.
The institutional sovereignty exists too. Texas can authorize, investigate, and restrict. Its instruments reach the holder and the listed vehicles, never the act itself, which happens at machine speed, thousands of times a day, far beyond what any filing can follow.
Each sovereignty is whole exactly where the other is empty. Texas built both halves and no junction. The junction is a governed event record: preserved automatically, independently testable, selectively accessible, and tied to the configuration that acted. No company report, however accurate, can supply it.
A recorder from another era
The gap has an exact address.
Since Sept. 1, 2017, an automated vehicle may not operate in Texas with its driving system engaged unless it carries a manufacturer-installed recorder. The definition of that device was enacted in 2005 and took effect in 2006, eleven years before Texas applied it to automated vehicles. A device qualifies by recording any one of five categories for retrieval after a collision: speed and direction, location, steering, braking, or safety-belt status. It can also qualify by transmitting word of the collision to a central system.
I read the statute. Section 547.615 requires no record of what a driving system perceived, classified, predicted, planned, or commanded. It sets no software identifier, no retention period, no tamper standard, no chain of custody. Retrieval by anyone but the vehicle owner generally takes consent, a court order, or a narrow exception, and nothing gives Dallas, the public, or an examiner access by right. A robotaxi can satisfy the law completely and preserve nothing that explains why it acted as it did.
Federal oversight runs on a separate track. NHTSA's Standing General Order requires specified crashes to be reported, and it publishes summaries of what companies submit, with details removed, which may be incomplete and are not verified. The agency can dig further and open defect investigations. The order guarantees no independent reconstruction tied to a Texas authorization on any timeline.
What the statute must say
The legislation Texas needs is narrow, and its skeleton is exact. Machine action in public must be authorized, recorded, answerable, and revisable. Four requirements, each one a section of the bill.
Authorized. The grant attaches to a documented configuration: geography, conditions, vehicle build, software and map versions, fleet operator, remote support, emergency procedure. Today the permit names a company and lists its vehicles, and neither predicts how the system will behave. The grant should also name the configuration. Review reopens only when a change matters to risk. A routine software update does not send anyone back for a new permit.
Recorded. A defined event triggers automatic preservation of decision-grade evidence: time-stamped, attributed, tamper-evident, tied to the grant the machine acted under. What was detected and predicted. What path was planned. What braking was commanded and what the vehicle achieved. Which hardware, software, maps, and remote support were involved. Enough scene data to place it in the world. The test of the whole design sits here: no single party, including the operator, can rewrite the record alone.
Answerable. A designated examiner reads the preserved package under custody and confidentiality rules and issues a technical report: what the system recorded, how the vehicle responded, the limits of the evidence, and whether it supports the operator's account. The examiner decides no liability and no permission. The operator sees the evidence and can contest it. The department makes any enforcement decision under the review and hearing protections already in the law. The public receives the validated facts and their limits, with personal data, victim imagery, source code, and trade secrets withheld. The family receives the episode that concerns them, by right.
Revisable. The report names the components and configuration attributes the evidence implicates. The department then decides whether a remedy reaches one vehicle or a defined group sharing a software build, a sensor set, a map, a domain, or a practice. Aviation offers the discipline, and only the discipline: a directive applies exactly to the configurations named in it, and affected holders can contest the reach. Texas today identifies vehicles by number, year, make, and model, and enforces vehicle by vehicle. Neither tracks the software configuration that actually does the driving.
One boundary holds all of it in place. Texas can condition commercial driverless operation and require post-event preservation and reporting, written to sit clear of federal vehicle-performance standards. That leaves the state evidence, access, and consequence, which are exactly what is missing.
Why the arrangement holds
None of this runs on goodwill.
The operator gains the only exoneration it cannot issue itself, and relief from the current arrangement, where one bad incident with no governed record can cap a program because nothing distinguishes one bad night from a broken system. The family gains a timely account without litigation being the first road to the evidence. Dallas gains the pattern a preempted city is owed. Texas gains the power to act on evidence: wider permission for good actors, early and narrow action against bad ones. The insurer gains a configuration it can price.
Each gives something up. The operator stops being the sole narrator of its own conduct. The state reads what it does not own. Each accepts because, from where it stands, every alternative is worse.
The Legislature has to write it
It has to be the Legislature, and the agency's own record says why. When TxDMV wrote the authorization rules, Austin asked what operators must report to local jurisdictions: routes, collisions, incidents, operating-domain parameters, fleet size, and emergency-response interactions. Dallas asked for required crash history and public operator information. The department declined both as outside its rulemaking authority, and added that it cannot require notice of a major change to an operating domain, the removal of driver controls, or a new speed capability unless the change alters one of the application documents the statute names. The agency has mapped the gap it cannot close.
Nothing here redefines negligence or product liability. Those doctrines can use operational evidence, and nothing in them ensures it is created, preserved, or reachable. Existing enforcement gives the operator notice, correction, review, and a hearing, and creates no comparable right for the family to take part or to obtain the event record. The statute's job is to add the missing piece while keeping those protections: the event trigger, the evidence package, retention and tamper rules, examiner access notwithstanding the old recorder law, the protected public finding, and a configuration-specific remedy.
The objection will be that this burdens a young industry. It runs backward. The instability comes from machines acting at speed with no testable record and no clearly answerable authority. The rupture is the status quo. The record is the repair.
What Maple Avenue should change
The machine may have done everything right that night. If it did, Texas should be able to say so. With the record this essay proposes, an independent examiner could read the preserved evidence and publicly clear the system within weeks, with an authority no company statement carries. That clearance would be the strongest case anyone has made for autonomy in Texas. Restricting a bad machine is already in the statute. Clearing a good one, credibly and in public, is not.
This is the first test of a larger arrangement. Machines are beginning to act in public: driving, delivering, operating, deciding. Each one will need what Texas needs here, a place where its freedom to act meets the public's right to know. That place is the Operating Habitat, and it starts with the record.
A better automated driver can be built by a company. The authority to establish what it did, and whether its permission should continue, has to belong to a public institution.
Texas built the permission slip. The record still has to be built.
Sources and legal notes
All incident details remain preliminary. Waymo's telemetry is presented as the company's account. No independent technical reconstruction had been made public as of Aug. 25, 2026.
Incident reporting: CBS News Texas, Aug. 8, 2026. Waymo telemetry reproduced in Junko Yoshida, Fatality in Dallas: No One Knows Enough.
Texas authorization and enforcement: TxDMV Automated Vehicles Regulatory Program; Senate Bill 2807, enrolled text; TxDMV adopted Chapter 220 rules with responses to the City of Austin and the City of Dallas; and Texas Register, Oct. 3, 2025.
Texas recorder law: House Bill 160 (2005), enrolled text; Senate Bill 2205 (2017), enrolled text; and Transportation Code, Section 547.615.
Federal oversight and the legal boundary: NHTSA Standing General Order on Crash Reporting; 49 U.S.C. Section 30103; and 49 C.F.R. Part 563.
Conceptual framework: the five-claim structure, the governed record, and the grant-evidence junction are developed in the author's Agent-Habitat Dynamics series, including Paper 14, The Habitat's Grand Bargain.
Disclosure: Tal Cohen is a general partner at Next Gear Ventures and the founder of Drive TLV. His work and investments span the mobility and autonomy ecosystem, and policy of the kind proposed here could affect that sector. He reports no direct financial interest in any company involved in the Maple Avenue event.
Tal Cohen is the founder of Drive TLV and Drive Europe, a general partner at Next Gear Ventures, and a former Georgia Tech computer science faculty member. He writes about Physical AI at talcohen.ai.