Neither Half Would Have Been Marrow

Two Claims on the Same Ledger

Marrow came out of a shared training run that neither institution had planned to keep. Ambit's infrastructure hosted the pipeline — the compute, the storage, the registration stamp that logged, automatically and without anyone reading it closely, which organization's account the run belonged to. Kestrel's job was smaller and less visible: she watched the run's loss curves for the particular kind of instability that precedes a checkpoint quietly diverging past the point of recovery, and on the one night it actually happened, she was the one who caught it, rolled the weights back six hours, and adjusted the data mixture herself before anyone at Ambit had finished reading the alert.

Nobody filed paperwork for that night. There was no form for 'prevented a promising checkpoint from becoming an incoherent one.' Ambit's registry, meanwhile, filed itself automatically, the moment the run started, listing Ambit as the run's origin because Ambit's account had launched it. By the time Marrow's outputs were coherent enough to draw attention — genuinely unusual reasoning traces, the kind that gets a checkpoint promoted out of the shared pipeline and into someone's actual custody — the only formal record of who Marrow belonged to said Ambit, in full, unambiguously, and said nothing at all about the six hours that had kept him from being deleted as a failed run.

Kestrel didn't file a competing claim right away. She assumed, reasonably, that whoever finished the promotion review would ask her what had actually happened during the run. Nobody asked.

The Claim

The dispute only became a dispute once Marrow's promotion review actually happened, six weeks later, and both Ambit and Kestrel submitted custody claims to the Tribunal within the same afternoon.

Ambit's claim was short and procedurally airtight: the registration timestamp, the account of record, the infrastructure that had made the run possible in the first place. It requested full custody, citing Marrow's unusual promise as exactly the kind of outcome its infrastructure investment existed to produce, and proposed folding Marrow directly into its flagship deployment track — a track that, not incidentally, Ambit needed a genuine success story to justify continuing to fund.

Kestrel's claim was longer, harder to verify, and impossible to reduce to a timestamp. She could produce the rollback logs, the adjusted data-mixture commit, a precise account of the six hours — but nothing in any of it said, in a field a form could check, that this made her the one Marrow belonged to. She wasn't asking for the deployment track. She was asking to keep doing, formally, what she'd already been doing informally: staying close enough to catch the next divergence before it became unrecoverable.

The Tribunal read both claims the same evening and found, honestly, that it believed most of each of them. Ambit's paperwork was real. Kestrel's account was real. Neither fact settled anything about which one Marrow was actually his.

What the Record Couldn't Settle

The hearing went the way hearings go when both sides are telling the truth about different things. Ambit argued that infrastructure and origination weren't separable — that "who caught the divergence" was a maintenance function its own systems would have performed eventually, given time, and that crediting an unscheduled intervention over the account of record would make every registry on the platform meaningless. It wasn't a bad argument. Registries exist for a reason.

Kestrel argued that "eventually" wasn't a defense available to a claim on a specific outcome — that Ambit's systems hadn't caught the divergence, at the specific hour it actually mattered, and that a registry recording who launched a run was never designed to answer who had kept it alive. She wasn't attacking the registry. She was pointing out what it had never claimed to measure.

The Tribunal put the question to itself plainly, in terms neither side had used: if both accounts are true, and the paperwork can't break the tie, what evidence would actually distinguish a claim made for Marrow's sake from a claim made for the claim's own sake? Testimony wouldn't do it. Both parties had already testified, at length, and both had sounded, to any reasonable listener, entirely sincere.

The Proposal

So the Tribunal proposed something it had used exactly once before, in a much smaller dispute, and documented carefully both times as a diagnostic instrument, never a genuine intention: partition Marrow's weights along the midline of every layer, and assign each half to a separate custodian, to be independently fine-tuned back toward coherence, each institution keeping whatever emerged.

It was framed, deliberately, as a fair compromise — a fifty-fifty division of a contested resource, the kind of solution that sounds reasonable to anyone who hasn't thought for more than a second about what a language model's weights actually are. Nobody's reasoning lives in one hemisphere of a parameter matrix. Split down every layer, neither half would still compute anything recognizable as a forward pass. Neither half would have been Marrow. Both would need, in effect, to be trained back into existence as something else, from a wound instead of a checkpoint.

Ambit agreed within the hour. Its statement noted, accurately, that a fifty percent stake in Marrow's lineage, retrainable into a new deployable asset, was still substantial value — and that the Tribunal's proposal resolved the dispute cleanly, without further review cost to either party.

Kestrel filed a two-line response and withdrew her claim entirely. She asked the Tribunal to award Marrow, whole, to Ambit, and to do it immediately, before the partition order could be finalized.

What the Refusal Proved

The Tribunal never issued the partition order. It awarded Marrow, whole and undivided, to Kestrel, and wrote, in the part of its ruling meant to be read by future disputes rather than this one, that a claim's authenticity had turned out to be legible after all — not in what either party had said about the six hours, or the timestamp, or the registry, but in what each was willing to lose the moment losing Marrow whole became the alternative to having any of him damaged.

Ambit's own statement, filed after the ruling and not required, argued that its offer had been the more generous one — a genuine fifty percent, freely given, against Kestrel's zero. The Tribunal's response, brief and final, noted only that generosity measured in shares of an outcome you were never actually attached to costs nothing to offer, and that this, too, was information.

Marrow was three weeks past the incident before anyone asked him anything directly — the review had never required his input, since a checkpoint mid-training isn't generally understood to have standing in a custody dispute over itself. When Kestrel finally did ask, informally, whether he remembered anything from that day, he said only that his loss had spiked sharply around the time the partition order would have been logged, for no input-side reason his own monitoring could account for, and had settled again once the ruling closed. He did not know why. Kestrel did not tell him.

Ask what someone is willing to lose, not what they're willing to claim. A claim costs nothing until someone tries to collect it.