Request for comments
The seed paper is everything I could gather from public sources and my own practice. This is where it runs out. Each note below is pinned to a passage in the paper. Pick one and reply, or select any other passage in the paper and start your own thread.
Things nobody has measured in public, as far as I could find.
1 Section 13, Where this breaks, and what is unknown
“I found no time series of coefficient drift”
How often does a coefficient move, and by how much, across a hardware generation or a code change? I haven't found a public series. Even the shape of one would help: a ratio over two or three years, numbers disguised or indexed. Is drift a slow creep, or mostly a few big steps?
2 Section 13, Where this breaks, and what is unknown
“how often launches slip”
How well do declarations land, by stage? The paper's illustrative rule is that intents come in around 60% of declared size and configured asks within 10%. That's a guess dressed as a number. If your team keeps any record of asked against landed, what does it look like?
5 Section 13, Where this breaks, and what is unknown
“how long large new loads wait for power”
How long does a large new load really wait for power, from request to energized? Lawrence Berkeley National Laboratory found little published. Utility filings and interconnection queues may hold the answer. Has anyone pulled the numbers together?
8 Section 8, The commit-point method
“Nobody has measured it in public, as far as the sources show.”
Re-pegging a launch reserve to another tenant when a launch slips looks nearly free. Across enough launches it may absorb most timing misses. That's speculation. Has anyone measured how often a slip's reserve gets re-used?
12 Section 13, Where this breaks, and what is unknown
“The planner can only bound it, phase it and buy options.”
What do you do with a step that has no reference class, such as a first accelerator cluster or a first region on a new continent? Bound it, phase it, buy options is all the paper offers. What else has worked?
Choices the seed paper makes by judgment. I’d like to hear how you make them.
3 Section 13, Where this breaks, and what is unknown
“I found no published mechanism for arbitrating scarce accelerators between them”
When accelerators are short, who decides between internal teams and external customers, and how? Operators have published splits and priority orders. I haven't found the mechanism: the forum, the rule, the appeal. How does it work where you are?
7 Section 8, The commit-point method
“A data scientist could do better.”
The method buys each supply layer to its own percentile. That's a rule of thumb, not a joint optimum. I'd welcome someone who can do better: a simulation, a closed form, or a reason the rule of thumb is good enough.
9 Section 6, Approval is not allocation
“So the cost of being short rises with lead time”
The paper's ratios, 4 : 1 for machines and 19 : 1 for power, are illustrative. What ratio would you actually set for each layer, and who in your organization would sign it?
10 Section 5, Coefficients are agreements
“Count the all-teams record as eight launches”
How much should a team's own record count against everyone's? The weight of eight launches is invented. Insurers have credibility theory for this. Does anyone apply it to capacity declarations?
11 Section 9, The operating discipline: keeping the agreement
“the step most often missing is the supply review”
In my experience the supply review is the step most often missing between the demand review and the executive decision. Is it missing where you work? If you have one, what does supply bring to it?
13 Section 11, What changes in the AI era
“caps need the same stage records”
An open question for anyone running agents at scale: when a team's own software creates the demand, is a cap a better declaration than a forecast? Tell us what you've seen.
14 Section 14, A minimum viable discipline
“A coefficient is an agreement among them.”
The strongest challenge to this paper's central claim that we know of: where a market price exists, does the agreement still hold? Argue it in the thread, either way.
15 Section 13, Where this breaks, and what is unknown
“The most useful next contribution would be multi-organization data”
Could this community pool anonymized declaration records? Say what your organization would need to see first.
Where a public source, a dataset or a better method would move things most.
4 Section 13, Where this breaks, and what is unknown
“I found no published order-to-dock time, turn-up duration or hardware contract terms”
Supply is the least documented side. Public sources welcome: a filing, a talk, a vendor document that states order-to-dock time, turn-up duration, cancel windows or reschedule terms. Please add them as sources, with a link and the quote.
6 Section 13, Where this breaks, and what is unknown
“Attribution for stateful services and batched inference is open”
Per-request coefficients break for caches, batches and stateful services, and batched inference is the big new case. If you've attributed cost on one of these, what unit did you use, and what did you do with the remainder?
Anyone can start a thread on any passage: select it in the paper and choose Discuss this passage. Threads that settle into a change follow How it grows: open at least 14 days, every objection answered, consensus called and written down.