The Investor

The word doing the work is "definition"

The most portable claim in the whole case has nothing to do with AOL or Vimeo. It is that a headline organic-growth figure rests on a definition the company picked itself and outsiders cannot rebuild. The Value Investors Club write-up relayed by The Bear Cave never says how that definition departs from standard practice, which reads like a hole in the reporting, or rather, it is the finding. A growth number nobody outside can reconstruct cannot be disproved from outside. It also cannot be defended from inside. Every consolidator ever assembled has the same joint somewhere: revenue that arrived by wire transfer, sitting in a base later described as growing on its own.

The enforcement half changes who gets to ask. The SEC's Financial Reporting and Accounting Unit sits inside the Division of Enforcement and was built for accounting and financial-reporting fraud plus auditing misconduct. The same Bear Cave issue is the only source, it never says when the unit was stood up, and it never connects the unit to this bear case, so it belongs in the standing-context column rather than the catalyst one. The transmission into a book runs through triage rather than litigation: a dedicated unit turns published activist and short-seller work into inquiries at a higher rate than a generalist docket does, and an S-1 is where KPI definitions and unremediated material weaknesses stop being awkward and become federally interesting. Timing skepticism is warranted: enforcement units take years to show up in outcomes. Direction is the claim, not velocity.

The comp that actually moves the marks

The same roundup drops a number with more effect on carrying values than Bending Spoons will ever have. Cloudflare trades above 35x revenue on 12% incremental operating margins, twelve cents of each new revenue dollar reaching operating profit, and it is growing more slowly than Palantir while doing it. A 2025-vintage Series C security or edge mark anchored to that name's headline EV/Revenue is asking one multiple to price growth and price the quality of that growth simultaneously. LPs tend to find that seam before the GP writes it down.

The AI line is a discount, not a premium

One element runs backwards against the pitch of the past two years. AI appears here as terminal-value risk to legacy software assets, not as upside narrative. Through 2024 and 2025 an AI roadmap was worth points of exit multiple. In this write-up AI exposure is the reason a public buyer marks legacy revenue down. If that framing holds among public buyers, the classic consolidator trade of buying declining assets cheaply, running them for cash and capitalizing the aggregate at a growth multiple has a shorter runway than most platform models assume.

Calibration matters before anyone re-marks anything. This is one pseudonymous author rather than an audited forensic review, relayed by an activist outlet that advertises vindication on its own prior work and disclaims investment advice in the same breath. The correct use is not a valuation call on any name. It is a diligence checklist, because the questions survive being wrong about the target, and the unit that makes the last question expensive exists whether or not the bear thesis pays.

Two sprints of work, priced as defensibility

These numbers arrive by way of Daily Dose of Data Science resurfacing Plan-and-Act (arXiv 2503.09572), which is not the same as a new release, and the difference matters for underwriting: the recipe has been sitting in public long enough that anyone who wanted to copy it has had the chance. The three moves producing the gain are structural and fully published. Split the loop into a planner and an executor, strip stale HTML out of the execution context after every action, rewrite the plan as the environment's state changes. No frontier-model access, no data partnership, no additional training compute. That is a pricing input rather than a technical curiosity. A seed or Series A deck calling proprietary agent orchestration its moat is describing a documented, reimplementable delta, and the premium in those rounds is paid for defensibility, not capability.

The failure case is where the underwriting actually lives, or rather where the interesting version of it lives. A planner finetuned without ever seeing the target sites produced steps that read beautifully and matched nothing on the page, and the executor followed them anyway. That configuration landed at 20.60%, roughly sixteen points below shipping no planner at all. Plan quality is a function of environment exposure. Which rewrites the first diligence question. Not "what is your architecture" but "whose environments have you seen, and can you keep getting more of them?"

That points the defensible companies away from the open web and toward permissioned private surfaces: internal enterprise applications, ERP, claims systems, brokerage back-office. Trajectory data, meaning recorded sequences of steps that completed a real task, compounds inside a customer relationship there and cannot be scraped by a competitor. An agent company working the public web is running a public recipe against a public dataset.

Accuracy is a dial on gross margin

The largest single gain, +10.31 points from dynamic replanning, costs one extra planner call for every executor step. The paper's authors flag that linear overhead as unresolved and float letting the executor decide when to replan. For an investment committee the consequence is blunt: gross margin in agent products is a dial set by the accuracy target, not a constant that improves with scale. Cost-per-call reporting conceals this completely. The figure worth demanding is cost per successful task, with a margin sensitivity table at 70%, 80% and 90% target success.

The ceiling deserves equal weight. 53.94% was the best number in that paper's own comparison rather than a current-quarter frontier print, and under it nearly half of runs still fail. Any pitch premised on full autonomy replacing a human at enterprise service levels is pre-product, however clean the demo looks. Demos deserve extra suspicion here, because planners degrade sharply on environments they were not trained on. The demo environment is usually the trained one.

Calibration: this is one paper on one benchmark, and not a recent one, and no round sizes, revenue figures, or multiples appear anywhere in the material. Call it a thesis input rather than a comps update, and what it changes is what underwriting asks for. The honest counter-case is cheerful, and this is probably where the argument is weakest: if capability costs harness engineering rather than training runs, agent companies need less capital to reach useful accuracy, which improves capital efficiency across the category even as it compresses the defensibility premium in any single round.

Leg one: pricing, and where the 30% comes from

FISN went out looking for $24 a share and priced at $16, per The Bear Cave's roundup, which is a 33% gap between the number the sellers defined and the number the market was actually willing to fund. Blue Owl put $20M into the same name. The same roundup carries a bear price target of zero on it. That is the useful bracket: institutional capital will fund a listing at a third off, and the discount buys no consensus whatsoever about the asset underneath. The modeling consequence for base-case listing assumptions is a 30% pricing haircut, set deliberately just inside the observed 33%, because one print establishes a direction and not a distribution. This is probably too conservative. If the next two prints clear at target, the haircut comes back out.

Leg two: one-third sellable at pricing retires the lockup line

One-third of FISN's shares were immediately sellable at pricing, which is the kind of deal term that reads like housekeeping and then quietly eats a model. It collapses two events every exit model treats as separate: the clearing price and the first genuine trading price. When a third of the register can sell on day one, the comfortable assumption that a lock-up delays real supply stops describing anything that happens. So the second change is a no-effective-lockup float scenario. Model distribution against day-one supply, not against a post-lockup window that in this structure never arrives.

Leg three: the financing failures decide who waits, and for how long

The two financing items in the same roundup are not IPO events, and neither is a fresh print. Sunrun's already-reported second quarter burned $620M with EBITDA again failing to cover interest expense and second-half guidance cut. UWM, America's largest mortgage lender, eliminated its dividend and announced rescue financing. What they set is the hurdle for anyone hoping to list into this window. The buyers repricing a listing by a third and demanding immediate float are the same buyers refusing to fund a coverage gap, which leaves a non-top-decile issuer two options: accept the haircut and the day-one supply, or wait. Most sponsors wait. Waiting is also a decision not to return capital for another year, which is the part that rarely survives into the committee memo. The 12-month hold extension for non-top-decile names is the convention that follows from waiting, one further annual valuation cycle before the alternative gets re-tested, and it should be described at committee as a chosen convention rather than a figure derived from FISN's print. The 30% and the float scenario come from the print. The extra year comes from the decision the print forces.