Concept prototype for venture debt loan origination
Good prose hides bad numbers.
A verification workspace for the concept memo. Siddharth Mehta for Obin.
What the agent actually changed
Four hours became four minutes, and got worse.
The analyst now holds a well written memo with thirty eight numbers in
it and no way to tell which the machine invented. So they check all thirty eight.
The one rule everything is downstream of
The better the model writes, the more dangerous it gets.
Fluency is not evidence.
This tool cannot stop a bad number. It can stop one from looking like a
good one.
The signature move
You can see what a machine claimed without reading it.
Serif for what a person wrote. Monospace for what the machine asserted.
The underline carries the state, so it survives being printed in black and white.
Cash of $9.14M against monthly net burn of
$380,000. On the facility that gives 39 months
of post draw runway. The company reports 112% net dollar retention.
provedopenconflictunsupported
Ordering is the product
The document decides what you read first, and it is wrong about what
matters.
Sorted by leverage crossed with reliability, the undisclosed lien is
first. In the order a person reading top to bottom would meet it, it is ninth.
Thirty eight figures, and the two you are not told about
Accuracy is measured on figures the analyst chose to check, not on figures
the agent chose to show. The count is disclosed. Which two never is.
Measured, never predicted
Model confidence appears nowhere in this
product.
99when a tool ran the arithmetic and the model never touched the number
77when it read the figure off a slide someone wrote to persuade you
Backtested against two hundred and fourteen closed memos where the
analyst's own number exists to compare against. Probability measures fluency, and a hallucinated
figure often scores higher than a correct hedged one.
The metric the firm buys on
Throughput counts the memo that was never written.
circulatedkilled earlythis dealqueued
Two of these memos were stopped partway because the agent surfaced a
disqualifying fact early. A killed deal with a defensible reason is throughput, not a gap.
Said first, so it cannot be used against me
assumptionThe credit box is market practiceNot Obin's actual thresholds. If yours differ, the ordering changes and the design
does not.
assumptionThe data room is a guessAt what a venture debt deal actually arrives with. The gap in it is deliberate.
assumptionCalibration is silent on noveltyA genuinely new kind of claim has nothing to score against, and the product should
say so rather than invent a number.
assumptionNo user researchEvery workflow assumption is inferred from Sindhu's framing and your site, not
observed. First thing I would test.
The question I actually want answered
Is the bottleneck producing the memo, or the committee trusting it once it
arrives?
I built for the first. I am not certain that is right.