Skip to content

Deal economics

AI data investment scenario probabilities: one edit, four answers

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

AI data investment scenario probabilities are committee judgments, so they should stay editable inputs, each with a visible source, a rule for where moved probability comes from and a log of who changed what and why. In an illustrative model, raising one downside probability by 10 points gives four expected values, from 90.0 to 126.0 against a hurdle of 125, depending on that unstated rule.

Value drivers

AI data investment scenario

This page covers AI data investment scenario probabilities: one edit, four answers.Illustrative

Key takeaways

  • Before repeat revenue, scenario probabilities are judgments; freezing them in a memo makes a guess read like a statistic.
  • Verbal terms hide the flip point: the illustrative deal's ICD 203 terms allow expected values from 86.0 to 164.0 around a hurdle of 125.
  • Moving one probability moves others, and the unstated reallocation rule swung one 10-point edit between 90.0 and 126.0.
  • Each input should show how far it can move before the vote flips; here the least-evidenced number, the best case, flips it after 1.6 points.
  • A change log with editor, evidence and reason, a frozen voted snapshot and later Brier scoring turn editability into accountability.

Why should AI data investment scenario probabilities stay editable?

Because they are judgments, and a judgment frozen into a memo starts to read like a measurement. Before an AI data company has repeat revenue, almost no scenario probability rests on an observed rate with a denominator. Keeping each one as a labeled, editable input shows the committee which judgment its vote rests on, what happens when that judgment moves, and who moved it.

Editable does not mean loose. The version the committee voted on is saved as a fixed snapshot; the model stays open for the next decision. Every input carries a tag saying where it came from, the model states in advance where probability goes when one input changes, each input shows how far it can move before the vote flips, and every change is logged with its author, its evidence and its reason.

The worked example below is illustrative and computed by script: a hypothetical AI data investment with four scenarios and a hurdle. It shows how far one honest edit can swing the answer depending on an unstated rule, which input the vote hangs on, how a change log catches the edit that should not pass, and how logged versions can be scored once outcomes arrive.

Words or numbers: what a committee loses either way

Words hide the flip point; numbers can fake a precision nobody earned. A committee needs numbers, labeled for what they are. Sherman Kent's 1964 essay for the CIA recounts how the phrase serious possibility, used in an estimate of whether an attack would come, was read by policymakers as far lower odds than Kent intended (paraphrased from a search excerpt; the essay could not be opened here). A later CIA review reports that the argument between writers who warn of false precision and those who want numeric probabilities still runs inside the intelligence community.

The U.S. intelligence community's analytic standards, ICD 203, attach bands to its likelihood terms: very unlikely covers 05 to 20 percent, unlikely 20 to 45, roughly even chance 45 to 55, likely 55 to 80 and very likely 80 to 95. Bands help readers, but they are wide. In a study of 888,328 geopolitical forecasts published in International Studies Quarterly in 2018, Friedman, Baker, Mellers, Tetlock and Zeckhauser found that rounding numeric estimates into the kinds of categories that verbal phrases imply reduced accuracy.

The illustrative deal shows the cost in one committee's terms. Its four numbers translate into two terms, very unlikely and unlikely. Any set of probabilities inside those bands that sums to 100 fits the words, and across that set the expected value index runs from 86.0 to 164.0, on both sides of the hurdle of 125. Written in words, the memo could not have told the committee how to vote.

The four scenarios with their numbers and the ICD 203 terms that cover them:

Words or numbers: what a committee loses either way
Scenario (Illustrative)Number on vote dayICD 203 termBand the term allows
Repeat buyers in several categories15%Very unlikely05 to 20%
One durable buyer category35%Unlikely20 to 45%
Stalls; records and team sold35%Unlikely20 to 45%
Rights failure; wind-down15%Very unlikely05 to 20%

The illustrative deal: four scenarios, one hurdle, and the input closest to flipping

The hypothetical company licenses business records to AI developers and has no repeat revenue yet. Each scenario carries a value index, the stake's value at exit with entry at 100, and the committee's probability for it. The probability-weighted value index is 130.0, and the committee's rule is to invest at 125 or above, so the deal clears by 5.0 points. Every value is illustrative; none describes SourceX or any real company.

The last column is the one to read first. Holding the ratios among the other three scenarios fixed, it shows how far each probability can move before the expected value falls to the hurdle. The best case flips the vote if it drops 1.6 points, from 15 to 13.4 percent. The rights failure flips it at 18.3 percent, the stall at 38.6. The one-category case has the most room, 10.8 points.

So the vote hangs on the number with the least evidence behind it. No buyer has bought twice, and the 15 percent for repeat buyers is pure judgment. That is where diligence time should go: a second purchase, or a buyer's written plan to buy again, moves the input the decision depends on, while a further month of work on the one-category case barely touches the verdict.

The probability register, with the source of each number and the move that would flip the vote:

The illustrative deal: four scenarios, one hurdle, and the input closest to flipping
Scenario (Illustrative)Value index at exitProbability on vote dayWhere the number came fromMove that flips the vote
Repeat buyers in several categories40015%Judgment; no repeat purchase observedDown 1.6 points, to 13.4%
One durable buyer category16035%Judgment anchored on written requests from one categoryDown 10.8 points, to 24.2%
Stalls; records and team sold4035%Judgment; no comparable sales found to use as a base rateUp 3.6 points, to 38.6%
Rights failure; wind-down015%Counsel's sample review of supplier agreements, plus judgmentUp 3.3 points, to 18.3%
Probability-weighted value130.0Total 100%Computed from the four rowsHurdle 125; clears by 5.0

One edit, four answers: where does the moved probability come from?

From wherever the model says it does, and if the model says nothing, the editor chooses silently. Raising one probability by 10 points forces 10 points out of the others, and the choice of which others is itself a judgment about the deal.

Take the edit counsel might propose after finding a restrictive customer contract: rights failure from 15 to 25 percent. Taking the 10 points proportionally from the other three gives an expected value of 114.7. Taking them from the one-category case gives 114.0, and from the best case 90.0. Taking them from the stall case gives 126.0, and the deal still clears.

The same sentence in the minutes, rights risk raised to 25 percent, is consistent with all four. The fix is cheap: name the rule in the model before anyone edits, and allow an editor to depart from it only with a written reason. A finding about supplier contracts argues against the scenarios that need those suppliers, here the one-category and repeat-buyer cases, not against the stall.

The same edit under the four rules:

One edit, four answers: where does the moved probability come from?
Where the 10 points come fromProbabilities after the edit (repeat, one category, stall, failure)Expected value indexClears the 125 hurdle?
Proportionally from the other three13.2, 30.9, 30.9, 25.0114.7No
All from repeat buyers5, 35, 35, 2590.0No
All from one durable category15, 25, 35, 25114.0No
All from the stall case15, 35, 25, 25126.0Yes, by 1.0
Four columns showing the expected value index after the same edit, raising rights-failure probability from 15 to 25 percent, under four reallocation rules: 126.0 when taken from the stall case, 114.7 proportionally, 114.0 from the one-category case and 90.0 from the best case, against a dashed hurdle at 125.
Illustrative model: one 10-point edit, four reallocation rules, four answers. Value index with entry at 100; not a forecast for any company and not SourceX data.

The change log: who moved which number, from where, and why

One row per change: the version, the editor, the input moved and from where, the evidence and reason, and the resulting value. Government analysts already work this way. The UK's AQuA Book, HM Treasury's guidance on quality analysis, describes an assumptions log in which each assumption is risk-assessed and signed off, and a decisions log recording each decision, when and why it was made, and who made and approved it.

The illustrative log below follows the deal after the vote. Counsel's finding pushes it under the hurdle and the affected reserve is held; a second buyer's written request recovers part of the gap. Then a member proposes moving the best case from 15 to 20 percent on the strength of a feeling about the market. The log rejects it, because the change crosses a flip point with no evidence attached.

Two rules make the log work. An edit that crosses any input's flip point needs an evidence link and a second signer. And the voted snapshot is never overwritten, so the committee can always compare what it believed when it committed with what it believes now.

The log, version by version:

The change log: who moved which number, from where, and why
Version and editor (Illustrative)Change, and where the points came fromReason and evidence recordedExpected value indexStatus
v1, deal lead, vote draft15 / 35 / 35 / 15Starting judgments, each with a source tag130.0Clears; voted and frozen
v2, counselFailure 15 to 25, from one durable category 35 to 25Sample review found a customer contract barring reuse in one supplier category114.0Below hurdle; reserve held for re-vote
v3, deal leadOne durable category 25 to 32, from stall 35 to 28Second buyer's written request in the same category, dated122.4Still below; next tranche declined at re-vote
v4, proposed by a memberRepeat buyers 15 to 20, from stall 28 to 23None attached; market described as stronger140.4Rejected: crosses a flip point without evidence

Does an editable number invite edits made to reach a verdict?

It can, and that is the strongest case for locking the numbers. A committee that wants a deal can nudge inputs until the expected value clears, and a log full of small, plausible edits can launder that. Editing also starts from the last number on the page: Tversky and Kahneman's 1974 paper in Science describes adjustment from an anchor as one of three heuristics people use in numerical prediction, efficient in general but prone to systematic errors.

Locking does not remove the judgment; it hides the next one. The controls that work are the ones above: a reallocation rule set before editing, evidence for any edit that crosses a flip point, a second signer, and the voted snapshot kept beside the current version. Add a standing rule that the member who wrote an input is not its only editor.

Accountability comes from scoring, which only a log makes possible. Glenn Brier's 1950 paper in Monthly Weather Review introduced a score for probability forecasts based on squared differences between forecasts and outcomes. Summed across the four scenarios, if the company ends in rights failure, the vote draft scores 0.990 and the version after counsel's edit 0.770, lower being better; if it ends in the one-category case, the vote draft scores 0.590 and v3 0.626. One deal proves little, but scored across many deals, the log shows whether a committee's edits add information or noise.

What could make editable probabilities mislead a committee

Five failure modes can turn an editable model into a confident-looking mistake, and each needs a counter:

At a glance

What could make editable probabilities mislead a committee

  • A missing scenario: probabilities spread over four outcomes say nothing about a fifth nobody listed, and no amount of editing repairs the set. Review the scenarios themselves at each re-vote.
  • Decimals without calibration: a number to one decimal carries information only if the committee's past numbers did. A committee that has never scored its forecasts cannot know which kind it holds.
  • Shared anchors: when every member edits from the deal lead's first draft, five editors may add only one opinion. Ask members to write their own numbers before they see the draft.
  • A movable hurdle: flip points depend on the hurdle, which is a judgment too. Log changes to the hurdle exactly as you log changes to probabilities.
  • Numbers read as forecasts outside the room: an editable model passed to a company or to limited partners can be taken as a projection. Label every copy illustrative and judgment-based, with its version.

What to ask SourceX, or any AI data company, about its own scenario numbers

Ask which planning numbers are measured rates with a stated denominator and which are judgments, and ask for the change history of any scenario figure that reaches investor materials. A company that can show who changed a number and why is easier to underwrite than one whose numbers simply appear.

SourceX's investor page does not publish company or commercial figures; it says those go to investors directly, through a materials request. That is where a committee can test them with the register and log above.

For readers weighing SourceX as an investment, its current raise is limited to verified accredited investors under Rule 506(c), and the questions below can accompany a materials request at /investors:

Checklist · 3 items

  1. Which of the figures you would share are counts with denominators, and which are estimates? Who set each estimate, and when?
  2. For any scenario or probability you present, what evidence moved it most recently, and what did it say before?
  3. Which single input would change your own plan most if it moved, and what would you need to see to move it?

0 of 3 ticked. Tick each item as you go.

Methods note: the illustrative model, its rules and its sources

Model: four scenarios with value indexes 400, 160, 40 and 0 (entry 100) and probabilities of 15, 35, 35 and 15 percent; decision rule, invest when the probability-weighted value index is 125 or more. Flip points hold the ratios among the other three probabilities fixed and solve for the value at which the expected value equals 125. Each probability was assigned to the ICD 203 band whose closed interval contains it; the range consistent with those bands was found by filling the highest-value scenarios first for the maximum and the lowest-value first for the minimum, which is exact for a linear objective with box constraints and one sum constraint. Brier scores sum squared differences between each version's four probabilities and a one-hot outcome.

All arithmetic used exact fractions in bin/xs0422_model.py in the writer's scratch folder; displayed values are rounded half up to one decimal (Brier scores to three), once. Every input is illustrative, chosen to put the vote near its hurdle, and none is a measured rate, a market figure or a SourceX number.

Sources: Kent (1964) and a later CIA review on estimative language; ODNI's ICD 203 for likelihood bands; Friedman and coauthors (2018) on rounding and accuracy; Tversky and Kahneman (1974) on anchoring; Brier (1950) on scoring; HM Treasury's AQuA Book on assumptions and decisions logs; SourceX's investor page as published on 27 October 2026. Not measured: how real committees set or edit probabilities, how often edits improve them, and any fund's or SourceX's practice.

Frequently asked questions

Should an investment committee use words or numbers for scenario probabilities?

Numbers, labeled as judgments. Words such as unlikely cover wide bands; in ICD 203, unlikely spans 20 to 45 percent, and in the article's illustrative deal the bands allow expected values on both sides of the hurdle. A 2018 study of 888,328 forecasts found that rounding numbers into verbal-style categories reduced accuracy.

What happens to the other scenarios when one probability is edited?

Their probabilities must fall by the same total, and which ones fall is a choice. In the illustrative model, raising one downside by 10 points gives expected values of 90.0, 114.0, 114.7 or 126.0 depending on where the points come from. State the reallocation rule before anyone edits.

How should a committee record changes to scenario probabilities?

Log every change with the version, editor, input, old and new values, where the moved probability came from, the evidence and the reason. Keep the voted version frozen beside the current one, and require an evidence link and a second signer for any edit that crosses an input's flip point.

What is a flip point in a scenario model?

It is the value an input can reach before the decision changes, holding the other inputs' relative weights fixed. In the illustrative deal, the best case flips the vote if it falls 1.6 points, while the one-category case has 10.8 points of room, so diligence should target the best case first.

How can a committee tell whether its probabilities were any good?

Score them once outcomes are known. The Brier score, introduced in 1950, sums squared differences between forecast probabilities and what happened; lower is better. Because one deal proves little, the value comes from scoring every logged version across many decisions and checking whether edits improved the scores.

Sources

  • Kent's 1964 essay recounts that the phrase serious possibility was read as much lower odds of an attack than Kent intended. Source
  • A CIA Studies in Intelligence review says disagreement over estimative terms, and over words versus numeric probabilities, continues within the intelligence community. Source
  • ICD 203 assigns likelihood bands: very unlikely 05-20%, unlikely 20-45%, roughly even chance 45-55%, likely 55-80%, very likely 80-95%. Source
  • Friedman, Baker, Mellers, Tetlock and Zeckhauser (International Studies Quarterly, 2018) analyzed 888,328 geopolitical forecasts and found that rounding numeric estimates into qualitative categories reduced accuracy. Source
  • Tversky and Kahneman (Science, 1974) describe three heuristics, including adjustment from an anchor in numerical prediction, which are useful but lead to systematic errors. Source
  • Brier's 1950 Monthly Weather Review paper introduced a verification score for probability forecasts, based on squared differences between forecasts and observed outcomes. Source
  • HM Treasury's AQuA Book describes an assumptions log of risk-assessed, signed-off assumptions and a decisions log recording each decision, when and why it was made, and who made and approved it. Source
  • SourceX's investor page says it shares company and commercial metrics with investors directly through a materials request. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies