The marginal cost of information retrieval has fallen to roughly zero. Every allocator now works from the same omniscient baseline, which means information asymmetry has stopped generating alpha. What generative AI added on top of that is synthetic polish: fluent prose, clean narrative arcs, and confidence uncoupled from evidence.
The problem in private capital was never bad judgment. Most of the judgment in this market is good. The problem is that it is invisible — formed in forty minutes inside one person’s head, recorded as a verdict, and unavailable for inspection the moment the meeting ends.
Capital under management scales. Analyst hours do not. When those two lines diverge, diligence quietly degrades into pattern matching, consensus fatigue, and a longer memo.
LLMs optimize for persuasion. askOdin compiles for physics.
That is the whole argument, and the rest of this piece is the evidence for it.
I. What 2,488 decks actually score
We scored 2,488 venture pitch decks with a single engine version inside one four-week window, 21 February to 19 March 2026. One version, one window, on purpose: askOdin holds a reference corpus of more than 110,000 scored documents compiled on public deal data, but that corpus spans several engine versions and scores from different versions are not comparable. Pooling them would produce a bigger sample and a meaningless median. The smaller number is the one that can be reproduced.
Here is the number.
The median Clarity Score™ was 35 out of 100. The mean was 33.2. 70.0% of decks drew a PASS verdict — the compiler’s floor, below 50. And 31.7% scored exactly zero, which we keep in the median rather than pruning, because a zero is a terminal structural finding and not a missing value. Dropping the tail would have moved the median up and made the report worthless.
Then the more interesting result. The Clarity Framework™ scores five sections out of 20 each. The spread between the strongest and the weakest is 3.7x, and it points the same direction in almost every deck:
- Business Model Physics is the weakest section in the market. 67.0% of decks scored below half marks on it, with a median of 6 out of 20. Unit economics that invert at scale, acquisition cost math that assumes a channel stays cheap forever, gross margin profiles drawn as though friction were optional.
- Problem Definition is the strongest. Only 17.9% fell below half marks, median 12 out of 20.
Read those two lines together, because the pairing is the finding. Founders describe what is broken in the world well. They describe how the company makes money badly. The eloquence is real and it is concentrated entirely on the half of the document that does not determine whether the investment returns capital.
That is not an edge case in a long tail. That is two thirds of the market, and it is the part a fluent summary is least likely to catch — because a summary reads the narrative, and the narrative is the strong half. The full method, thresholds and per-section distributions are published in Where Pitch Decks Break.
II. Summarization is not verification
The industry’s response to diligence fatigue has been to wrap a language model around the data room. It is an understandable instinct and it introduces a new fiduciary risk on top of the old one.
An LLM is a probabilistic token-prediction engine. Pointed at an operating model, a cap table, or a capitalization history, it returns the statistically plausible answer. In private capital, plausible is precisely the shape a bad allocation takes on the way past. Worse, a probabilistic model will smooth a fatal contradiction into a coherent sentence, because narrative cohesion is what it was built to produce. The contradiction was the signal. It is now gone, and nothing in the output tells you it was ever there.
There is also nothing left at the end that you can defend. You cannot audit a statistical guess. You cannot walk an investment committee, a limited partner, or a sovereign authority through a neural network’s hidden weights and call it a record.
Verification is a different operation, and it has four parts:
- Extract the natural-language claims into discrete, machine-testable assertions.
- Stress-test each assertion against the structural constraints of the business model.
- Cross-examine claims across documents — deck assertion against the actual cell in the model.
- Compile a sealed record of the reasoning, before the capital moves.
Every one of those steps has to produce an artifact. That is the difference between an answer and an audit.
III. The four-protocol compiler
AI Judgment Infrastructure™ is not a chatbot with a finance prompt and it is not an automated note-taker. It is a deterministic compiler for investment reasoning, and it runs on four patent-pending protocols.
RUNE Protocol™ is the compiler itself. It strips persuasive rhetoric, semantic inflation and storytelling out of unstructured materials and renders what survives as an auditable logic graph, propagating a Brittleness Score through every variable to the source text that asserted it. It does not summarize and it does not guess. It compiles.
U.S. PATENT PENDING 63/948,559RAVEN Protocol™ is the triangulator. It cross-examines the pitch narrative against the spreadsheet math, the cap table and the rest of a heterogeneous data room, so a number appearing three different ways in three different files becomes a finding rather than an inconsistency nobody indexed. The architectural mechanics of RAVEN’s triangulation engine are protected under U.S. Provisional Patent No. 63/994,876 and are not publicly disclosed.
NORN Protocol™ is the temporal layer. It isolates Narrative Inflation: the divergence, across chronologically sequential documents, between how confidently a venture presents and what its underlying capital efficiency is actually doing. Round over round, the deck gets better while the business gets worse, and no single document contains the evidence. — U.S. Prov. Patent No. 64/011,252
JUDGE Protocol™ is the halting condition. When compiled reasoning violates a structural constraint, JUDGE stops inference rather than producing a confident number anyway. It does not return a probability. It returns a verdict. — U.S. Prov. Patent No. 64/017,488
The terminal output is the Defensible Audit Log™: every claim tied to the page and line that asserted it, every contradiction preserved rather than silently reconciled, and the whole record anchored to SHA-256 fingerprints of the exact files submitted. Change a byte of the deck and the anchor fails loudly. That is what makes it checkable by someone who was not in the room — the architecture is documented in full.
IV. The shield: what this does not claim
A standard that overclaims is worse than no standard, so this part is stated plainly and it does not move.
askOdin audits reasoning, not truth.
We do not predict market outcomes. We do not certify that a company will succeed, and a high Clarity Score is not an endorsement or a substitute for your own diligence. Founders lie, markets turn, and a structurally immaculate thesis can still lose money. What the compiler establishes is narrower and more durable: whether the reasoning presented holds against the structural constraints it invokes, and whether anyone can check that finding later.
That narrowness is a feature, not a bug. It is the reason the output survives contact with an investment committee.
V. The standards gap
Every institutional function has acquired a governing standard except one.
Financial reporting has GAAP and IFRS. Audit has GAAS and the PCAOB. Information security has SOC 2 and ISO 27001. Debt markets went further and built an entire industry of external raters, precisely because “we looked at it carefully and we’re comfortable” stopped being an acceptable answer at scale. Credit is the useful comparison here: what a rating agency sells is not a prediction of default, it is a published methodology that lets a third party reconstruct how the opinion was formed.
Investment judgment has none of this. It has each firm’s private convention and a memo. Every fund improvises its own internal standard for what counts as sufficient reasoning, and none of those standards are auditable across firms, across LPs, or across years.
That was survivable when the constraint was information. It is not survivable now that information is free, narrative is machine-generated, and the only remaining scarce input is the rigor applied to the logic. No enterprise closes its books on an unstandardized ledger. Private capital allocates billions on one.
Judgment is the last unscalable asset. Infrastructure is how an unscalable asset gets scaled — and there is a version of this argument aimed squarely at GPs, and another about why unverifiable rigor is indistinguishable from none.
Venture capital is the last unaudited asset class. askOdin provides the infrastructure to close the gap.
The Theory is now Infrastructure.
We have moved from thesis to execution. The framework is running live on our platforms.
Choose your path: fix your narrative, or scale your judgment.
This essay builds on the SSRN working paper “The Last Mile of AI: Judgment Infrastructure, Defensible Audit Logs, and the End of Information Retrieval” (DOI 10.2139/ssrn.6664200). Empirical figures are drawn from Where Pitch Decks Break, n = 2,488, scored 21 Feb – 19 Mar 2026.
U.S. Provisional Patents Pending: 63/948,559 · 63/994,876 · 64/011,252 · 64/017,488.