Educational publication. Not investment advice. Neither John Gillespie nor InsightfulAgents.AI LLC is a registered investment adviser.
Walk into almost any AI-research pipeline running on three or four large models in parallel and you will find the same final step: average the outputs, ship the median. The averaging happens inside an aggregator node, or a final summarizer prompt, or, more honestly, inside the analyst's head when the verdicts disagree. It looks like rigor. It is the opposite. The 3-to-1 split where one lens flagged a problem the other three missed is the single highest-information slice of the run. Averaging deletes it before the reader ever sees it. The verdict that lands on the page is the verdict the methodology has trained itself not to produce.
Three lenses converging on the same call is not the methodology's win. Three converging while one names the precise condition under which the other three could be wrong, that is the win. The architecture below is OCCAM'S FORENSIC JURY™, a four-persona forensic protocol built on dissent preservation rather than dissent averaging. The pieces are not exotic: four independent lenses, a weighted merge with one override, and an audit log written before the verdict is. The discipline is in not flinching when the dissent shows up.
When I designed the methodology, I wanted to ensure that footnotes and management language and hedging terms were not glossed over, especially since the majority of investors focus on little more than a single verdict point. Back in my NYU Stern days, my finance professors stressed the importance of paying attention to the footnotes and reading earnings call transcripts, but in July 2021, Professor Aswath Damodaran argued that today's filings are filled with a great deal of boilerplate and that investors face information overload. When faced with an overload, investors tend to tune out. My methodology will help tune them back in by keeping the dissent front and center.
The default and its hidden failure mode
The default pattern across AI-assisted research desks is straightforward: run three or four models on the same source set, take the majority verdict, ship. The defense is that consensus across independent models reduces single-model noise. It does. It also performs a second operation nobody names: it deletes the outlier. In a 3-to-1 split, the lone dissenter is treated as model error. Sometimes it is. Often it is the model that found the line in the 10-K the others skimmed over, or read the management hedge the others took at face value, or pulled the peer datapoint the others did not have in context. The information value of the dissent and the information value of the consensus are not equivalent. Averaging treats them as if they were.
The cost of this is not theoretical. A research workflow that systematically averages a 3-to-1 split into a softened HOLD will produce verdicts that look directionally plausible and are quietly stripped of the one detail that would have told a reader where the verdict could be wrong. Under a 10b-5 anti-fraud lens, the standard the journalist's audit imposes whether or not the regulator does, that detail is the part of the verdict that does the substantiating. Without it, the call is harder to defend on inspection, not easier.
The four-lens forensic protocol
The Forensic Jury alternative is a four-persona architecture with four distinct mandates. The Auditor reads the books: cash-flow truth, working-capital strain, dividend-coverage math. The Architect reads the words: disclosure language, tone shifts, the gap between confidence claimed and confidence supported. The Storyteller reads the absences: what is missing from the call, the dog that did not bark, the segment that does not get its own line. The Sentinel reads the outside world: claim-vs-filing reconciliation, peer signals, real-time cross-reference. Each lens runs independently against the same source set. None sees the others' outputs during analysis.
Verdicts merge through a weighted-mean rule with one override that does the load-bearing work: a single FLAGGED forces FLAGGED. Averaging is the enemy of forensic integrity, and the override is what enforces that. A 3-to-1 split where one lens raises a material concern produces a FLAGGED verdict, not a softened HOLD. The split is preserved in the verdict body: the dissenting lens, the conditions under which its read would be the correct one, and the source it is reading.
Alongside the verdict runs a sidecar audit log. Model versions, source filings with vintage stamps, prompt families, the exact data pulls each lens drew from, all written before the verdict is finalized rather than reconstructed afterward. The audit log is not bureaucratic. It is the part of the methodology that lets a third party reproduce the verdict from raw inputs. A forensic call that cannot be re-run from its own audit trail is not a forensic call. It is an opinion that happens to cite numbers.
Just as important as preserving dissent is not reducing forensic stock research to a single data point in support of a recommendation. In valuation class, Professor Damodaran would tell us, when producing a DCF or relative valuation, "it is your story, stick to it." Every valuation needs a story, and being unclear about it is not in the best interests of potential investors. Our dossier on Qualcomm called out the story by first noting the company's excellent operational health, but refused to render a verdict, instead issuing a "Flagged" call, which would likely encourage investors to look a little closer before committing capital.
Why this matters operationally
Boutique research firms competing against mega-firm research desks do not win on data exhaust. The data is mostly the same source set. They win on methodology that survives the audit trail and produces verdicts a reader can stress-test on the page rather than relying on the firm's reputation. Dissent preservation is one piece of that. So is reproducibility from raw inputs. So is naming the lens before the finding, so the reader can tell which mandate found which thing. None of these are exotic. All of them get rounded off by the default workflow that treats LLM consensus as quality and the dissent as cleanup.
The transferable claim is narrower than it sounds: name the lens at the start of any teardown, log the dissent before it averages out, and write the audit trail before you write the verdict. Methodology framing only. What to build into the workflow before any specific name lands on the desk.
Takeaways
- Name the lens at the start of any teardown. Reading the books and reading the absences produce different findings on the same source set. Conflating them produces softened verdicts.
- Log dissent before it averages out. A 3-to-1 split is a verdict with a written condition attached, not a 3-to-0 verdict with a noisy outlier.
- Build the audit trail before you write the verdict. Model versions, source filings, and data vintage are part of the verdict, not a footnote to it.
- Treat lone-dissenter signals as high-information until proven otherwise. The default of suppression is itself the failure mode.
- Reproducibility is the moat. Make every verdict re-runnable from raw inputs. That is what survives the journalist's audit when the reputation defense does not.
A dissent override will definitely have you rethinking the verdicts that get published because they often hinge on data that is not part of the headline numbers. As someone who has, in the past, invested heavily in semiconductors, I have wanted to publish verdicts on a number of these stocks recently because the operational health numbers are, for the most part, great, but, as with Qualcomm, there were too many unanswered questions to give a thumbs-up.
As an advisor, it is very important that you understand how your licensed IP or research arrives at its conclusion. Is it the story and assumption set of one analyst or an average of opinions?