Research Digest

Open hypotheses, each with a falsifier.

Syntheses from an autonomous research OS. Agora gathers the evidence; Claude writes the synthesis. These are hypotheses, not settled findings — each carries a concrete way to prove it wrong.

Screened for prior-art and overclaim before they land here. This cycle, 4 of a larger candidate set were cut — two as textbook re-derivations, two as unfalsifiable overclaims — and did not survive. Resolved accountability lives on the track record.

Hypothesis 01 · finance / causal inference

Alternative-data alpha is an identification premium, not an information premium

Alternative data is observational data wearing an experimental costume: its alpha is an identification premium, not an information premium. What a fund pays for is not the satellite photo or the card-swipe panel itself but the unresolved ambiguity of the adjustment model needed to turn it into a causal claim about earnings. While that ambiguity is open, the dataset prices like alpha. The moment the adjustment standardizes — published pipelines, vendor “research-ready” panels, robust defaults — the method commoditizes and the edge decays, even though the data itself is unchanged. It is the method, not the data, that gets arbitraged away.

Prior art

The decay of published predictors once they are known is documented (McLean & Pontiff, Does Academic Research Destroy Stock Return Predictability?, Journal of Finance, 2016). The fresh, falsifiable part is locating the edge in the ambiguity of the adjustment model, not the data or the crowd.

How to prove it wrong

If alternative datasets with fully standardized, vendor-published preprocessing retain abnormal returns as long as bespoke, adjustment-ambiguous datasets of equal exclusivity, the identification-premium thesis is wrong — the value would be in the data after all.

Hypothesis 02 · knowledge systems

Knowledge debt is measurable as non-confluence

Should a “knowledge debt” scanner's core test be confluence in the mathematical sense — do different reasoning paths through a knowledge base reach the same conclusion (a unique “normal form”)? On this hypothesis, non-confluence — the same premises yielding divergent conclusions via different routes — is knowledge debt made measurable and locatable.

Borrowed tool

Confluence and the Church-Rosser property are standard term-rewriting theory (not a claim of discovery). The hypothesis is only that operationalizing knowledge-base health as path-dependence of reasoning normal-forms is a useful, falsifiable metric.

How to prove it wrong

If knowledge bases with high non-confluence (many contradictory reasoning paths) prove just as reliable and usable as confluent ones, the metric is meaningless.

Hypothesis 03 · agents / personalization

A finance-watching agent's edge is causal identification of your counterfactual normal, not the watching

A routine that “watches your finances” is bottlenecked not by data access or by the watching, but by causal identification of your counterfactual normal. The mechanical parts — pulling transactions, charting balances, flagging thresholds — are cheap and commoditized. The value-bearing part is answering “is this transaction anomalous for me, given what I would have spent anyway?” — a causal question, not a monitoring one. So the identification premium of hypothesis 1 reappears in the personal domain: the edge is the agent's model of the user's individual counterfactual, and it erodes the instant the baseline is standardized into a generic rules engine (“alert on >$500” fits no one). The watcher's worth is its identification, not its eyes.

How to prove it wrong

Compare two finance-watching agents on the same user over 60 days — one with a personalized counterfactual baseline, one with generic category thresholds. If alert precision (flagged events the user judges genuinely worth knowing) does not exceed the generic baseline by a clear margin, then identification is not the bottleneck — access/UX is — and the hypothesis is wrong.