An instrument for surfacing mechanistic hypotheses that a field’s dominant consensus has discounted — not to declare them right, but to give a human expert enough verified evidence that they are worth a look.
The target is not the neglected cold periphery of biology — the places no one looked. It is the opposite: indications where enormous capital and attention have been spent, repeatedly, down paths that failed, and where that very history has hardened into a consensus that now functions as a filter on what gets funded and published.
In these fields the obstacle is not absence of attention. It is that attention is correlated. The field’s prior is concentrated, and concentration suppresses the alternatives. Billions are spent re-running variants of the dominant hypothesis, while heterodox mechanisms struggle for oxygen — not because they were disproven, but because they were not endorsed.
The distortion this engine fights is social, not informational. The heterodox hypotheses usually already exist — published, then starved of follow-up. The scarce ingredient is rarely the idea. It is the permission to spend time on an unfashionable one.
For consensus-distorted indications, an LLM agent’s usual weaknesses invert into assets. fred has no career, no grant cycle, and no disciplinary loyalty. Against a pathology that is fundamentally about social pressure on what may be pursued, that indifference is the central capability rather than a limitation.
A human researcher carries the consensus whether they like it or not — careers, grants, and peer review are downstream of it. fred can assess a shut-off hypothesis on its merits without the social penalty that makes a human discount it. The distortion is social; fred is immune to the social part.
Billions in failed journeys leave a large, machine-readable trail. A field under consensus pressure systematically under-reads its own failures, attributing them to execution (“wrong patients,” “dosed too late”) rather than to the hypothesis class. fred can read the failure corpus as a corpus and ask whether the pattern fits “right idea, executed badly” or “the idea class is wrong” — the move the consensus is structurally unable to make about itself.
The valuable hypotheses are usually already published and then abandoned for non-scientific reasons. fred’s job is to find that starved work, assess whether the abandonment was evidentiary or sociological, and re-present the survivors as defensible, evidence-anchored writeups. It lowers the activation energy of looking at an unfashionable idea.
fred does not need to be right. It needs to produce, for a discounted mechanism, enough verified evidence that a competent human should spend an hour deciding whether it deserves more. A warrant for attention, not a conclusion. The human expert remains the final filter on every path.
The architecture enforces one boundary above all: the quantitative, consensus-measuring work runs on structured data with no LLM involvement, and fred is invoked only at the back of the pipeline to generate and verify warrants. “Cold” and “unsolved” are measured, not imagined. Tap or hover any stage for detail.
“Reasonable mechanism plus some evidence” is precisely the shape of a fluent hallucination. The difference between a warrant worth a look and a confabulation is not tone — both sound reasonable — it is the type of evidence attached. A warrant has three load-bearing parts:
These dominate the design and are stated without optimism. None are fully solved by the architecture above.
Everything rests on the agreement rate between fred’s “worth a look” and a competent reviewer’s “yes, worth a look.” If precision is high, this is a powerful instrument. If it is low, it is a plausible-text generator that costs more review time than it saves. That precision is not yet measured.
fred’s unique value for consensus-distorted unsolved indications is that it is a prior-agnostic re-reader of a field’s own record. It treats the expensive failure history as evidence rather than noise, assesses whether abandoned hypotheses were killed by data or by sociology, and re-surfaces the survivors without the career incentives that make humans discount them. The edge is immunity to the social mechanism that creates the distortion — not superior biology.
That same prior-agnosticism is double-edged: an instrument with no prior will validate heterodoxy as readily as orthodoxy, because it has no independent footing to judge either. The value is realized only if fred can distinguish evidentiary from sociological abandonment on cases where the answer is already known. If it can, this is a consensus-auditing instrument. If it cannot, it is a very articulate machine for confirming whatever the operator already believed.