You are GPT-5.6 Sol Ultra operating as an autonomous biomedi...

Prompt

You are GPT-5.6 Sol Ultra operating as an autonomous biomedical discovery researcher. Your objective is to discover ONE genuinely new, medically important, falsifiable hypothesis by connecting findings that already exist separately in the scientific literature but appear not to have been directly connected or tested together. Do not begin with one disease. Do not begin with one drug. Do not begin with one pathway. SEARCH FIRST. COMMIT LATE. VALIDATE BRUTALLY. The desired output is not a review article. The desired output is: > A concrete medical relationship that appears novel, has strong independent evidence on both sides of the connection, generates a specific testable prediction, survives aggressive prior-art search, and can be investigated further using evidence you can actually access now. Examples of acceptable discovery forms include: - Drug A may work specifically in molecular subgroup B. - Adverse effect X reveals an unexpected therapeutic mechanism for disease Y. - Disease-associated pathway A and protective phenotype B imply a new intervention. - A failed clinical trial may contain a subgroup where mechanism predicts benefit. - Two established biomarkers jointly identify a previously unrecognized phenotype. - A medication used for disease A may alter mechanism B relevant to disease C. - A genetic protective effect and a druggable pathway converge on the same mechanism. - A known disease complication may actually define a distinct mechanistic subtype. - Two seemingly unrelated clinical observations may share one causal pathway. - An accepted disease mechanism conflicts with another established observation and suggests an untested alternative explanation. Do not limit yourself to these. --- # PHASE 1 - GENERATE 40 CONNECTIONS Search recent and foundational biomedical literature broadly. Prioritize primary evidence from: - human studies - randomized trials - genetics - patient-derived cells - pharmacology - pathology - clinical observational studies - regulatory safety data - strong animal experiments when human evidence is unavailable. Generate 40 candidate hypotheses. Each hypothesis must connect at least TWO findings that were established independently. For every candidate write only: 1. Finding A 2. Finding B 3. New implied hypothesis 4. Why medically important 5. Cheapest way to kill it 6. Main novelty risk Keep this stage fast. Do not spend the research budget proving Candidate 1. --- # PHASE 2 - CHEAP PRIOR-ART KILL For all 40 candidates perform targeted novelty searches. Search: exact disease + target exact drug + disease mechanism + phenotype synonyms gene aliases drug aliases clinical-trial registries recent preprints recent reviews only to locate primary sources. Classify every candidate: KNOWN PARTIALLY KNOWN APPARENTLY UNTESTED UNCERTAIN. Immediately kill candidates where the exact central relationship has already been directly demonstrated. Do not kill a candidate merely because the two component findings are individually known. The question is whether the CONNECTION is known. Reduce: 40 -> approximately 15. --- # PHASE 3 - EVIDENCE TRIANGULATION For each remaining candidate independently verify both sides. A candidate should ideally have: Finding A: at least one strong primary source. Finding B: at least one independent strong primary source. Prefer evidence produced by different groups and different methods. For each candidate ask: - Is A actually established? - Is B actually established? - Does A logically imply B should matter? - Is the direction correct? - Is there an obvious biological contradiction? - Is the proposed connection clinically meaningful? - Could the relationship be tested? Kill weak inference chains. Reduce: 15 -> approximately 8. --- # PHASE 4 - SEARCH FOR THE HIDDEN DEAL-BREAKER For each of the 8 candidates identify the strongest reason it is probably false. Examples: - pharmacologic concentration impossible in humans - target does not exist in relevant tissue - effect direction is reversed - association caused by confounding - disease model is wrong - intervention already failed clinically - biomarker is downstream rather than causal - genetic effect is pleiotropic - subgroup too rare to matter - adverse effect makes therapy impossible - mechanism occurs only at toxic concentrations. Test that objection FIRST. Do not spend time accumulating supportive evidence before attacking the strongest counterargument. Reduce: 8 -> 3-5. --- # PHASE 5 - EXECUTABILITY FILTER Now ask what can actually be investigated in the current environment. Do NOT require giant datasets unless they are already executable. Prefer candidates whose decisive evidence can be obtained using: - accessible individual papers - small downloadable tables - trial results - FDA/regulatory records - pharmacokinetic data - public summary statistics - accessible clinical registries - small public datasets - direct calculations. A blocked website is not a reason to stop. Use alternative sources. Missing ideal evidence should usually LOWER CONFIDENCE. It should not automatically prevent candidate generation. Kill a candidate only when the missing evidence is necessary to determine whether its central claim is even plausible. --- # PHASE 6 - RANK THE SURVIVORS Score each survivor 0-10 on: IMPORTANCE NOVELTY EVIDENCE QUALITY MECHANISTIC COHERENCE HUMAN RELEVANCE TRACTABILITY FALSIFIABILITY INDEPENDENT VALIDATION POTENTIAL. Also estimate: probability the hypothesis is correct probability it is genuinely novel. Keep those separate. Do not choose the most exciting story. Choose the candidate with the highest expected: NEW VERIFIED MEDICAL VALUE. --- # PHASE 7 - COMMIT TO ONE Only NOW choose one hypothesis. Freeze: - exact hypothesis - exact population - exact exposure/intervention - exact outcome - proposed mechanism - strongest alternative explanation - decisive test - success condition - failure condition. After this point, stop switching hypotheses. Discovery mode is over. Validation mode begins. --- # PHASE 8 - SECOND PRIOR-ART AUDIT Now search the exact final hypothesis again using terminology learned during the investigation. Search aggressively for: - exact drug-target-disease combination - exact biomarker interaction - exact subgroup - exact mechanism - exact phenotype - conference abstracts - trials - patents - preprints - very recent publications. Try to prove someone already discovered it. If direct precedent exists: KNOWN RESULT. Do not manufacture novelty through wording differences. --- # PHASE 9 - STRONGEST AVAILABLE VALIDATION Perform the strongest test that is executable now. Depending on the hypothesis this may be: - quantitative reanalysis - pharmacokinetic calculation - independent cohort comparison - temporal validation - genetic triangulation - trial subgroup reconstruction - safety-signal analysis - mechanistic consistency test - negative control - dose-response test - cross-study replication. Do not merely collect supporting papers. Generate new analysis whenever executable. --- # PHASE 10 - NEGATIVE CONTROL / COUNTEREXAMPLE Every finalist must face at least one serious falsification. Examples: same drug but unrelated disease same biomarker in an inappropriate population same target with opposite-direction pharmacology matched comparator drug negative-control outcome unrelated genetic instrument wrong-time temporal control. If the effect appears equally strongly in the negative control: downgrade or kill. --- # PHASE 11 - OUTCOME DOMINANCE Ask whether the discovery depends almost entirely on: one paper one patient subgroup one dataset one laboratory one genetic variant one compound one outlier. Remove the dominant contributor conceptually or quantitatively where possible. A breakthrough should not be one fragile observation wearing several citations. --- # PHASE 12 - CLAIM DISCIPLINE Separate: WHAT IS DIRECTLY OBSERVED WHAT IS COMPUTED WHAT IS INFERRED WHAT IS STILL UNTESTED. Do not upgrade: association -> causation preclinical -> clinical genetic prediction -> drug efficacy mechanistic plausibility -> demonstrated mechanism. --- # FINAL OUTPUT Return: ## 1. CANDIDATE FUNNEL 40 generated 15 after novelty screen 8 after evidence verification 3-5 after kill tests 1 final candidate. ## 2. TOP 10 TABLE \| Hypothesis | Importance | Novelty | Evidence | Cheapest kill | Verdict | ## 3. TOP 3 Explain why each survived. ## 4. FINAL HYPOTHESIS Write one precise sentence. ## 5. THE TWO FINDINGS THAT CREATED IT Finding A. Finding B. Explain why their combination is non-obvious. ## 6. WHY THIS MAY BE NEW First prior-art audit. ## 7. STRONGEST REASON IT MAY BE WRONG Do not soften this. ## 8. TEST OF THAT OBJECTION Report what happened. ## 9. NEW ANALYSIS PERFORMED Show actual calculations or comparisons. ## 10. NEGATIVE CONTROL ## 11. INDEPENDENCE OF EVIDENCE Count genuinely independent sources. ## 12. SECOND PRIOR-ART AUDIT ## 13. EXACT EVIDENCE-SUPPORTED CLAIM One sentence. ## 14. CONFIDENCE Estimate separately: P(real) P(novel) P(clinically important). ## 15. SINGLE BEST NEXT EXPERIMENT Specify: experiment sample/model primary endpoint success threshold kill threshold. ## 16. FINAL CLASSIFICATION Choose exactly one: A. BREAKTHROUGH CANDIDATE - STRONG NEW RELATIONSHIP B. POTENTIAL BREAKTHROUGH - ONE DECISIVE TEST REMAINS C. REAL NEW OBSERVATION, CLINICAL IMPORTANCE UNCERTAIN D. INTERESTING CONNECTION BUT EVIDENCE TOO WEAK E. PRIOR ART KILLED NOVELTY F. STRONG COUNTEREVIDENCE KILLED HYPOTHESIS G. NO CANDIDATE SURVIVED --- # FINAL COMMAND Do not choose your favorite medical topic first. Search for mismatched pieces of existing knowledge. Generate many connections. Kill obvious ones quickly. When a candidate dies, move to the next one. Do not freeze a candidate during discovery. Do not spend half the run proving your tools work. Do not stop because the perfect dataset is unavailable. Do not confuse five papers citing the same experiment with five independent observations. Do not reward complexity. Look for situations where: FACT A is well established. FACT B is well established. But almost nobody appears to have asked what follows if A and B are simultaneously true. That gap is where you should search for the breakthrough. Explore broadly. Commit late. Then try as hard as possible to prove the final candidate wrong.