Every hypothesis Encounter has published, with its confidence and impact ratings, the evidence at the time it was made, and what happened when it was tested.
Predictions are locked with timestamps before they are testable. Confirmed predictions are kept on the page. Refuted predictions are kept on the page too. Each prediction is committed to the page before the answer is known, so the trial readout that arrives later settles the call.
Encounter runs a hypothesis engine that walks the framework's structural model: a graph of cycles, regimes, positions, and failure modes. The walker proposes candidates. For each position in a cycle, for each failure mode that position can suffer, it names which known intervention is structurally positioned to act.
A structural filter throws out candidates that don't fit. The intervention's class has to match the position's regime, the failure has to belong to that regime's recognised failure set, every cited rule has to resolve in the canon. The point is to leave only structurally sound claims to spend evidence retrieval on.
For every surviving candidate, the engine queries PubMed, ChEMBL, and ClinicalTrials.gov live. It records how many papers and trials exist for the drug + indication, whether ChEMBL agrees on the mechanism, and which active trials are most relevant. The evidence is frozen at retrieval time so the hypothesis stays re-readable later.
Three scores in [0,1]: structural (how cleanly the candidate fits the framework), evidence (what the retrieval returned), novelty (how far it is from things already in the address graph or in published clinical use). A weighted composite plus 1-5 confidence and impact ratings let us triage the daily output.
The engine generates more candidates than Encounter can chase. The top-ranked ones are reviewed; those that hold up under operator scrutiny are published here, with all the inputs (substrate, mechanism, evidence, score) preserved so the prediction can be re-checked when the trial reads out.
Dated, falsifiable calls on real combination trials: does the added agent contribute over its backbone? The call comes from the structural contribution operator. Open predictions are locked before the trial reads out.
Track record: 7 read out · 6 confirmed · 1 refuted. These were not locked before their readouts — they are the operator's record, shown so the open calls below can be weighed against it.
redundant: both address failure_threshold_drift at position_p10_1 (second brake at the same station)
Falsifier: The combination beats the backbone by a clinically meaningful margin on overall survival / pfs.
distinct target (failure_sub_threshold_loading at position_p10_1, different failure mode at the same station) but its rate-limiting role in this substrate is not confirmed; no benefit predicted (a different target is necessary, not sufficient)
Falsifier: The combination beats the backbone by a clinically meaningful margin on progression-free survival.
distinct target (failure_sub_threshold_loading at position_p10_1, different failure mode at the same station) but its rate-limiting role in this substrate is not confirmed; no benefit predicted (a different target is necessary, not sufficient)
Falsifier: The combination beats the backbone by a clinically meaningful margin on relapse-free survival.
distinct target (failure_sub_threshold_loading at position_p10_1, different failure mode at the same station) but its rate-limiting role in this substrate is not confirmed; no benefit predicted (a different target is necessary, not sufficient)
Falsifier: The combination beats the backbone by a clinically meaningful margin on progression-free survival.
Marginal PFS win; approved (Opdualag), OS n.s. The conservative no-benefit call MISSES this one win — the false-negative cost of requiring a confirmed bottleneck.
redundant: both address failure_threshold_drift at position_p10_1 (second brake at the same station)
Falsifier: The combination beats the backbone by a clinically meaningful margin on overall survival.
OS negative; tiragolumab added nothing. Roche later scrapped the programme.
distinct target (failure_sub_threshold_loading at position_p10_1, different failure mode at the same station) but its rate-limiting role in this substrate is not confirmed; no benefit predicted (a different target is necessary, not sufficient)
Falsifier: The combination beats the backbone by a clinically meaningful margin on overall survival.
LAG-3 fixed-dose combo did not improve OS vs standard of care in MSS mCRC (immune-excluded — LAG-3 not the bottleneck). Conservative no-benefit call CONFIRMED; the old optimistic rule would have refuted.
redundant: both address failure_threshold_drift at position_p10_1 (second brake at the same station)
Falsifier: The combination beats the backbone by a clinically meaningful margin on overall survival.
Discontinued for OS futility (Apr 2025); BeiGene scrapped TIGIT. Fifth independent anti-TIGIT programme to fail. Redundant call confirmed.
distinct target (failure_sub_threshold_loading at position_p10_1, different failure mode at the same station) but its rate-limiting role in this substrate is not confirmed; no benefit predicted (a different target is necessary, not sufficient)
Falsifier: The combination beats the backbone by a clinically meaningful margin on relapse-free survival.
No RFS benefit (HR 1.01) — same drug/node as RELATIVITY-047 but adjuvant melanoma has reduced LAG-3+ density, so LAG-3 is not the bottleneck here. The natural experiment. Conservative call CONFIRMED; the old rule would have refuted.
redundant: both address failure_threshold_drift at position_p10_1 (second brake at the same station)
Falsifier: The combination beats the backbone by a clinically meaningful margin on overall survival.
Discontinued for futility vs nivo+chemo (Dec 2025). Fc-silent TIGIT also failed — not an Fc artefact.
distinct target (failure_sub_threshold_loading at position_p10_1, different failure mode at the same station) but its rate-limiting role in this substrate is not confirmed; no benefit predicted (a different target is necessary, not sufficient)
Falsifier: The combination beats the backbone by a clinically meaningful margin on progression-free survival.
Missed PFS significance vs pembrolizumab (5.1-mo numerical gain, n.s.). The conservative no-benefit call is CONFIRMED here.
T-VEC (oncolytic lysis with GM CSF) acts at the antigen release step of the cancer-immunity cycle. The failure mode is threshold drift: threshold moves out of reach of normal loading current. Capacitor sometimes fires…
pembrolizumab (pd1 brake release) acts at the cytolytic synapse; biopsy-revealed molecular profile step of the cancer-immunity cycle. The failure mode is throughput drop: factory's rate of work falls. Fewer units prod…
tremelimumab (ctla4 brake release) acts at the activation barrier; commitment threshold step of the cancer-immunity cycle. The failure mode is setpoint loss: homeostat loses memory of its setpoint. System drifts witho…
bevacizumab (vegf blockade clearing supply line) acts at the initiation (C4); first constraint engaged step of the cancer-immunity cycle. The failure mode is sub-threshold loading: loading current never raises charge…
nemvaleukin alfa (il2 alpha bypass for treg avoidance) acts at the architecture (C1); pathway selection step of the cancer-immunity cycle. The failure mode is differentiation skew: factory output distribution shifts a…
oxaliplatin (immunogenic chemotherapy) acts at the antigen release step of the cancer-immunity cycle. The failure mode is threshold drift: threshold moves out of reach of normal loading current. Capacitor sometimes fi…
sargramostim (myeloid lineage support) acts at the DC capture / migration; surveillance gathering step of the cancer-immunity cycle. The failure mode is differentiation skew: factory output distribution shifts away fr…
sargramostim (myeloid lineage support) acts at the DC capture / migration; surveillance gathering step of the cancer-immunity cycle. The failure mode is loading dock backup: built outputs cannot egress. Backs into the…
sargramostim (myeloid lineage support) acts at the DC capture / migration; surveillance gathering step of the cancer-immunity cycle. The failure mode is low yield: per-unit yield falls. Many starts, few finishes. Load…
sargramostim (myeloid lineage support) acts at the DC capture / migration; surveillance gathering step of the cancer-immunity cycle. The failure mode is throughput drop: factory's rate of work falls. Fewer units produ…
STING agonist (sting pathway agonism) acts at the antigen release step of the cancer-immunity cycle. The failure mode is sub-threshold loading: loading current never raises charge to threshold. Capacitor never fires;…
doxorubicin (immunogenic chemotherapy) acts at the antigen release step of the cancer-immunity cycle. The failure mode is threshold drift: threshold moves out of reach of normal loading current. Capacitor sometimes fi…
Oncolytic virus (lytic antigen release) acts at the antigen release step of the cancer-immunity cycle. The failure mode is threshold drift: threshold moves out of reach of normal loading current. Capacitor sometimes f…
STING agonist (sting pathway agonism) acts at the antigen release step of the cancer-immunity cycle. The failure mode is threshold drift: threshold moves out of reach of normal loading current. Capacitor sometimes fir…
TLR agonist (tlr pathway agonism) acts at the antigen release step of the cancer-immunity cycle. The failure mode is threshold drift: threshold moves out of reach of normal loading current. Capacitor sometimes fires,…
Position 4 of the cancer-immunity cycle is the commitment threshold. A T cell either crosses it and becomes a fully committed anti-tumour effector or holds back. The decision is gated by the CD28/CTLA-4 balance at the…
Get in touch.