Pareto-Optimum Endpoint Reduction — Cross-Oracle Report

Pareto-optimum reduction is the modelled best-case ceiling for each endpoint, composed by that oracle's own combiner (correlation-shrunk). Relative reduction = 1 − combined HR.

Two figures are reported. Optimum is a greedy forward selection that adds an intervention only while it improves the endpoint — the true achievable ceiling. All-in turns on every candidate at full dose, which is what the prior edition of this report tabulated. Where the two differ, the model is non-monotone: adding a beneficial intervention makes the modelled endpoint worse.

Every number is the oracle's own engine output, executed under jsdom — 17 via each dashboard's ATLAS_PARETO adapter, 39 via the newer series' combinedFD() combiner, and the all-cause-mortality oracle via its computeModel() directly.

Units differ by oracle — see the Unit column. Most report % relative reduction in a mortality or event endpoint. Osteoarthritis and BPH report % of maximum modelled symptom effect or relief, where 100% means the full modelled effect, not symptom elimination; those two are therefore not comparable with the mortality oracles and should not be read as the "best" performers. Osteoarthritis returns 100.02%, marginally above its own ceiling — a rounding/saturation artifact in that engine worth a separate fix.

Caveat on every figure. These are upper-bound ceilings, not evidence-graded expectations. Each assumes every effect is real, simultaneous, fully adhered, and correctly de-correlated by the engine's own discount. Units are % relative reduction in the oracle's stated endpoint; a few oracles report % of maximum modelled symptom effect rather than mortality, in which case 100% means full modelled effect, not elimination.

Sixteen oracles were newly wired for this edition via their own engines — hypertension, multiple sclerosis, schizophrenia, tuberculosis, hearing loss, HIV, immunology, dementia, cachexia and rare-disease targeting — taking coverage from 57 to 73. Four of the remaining entries are not intervention oracles at all. Every remaining oracle with a combinable endpoint is now wired; generic extraction returned zero or NaN for them, and no figure is published rather than an unverified one.

Note on precision: the six newly wired oracles are read from their rendered metric elements, which round to two decimal places, so their figures carry ±0.5 point of display precision rather than full engine precision.

• Atlas Coverage of US Mortality — Deduplicated Count atlas_us_deaths_count_dashboard.html — not an intervention oracle — no intervention set to optimise (calculator / harness / count dashboard). Correctly out of scope.
• Life Expectancy Shift Calculator life_expectancy_shift_calculator.html — not an intervention oracle — no intervention set to optimise (calculator / harness / count dashboard). Correctly out of scope.
• Longevity Fiscal Impact Model longevity_fiscal_impact_model.html — not an intervention oracle — no intervention set to optimise (calculator / harness / count dashboard). Correctly out of scope.
• MIMIC-III Mortality Harness mimic_mortality_harness.html — not an intervention oracle — no intervention set to optimise (calculator / harness / count dashboard). Correctly out of scope.
• Peripheral Neuropathy Oracle peripheral_neuropathy_oracle.html — no combiner exists — the page is an evidence catalogue whose Pareto control parses RR/OR out of text and highlights items above a cutoff; it never combines interventions into a single endpoint, so no combined optimum is defined
• Peripheral Neuropathy Oracle peripheral_neuropathy_oracle.html — not a combinable intervention oracle — effect sizes are stored as mixed-unit strings (risk differences, RR, NNT, SMD) with no common combiner, so no Pareto optimum is defined

This oracle appears in the coverage list as having no combinable endpoint, and this appendix is what it has instead. Only 5 of 17 of its interventions carry a ratio-scale estimate; the rest are annualised risk differences, standardised mean differences, NNTs, bare response rates, or narrative verdicts. More decisively, the ratios that do exist are responder ratios above 1 (RR 1.73 = 1.73× as likely to reach ≥50% pain reduction), not hazard ratios below 1, and they sit on different endpoints in different populations — incident neuropathy in T1DM, pain response in diabetic neuropathy, disability-grade improvement in Guillain–Barré, relapse in CIDP.

Composing them would give a number with no referent, so no combined optimum is published. Interventions are stratified by population and ranked within stratum only: rows are comparable within a block, never across blocks, and none of these figures belong in the main table above.

The concentration result — a median of 2 interventions reaching 80% of the ceiling — is driven by the distribution of effect sizes, not by any mortality model. It reflects that each oracle's catalogue is dominated by two or three large, well-evidenced effects and a long tail of small ones. It is not evidence that the tail is worthless: a small effect that is cheap and safe may still be worth adopting, and the greedy ordering is not a clinical recommendation.

The optimum figures inherit every assumption of the underlying oracle, including its ρ correlation discount and its saturation curve — the same machinery responsible for the non-monotonicity above. Until that is fixed, treat the optimum column as the better of two flawed estimates rather than as a corrected number.

Finally, these are cross-sectional ceilings on heterogeneous endpoints. They are not comparable to one another as health impact: 90% of a symptom score and 90% of mortality are different quantities, and no weighting across oracles is implied.