The scientific literature is a balance sheet of capital assets with unrealized residual value. Thousands of sound results were abandoned not because they were wrong, but because some binding constraint of their era — compute, instrumentation, cost, data, or attention — made them impractical. Many of those constraints have since lifted. ORION is the machinery that finds those results, prices them, and decides which are worth re-opening.
A discovery can be correct and still fail to produce value, because the world of its time could not act on it. That failure is not one social story — it decomposes into distinct, measurable causes: the instrument did not exist, the computation was intractable, the input data had not been collected, the cost curve was wrong, the adjacent field had not matured, the result was published where nobody relevant read it, or the author died before follow-up.
Each of those causes has a different half-life. Some have expired. Where a cause has expired and the underlying claim is still sound, there is recoverable value sitting unclaimed in the public record. ORION's job is to identify that condition at scale and rank what to do about it — a portfolio that finances its own next cycle.
The two compose. This site describes the evaluator.
Expected value here is a product, not a sum. A candidate's worth is the chance the science is real, times the chance modern capability actually unlocks it, times the chance nobody else has already done it, times the benefit if it works — divided by what it costs to find out. Scoring these as independent weighted criteria and adding them produces rankings nobody believes: under an additive form, a candidate that is known to be discredited can still place in the top band. ORION is multiplicative by construction, so a zero anywhere is a zero overall.
ERR = Pvalid × Punlock|valid × Popen × ( B × R × X ) ÷ CtestTerms are stratified into layers because they are not the same kind of object.
Pass all, or the candidate receives no score at all.
Anything that can single-handedly disqualify a candidate belongs here as a binary, not in the score as a large penalty. A penalty big enough to move a candidate more than one ranking band is a gate wearing a costume.
| Rank | Term | Question |
|---|---|---|
| 1 | P_valid | Does the original result have a scientifically sound core? Spans the full 0–1 range in practice and multiplies everything downstream. |
| 2 | P_unlock | valid | Does modern capability materially change the achievable result? This is the only term that distinguishes ORION from an ordinary grant panel. |
| 3 | P_open | Still unclaimed — and likely to remain so through our horizon? A hazard rate, not a checkbox. |
P_unlock comes out near 1.0 for most candidates, the
mapper is broken and ORION has no edge — everything looks unlockable to a fluent
narrator. A control class exists specifically to measure this. The framework is designed to be
able to report that it does not work.
P_open carries the entire economic justification for the foundry. If the
top-ranked opportunities are things a well-funded lab is already running, ORION's
counterfactual value is zero no matter how high they score. The scoring must be able to
drive such a candidate to zero.
| Rank | Term | Meaning |
|---|---|---|
| 4 | B | Benefit magnitude × scale, held as a vector over outcome dimensions. Severity times population — a product, as every real benefit metric is. |
| 5 | R | Residual fraction — the share of potential value still unrealized. |
| 6 | X | Cross-disciplinary option value: one result unlocking several fields. Bounded (cap ≈ ×1.25) so it can promote but never rescue. |
The benefit vector is kept uncollapsed inside the system. Different funders weight the same analysis differently; collapsing happens only at the report boundary, per funder, and is recorded as a separate artifact so any ranking is reproducible against a stated preference.
C_test is not the cost of the project. It is the cost of the
cheapest decisive test. In a staged foundry you are not buying the project — you
are buying the right to decide about the project. Each candidate is a call option and
C_test is the premium. Reproducibility and automation potential are not separate
scores; they are the inputs that determine C_test, and they earn their influence
through the denominator. Time-to-decisive-test and time-to-impact are tracked separately.
Deployment feasibility, tier / time horizon, discipline, geography and funder mandate are portfolio constraints and budget envelopes applied after ranking — not score components. Carrying them in the score costs reviewer time and cannot change any ordering.
latest. Elasticsearch is Elastic License v2 / SSPL and cannot be redistributed in a product container that ships to a marketplace.match_phrase + highlight offsets), not generated.doc_cid and fragment_id, regardless of which entry path produced it.Ten work items, each with a kill gate — a stated condition that ends the item if unmet.
| Item | Work | Kill gate |
|---|---|---|
| ORION-1 | Plan of record — architecture, rules, identity scheme, sequence | — |
| ORION-2 | Stand up search node; index orion-documents | offsets from highlight resolve against stored text |
| ORION-3 | Scaffold the agent — A2A + foundry endpoints | both entry paths yield the same doc_cid |
| ORION-4 | Revise six schema files to phrase_highlight | all six load; schema_id stable across restarts |
| ORION-5 | Ingest path — file, pmem source, index, embed | drop index, rebuild from pmem, identical fragment_id set |
| ORION-6 | Five extraction nodes | phrase recall ≥ 0.8; classification F1 ≥ 0.70 on 50 docs |
| ORION-7 | Record assembly + multiplicative ERR under a named weight profile | two reviewers reach Kendall τ ≥ 0.6 |
| ORION-8 | Dispatch + calibration loop | ≥1 of 5 Tier-1 runs publishable; Class A beats Class B better than chance |
| ORION-9 | Metrics, dashboard, service registration | dashboards hold real data |
| ORION-10 | Pilot corpus — bibliographic graph joined to full text on DOI | 50-doc labeled subset exists |
Blueprint stage. The evaluation framework, the architecture and the ten-item plan exist and are recorded. The corpus selection and acquisition scripts exist. No service has been built and no search node has been stood up. Confidence in the framework is deliberately held at moderate: the calibration set that would validate its weights is the first phase of the work, not a later refinement.