RRecktek · Initiative

ORION — Knowledge Foundry

orion.rrecktek.com · Blueprint — not yet built

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.

metasciencelatent value portfolio scoringknowledge extraction decision theoryOpenSearch pmem 2.0

The thesis

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.

Two halves

UNBUILT is the generator — it proposes candidates: claims from the historical record whose blocking constraint may have lifted.
ORION is the evaluator — it gates, scores, ranks, and dispatches them.

The two compose. This site describes the evaluator.

How candidates are scored

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 ) ÷ Ctest

Terms are stratified into layers because they are not the same kind of object.

Layer 0 — Gates (binary, unweighted)

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.

Layer 1 — Probability terms (multiplicative, 0–1) — these dominate

RankTermQuestion
1P_validDoes the original result have a scientifically sound core? Spans the full 0–1 range in practice and multiplies everything downstream.
2P_unlock | validDoes modern capability materially change the achievable result? This is the only term that distinguishes ORION from an ordinary grant panel.
3P_openStill unclaimed — and likely to remain so through our horizon? A hazard rate, not a checkbox.
Built-in falsifier. If 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.

Layer 2 — Value terms

RankTermMeaning
4BBenefit magnitude × scale, held as a vector over outcome dimensions. Severity times population — a product, as every real benefit metric is.
5RResidual fraction — the share of potential value still unrealized.
6XCross-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.

Layer 3 — Cost (the denominator)

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.

Layer 4 — Not criteria

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.

The honest limitation

There is no ground truth yet. Residual Scientific Value and Counterfactual Advancement are unobservable by construction — they are claims about a history that did not happen. Until a labeled outcome set exists, any weight vector is editorial preference formatted as measurement. Building that calibration set is Phase 0, not Phase 4. Every weight currently in the system is a prior, not a measurement.

Architecture rules

Plan of record

Ten work items, each with a kill gate — a stated condition that ends the item if unmet.

ItemWorkKill gate
ORION-1Plan of record — architecture, rules, identity scheme, sequence
ORION-2Stand up search node; index orion-documentsoffsets from highlight resolve against stored text
ORION-3Scaffold the agent — A2A + foundry endpointsboth entry paths yield the same doc_cid
ORION-4Revise six schema files to phrase_highlightall six load; schema_id stable across restarts
ORION-5Ingest path — file, pmem source, index, embeddrop index, rebuild from pmem, identical fragment_id set
ORION-6Five extraction nodesphrase recall ≥ 0.8; classification F1 ≥ 0.70 on 50 docs
ORION-7Record assembly + multiplicative ERR under a named weight profiletwo reviewers reach Kendall τ ≥ 0.6
ORION-8Dispatch + calibration loop≥1 of 5 Tier-1 runs publishable; Class A beats Class B better than chance
ORION-9Metrics, dashboard, service registrationdashboards hold real data
ORION-10Pilot corpus — bibliographic graph joined to full text on DOI50-doc labeled subset exists

Status

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.