From Predictor to Instrument Evidence-Bounded Prediction and Prospective Assessment with Learned Evolution ABSLTN Public Research Note v0.1 — October 2026 Abstract Predictive systems used in changing physical or dynamical environments face a problem broader than forecast accuracy. Observations may be incomplete or stale. Predictions may extend across materially different horizons and operating conditions. Controls and external influences may alter future evolution. Different predictive engines may expose different capabilities and uncertainty semantics. Once forecasts are used operationally, later evidence creates a second question: what did the system actually claim before the outcome was known, and under what conditions was that claim made? This note describes a public architecture for treating those distinctions as first-class. The ABSLTN Predictive Instrument separates prediction from assessment. A predictive engine owns its prediction; the Instrument owns the conditions under which that prediction is evaluated. The Instrument is model-neutral and can assess different predictive engines without requiring them to expose capabilities they do not possess. Learned Evolution is ABSLTN's native predictive-engine research program for recursive learned dynamics under incomplete and changing evidence. The contribution described here is architectural. It is not a claim of universal predictive superiority, a new general filtering optimum, or validated production performance on live consequential systems. 1. Prediction Is More Than a Score A predictive model is often summarized by a single metric. Operational use is rarely so simple. A model may perform differently: - at short and long prediction horizons; - in different operating regimes; - when observations are complete versus sparse, delayed, noisy, or missing; - under different control or external-input assumptions; - before and after its own predictions begin feeding recursively into subsequent predictions. The Predictive Instrument keeps these conditions explicit rather than flattening them into one number. The result is an assessment envelope: a map of where a predictive system meets a defined performance standard, how that performance changes across relevant conditions, and where the limits become visible. 2. Public Predictive Principles Observation is not state. A measurement may expose only part of the system. The existence of a measurement does not imply complete knowledge of the underlying state. Inference is not observation. A reconstructed or estimated state may be the best current representation available while remaining different from directly observed evidence. Prediction is conditional. A forecast depends on the evidence available when it was made, the inferred current state, the predictive engine, the requested horizon, and relevant future assumptions. Changing those conditions may produce a materially different claim. Insufficient evidence may require refusal. A predictive system should be able to report that available evidence does not support a requested state estimate or forecast. Refusal can be a successful predictive result. Uncertainty retains provenance. Different engines represent uncertainty differently. A system that produces a point forecast should not be made to look equivalent to a system that natively produces a calibrated probabilistic representation merely for interface consistency. The engine owns the prediction; the Instrument owns the assessment. The Instrument defines shared assessment conditions. The engine retains its own model form, inference method, prediction method, and supported capabilities. Prediction history is prospective. Later truth should evaluate a prior prediction rather than replace or retrospectively reconstruct what the system originally claimed. 3. The Predictive Instrument The ABSLTN Predictive Instrument is an engine-neutral assessment layer for predictive systems. Its public purpose is straightforward: Find where a predictive system can actually be relied upon. Instead of asking only, "How accurate is this model?" the Instrument supports questions such as: - How does performance change as the prediction horizon increases? - Which operating conditions materially change model performance? - What happens when the evidence available to the model becomes sparse, delayed, noisy, or missing? - Which predictive capabilities are actually supported by the engine being assessed? - What did the system claim before the later outcome became known? The Instrument does not require Learned Evolution. Existing forecasting systems, simulators, digital-twin predictors, learned models, mechanistic models, vendor services, and hybrid systems can in principle be assessed if their predictions can be exposed through an appropriate boundary. 4. Assessment Envelope A single performance number can hide the boundary of predictive capability. The Instrument therefore emphasizes three intuitive dimensions: Time How does predictive performance change as the requested horizon increases? Conditions Where in the operating space does the model continue to meet the required standard? Evidence How does performance change as the observations available to the system change? Additional conditions may matter in particular domains, but the principle is the same: materially different predictive claims should remain distinguishable. The assessment envelope is not a declaration that a model is trustworthy in the abstract. It is evidence about whether a defined predictive claim meets a defined standard under defined conditions. 5. Learned Evolution Learned Evolution is ABSLTN's native predictive-engine research program. It explores recursive learned dynamics in settings where predictions may need to carry state forward between incomplete or asynchronous observations and where new evidence can later update the system's best current estimate. The important distinction is between one-step prediction and continued recursive use. A system that performs well when repeatedly given the correct observed state may behave differently when its own predicted state becomes part of the next prediction. Learned Evolution investigates this harder operating case while preserving the distinction among evidence, inferred state, prediction, and later reconciliation. Learned Evolution is one engine family. It is not the definition of the Predictive Instrument and is not required for model-neutral assessment. 6. Evidence-Bounded Behavior ABSLTN has tested the architecture through controlled synthetic pressures designed to make specific boundaries visible. Those studies include cases involving partial observation, missing or asynchronous evidence, recursive prediction, state reconciliation, and multiple predictive engines evaluated under common assessment conditions. The public conclusions are bounded: - insufficient evidence can be represented as an explicit refusal rather than a hidden assumption; - recursive prediction can coexist with later evidence reconciliation without rewriting what was previously predicted; - combining complementary observations can improve state estimation relative to weaker evidence configurations in the tested synthetic settings; - predictive engines with different capabilities can participate in the same assessment without fabricating unsupported capabilities; - later outcomes can be attached to previously released predictions to create prospective assessment evidence. These are architecture and synthetic-prototype results. They should not be interpreted as evidence of production reliability in any particular real-world domain. 7. Prospective Assessment A predictive system becomes more inspectable when each released forecast remains connected to the conditions under which it was made. The Predictive Instrument therefore treats prediction history prospectively: prediction -> later evidence -> assessment The original prediction remains part of the record. Later evidence can resolve or assess it without changing what was known at the prediction cutoff. This supports a broader shift from episodic model evaluation toward longitudinal evidence about how predictive systems behave under the conditions they actually encounter. 8. Relationship to Digital Twins The Predictive Instrument can exist inside a digital-twin environment without attempting to replace the broader digital-twin category. A digital representation may contain observed values, inferred states, predicted futures, assumptions, controls, and external inputs. The Instrument's narrower concern is to keep the epistemic status of those quantities visible and to assess predictive claims under explicit conditions. That distinction becomes especially important when the twin moves from representation toward operational decision support. 9. What This Public Work Establishes The current work supports the following bounded claims: - predictive roles such as observation, inference, and prediction can remain explicitly distinguishable in one architecture; - insufficient evidence can become an explicit system result; - recursive prediction and later evidence reconciliation can coexist without collapsing prior prediction and posterior estimate; - different predictive engines can be evaluated under common conditions while remaining capability-heterogeneous; - predictive claims can be preserved prospectively and assessed after later outcomes become available. 10. What This Public Work Does Not Establish The current work does not establish that Learned Evolution is universally superior to established forecasting, filtering, state-estimation, or data-assimilation methods. It does not establish state-of-the-art performance on standard benchmarks. It does not establish calibrated uncertainty across arbitrary domains or robustness to arbitrary distribution shift. It does not demonstrate production reliability, safety, or autonomous-control suitability on a live consequential system. It does not turn predictive capability into authority to act. These limitations define the next empirical program rather than claims to be hidden. 11. Research Direction The next public questions are empirical: - Can independently developed predictive engines participate without changing the Instrument's core assessment semantics? - How does the assessment envelope behave on real external domains not constructed around the architecture? - How should uncertainty be evaluated across horizon, regime, evidence quality, and engine version? - How should predictive history be used for long-lived systems whose sensors, models, regimes, and controls change over time? 12. Conclusion Operational prediction is a temporal process. Evidence arrives. State is inferred. Conditional futures are predicted. New evidence arrives. Predictions are eventually tested. A trustworthy predictive architecture should preserve those stages rather than allowing each new stage to rewrite the meaning of the previous one. The ABSLTN Predictive Instrument is built around that principle. Observe without pretending to possess the hidden state. Infer without calling inference an observation. Predict without hiding the conditions of the forecast. Refuse when the evidence does not support the requested claim. Express only uncertainty the predictive engine actually owns. Assess different engines under explicit shared conditions. Preserve what was predicted before reality arrives. Let later evidence evaluate the claim rather than rewrite it. That is the transition from predictor to Instrument. Public claim boundary This note intentionally describes the public problem, architectural principles, bounded synthetic evidence, and research direction while withholding detailed implementation architecture, internal interfaces, source code, parameterization, operational state mechanics, and unpublished experimental methods. Learned Evolution and the Predictive Instrument remain under active development.