C2 · Publication Volume 12

Target Ranking

evidence weight, dependence, confidence and transparency

Learning goals

After this lesson, you should be able to construct an auditable ranking, protect necessary criteria with gates, distinguish evidence weight from confidence, identify dependent layers, model missingness, perform sensitivity analysis and communicate ties or rank instability rather than forcing false precision.

Ranking is a decision aid for allocating the next test. It is not a probability of discovery unless calibrated as one, and it is not economic value. A target can rank highly because it is scientifically discriminating, because it has strong geological support, because it is testable or because a particular objective values option preservation. These objectives should not be mixed invisibly.

Ranking frame and eligibility gates

Define the portfolio, decision date, target version, eligible actions and objective. Compare targets only at a compatible maturity and scale or state the adjustment. A regional lead with sparse data should not be punished as though it were a detailed target, nor should its uncertainty be hidden by generous scores.

Apply eligibility gates first. Necessary geological criteria, valid coordinate lineage, minimum data quality and a lawful, safe next test may be non-compensatory. Use pass, fail and indeterminate. An indeterminate target can remain in a gap-resolution queue instead of receiving a zero that implies negative evidence.

Keep geological support, evidence maturity and testability as separate dimensions. If a final ordering is required, show how the dimensions are combined and why. Also publish the component table, because a single rank conceals different pathways to the same total.

Weight, confidence and direction of evidence

An evidence weight expresses discrimination between hypotheses or importance to the declared objective. Confidence expresses trust in the observation and its transformation. A criterion can be highly diagnostic but poorly observed, or precisely measured but weakly diagnostic. Multiplying weight by confidence is one possible heuristic, not a universal probability rule.

Evidence may support, contradict or be neutral. Scores that only accumulate positive matches reward data-rich targets and cannot represent refutation. Use signed contributions or separate support and contradiction ledgers. Protect necessary contradictions through gates rather than allowing many weak positive features to offset them.

Weights should be elicited before target outcomes are known and documented with rationale. Use ranges where expert agreement is weak. Data-driven weights require coherent training labels, representative search space and validation that respects spatial structure. A mathematically fitted coefficient does not eliminate selection bias.

Dependence, redundancy and missingness

An auditable ranking separates gates, evidence groups, confidence and sensitivity before producing tiers
An auditable ranking separates gates, evidence groups, confidence and sensitivity before producing tiers

Evidence layers may share source data, causal processes or processing steps. Structural complexity, distance to faults and magnetic gradient can encode overlapping information. Several pathfinder elements from one sample medium may reflect one dispersion mechanism. Treating them as independent counts the same evidence repeatedly.

Create dependence groups and combine within groups before combining across groups. Test the effect of removing each group. If one data family determines the entire order, the result should be presented as conditional on that family. Correlation alone does not prove causal redundancy, but it signals the need for geological and lineage review.

Missingness can be uninformative, geology-related or exploration-related. Do not substitute average or zero without a model. Report coverage and missing-state codes alongside scores. Calculate a score interval if an unresolved criterion could plausibly take a range. Targets whose intervals overlap are effectively tied under current knowledge.

Sensitivity and validation

Sensitivity asks whether reasonable changes in assumptions alter the decision. Vary weights, thresholds, missing-data treatments, dependence grouping and confidence. Use leave-one-group-out tests and alternative hypotheses. A stable top tier is more defensible than a precise ordering that reverses under small changes.

Validation must match intended use. Randomly splitting spatially clustered occurrences can place near-duplicates in training and testing and inflate apparent performance. Use spatial blocks, geological domains, time-separated decisions or a genuinely prospective holdout where feasible. Preserve the chronology: a dataset acquired after target selection cannot validate the original selection.

Metrics must reflect action. Capture versus area, calibration, precision-recall, rank of held-out examples and information gained each answer different questions. Known occurrences are not an unbiased sample of all systems, and “non-deposit” locations are rarely verified true negatives. Report these label limitations.

Worked synthetic example

Three fictional targets pass necessary gates. Four criteria are scored from 0 to 1: source S, pathway P, trap T and independent footprint F. Declared weights are 0.30, 0.25, 0.25 and 0.20. Confidence values multiply the criterion scores as a transparent teaching heuristic.

| Target | S(c) | P(c) | T(c) | F(c) | |---|---|---|---|---| | A | 0.9 (0.9) | 0.8 (0.7) | 0.6 (0.8) | 0.7 (0.9) | | B | 0.7 (0.8) | 0.9 (0.9) | 0.8 (0.7) | 0.5 (0.8) | | C | 0.8 (0.6) | 0.6 (0.8) | 0.9 (0.9) | 0.8 (0.6) |

Scores are A=0.30(0.81)+0.25(0.56)+0.25(0.48)+0.20(0.63)=0.629; B=0.30(0.56)+0.25(0.81)+0.25(0.56)+0.20(0.40)=0.591; and C=0.30(0.48)+0.25(0.48)+0.25(0.81)+0.20(0.48)=0.563. The apparent order is A, B, C.

Suppose P and F derive from the same physical response for A and B, while C's footprint is independent. A dependence audit caps their combined contribution at the larger of the two rather than the sum. Revised teaching scores are A =0.243+0.140+0.120=0.503, B =0.168+0.203+0.140=0.511, and C remains 0.563 because its evidence is independent under the stated record. The order reverses. The lesson is not that a max rule is universally correct, but that dependence treatment can control the decision and must be explicit.

Interpretation workflow

  1. Define portfolio, objective, decision date and comparable maturity.
  2. Apply necessary geological, quality and practicability gates.
  3. Separate geological support, evidence maturity and testability.
  4. Assign diagnostic weight, direction and confidence distinctly.
  5. Group shared causal or source-data evidence.
  6. Encode missing, negative and indeterminate states explicitly.
  7. Calculate component scores, intervals and tiers rather than only ranks.
  8. Vary weights, gates, grouping and missing-data assumptions.
  9. Validate with spatial, geological or temporal separation appropriate to use.
  10. Record dissent, decision rationale and the next observation for each tier.

Practice and review

  1. Recalculate the original scores if trap weight increases to 0.40 and all weights are renormalised.
  2. Identify three potentially dependent evidence pairs in a mineral-system ranking.
  3. Explain why an untested criterion should be indeterminate rather than negative.
  4. Design a spatial validation split for clustered occurrences and state its remaining bias.
  5. Write a result in which two targets are tied because their sensitivity intervals overlap.

Sources and further reading