LodeCraft insight · Target generation
What a defensible target-generation study can and cannot establish.
A useful target-generation study narrows expert review, preserves its evidence and limitations, and refuses to turn a relative rank into a probability or discovery claim.
Publication details & scope
Direct answer
A defensible study can organise evidence and rank where expert investigation may be most useful. It cannot, by itself, establish mineralisation, discovery, economic value or a drilling decision.#
The strength of the study comes from traceability: the question, inputs, transformations, model roles, disagreement, target-selection rule and professional boundary stay attached to the result.
Begin with a geological hypothesis.#
Machine learning can search a large evidence space, but the project still needs a geological reason for combining particular datasets. The study should state the deposit or process concepts being examined, the expected evidence relationships and where those expectations may fail.
This does not require forcing every model to reproduce a preferred interpretation. It requires knowing what question a warm area is supposed to help answer.
Distinguish supervised prediction from unsupervised contrast.#
A supervised model fitted to labelled examples and an unsupervised model searching for unusual multivariate patterns answer different questions. If known occurrences influence model selection, threshold selection or version choice, they are no longer an untouched validation set.
Post-hoc overlap can be useful descriptive evidence, but it should not be reported as independent accuracy, enrichment or significance without an appropriate denominator and spatially valid comparison.
Keep support and disagreement visible.#
A final composite can hide whether its branches agree, whether one data family dominates or whether a model is extrapolating into weak support. Publish or deliver the relevant component views, coverage masks, branch disagreement and abstention areas alongside the final rank.
Where several branches use overlapping inputs, agreement is not independent replication. It may still be useful, but it must be interpreted honestly.
Define what the target geometry means.#
A percentile band or review polygon is a selection rule inside a stated study area. It is not automatically a probability contour. Ties, masks, smoothing, minimum-area rules and polygon simplification can change the realised footprint and should be documented.
The handover should explain whether an area is a relative review priority, a geological hypothesis, an investigation zone or another scoped output.
Challenge shortcuts before delivery.#
Roads, exposure, survey density, historic workings, drainage, cover and inherited mapping can align with real geology while also creating shortcuts. A defensible study tests plausible nuisance relationships and shows them as context or warnings rather than silently using them to improve the story.
A defensible study should retain#
Use this as your own preparation record. Selections last while this page is open; they do not validate project readiness.
What the study cannot establish alone#
It cannot convert incomplete data into certainty. It cannot substitute for a Competent Person where public reporting rules require one, and it is not automatically a public report of Exploration Results, Mineral Resources or Ore Reserves. The current JORC framework places responsibility for applicable public reporting on appropriately qualified and experienced Competent Persons.
Review MindSync’s Bynoe and Yamarna case studies for examples of ordinal ranks, explicit limitations and expert-review boundaries.
For a question about your own project, explore MindSync’s mineral exploration data analysis and target generation engagement, including the data-suitability gate and agreed handover.