Coarse To Fine Search¶
Essence¶
Coarse-to-Fine Search is a way to search a large possibility space without pretending that every possibility can receive full-detail evaluation immediately. It begins with a low-resolution pass over the whole bounded space, uses that pass to locate promising, uncertain, or high-risk regions, and then adds detail only where the added precision can change the outcome.
The archetype is not merely “narrow the list.” Its distinctive move is a change in search resolution: broad low-cost coverage first, selective high-resolution refinement second, and false-negative checks throughout.
Compression statement¶
When exhaustive fine-grained search is too costly, search coarsely first and progressively refine promising regions to concentrate effort without losing sight of the wider space.
Canonical formula: search_space + coarse_representation + promising_region_filter + resolution_transition_rule + refinement_step + false_negative_check -> focused high-resolution search
When to Use This Archetype¶
Use this archetype when the search space is too large for detailed review of everything, but a cheaper representation can still reveal where deeper investigation should go. It is especially useful when candidate options cluster into regions, categories, neighborhoods, stages, branches, hypotheses, sites, or parameter ranges.
It is weak when the early representation is arbitrary, when the cost of missing an option is unacceptable, or when rules require full detailed review before any exclusion. In high-stakes human, safety, medical, or legal contexts, the coarse stage should guide deeper attention rather than act as an unappealable final gate.
Structural Problem¶
The structural problem is the tension between coverage and depth. A broad search protects against tunnel vision but is too shallow to decide well. A fine-grained search gives richer evidence but cannot be applied to every possibility. Without a staged resolution strategy, teams either waste detailed effort on low-value options or use coarse screens as if they were final judgments.
The common failure is not just inefficiency. It is also invisibility: once a region is filtered out, nobody can tell whether it was genuinely low value, merely poorly represented by the coarse model, or never considered seriously.
Intervention Logic¶
The intervention starts by defining the search space boundary. The team then chooses a coarse representation that can cover the whole space cheaply while preserving enough signal to guide attention. A filter identifies promising, uncertain, risky, or strategically diverse regions. Those regions receive a refinement step: more evidence, more detailed modeling, deeper review, local testing, or higher-fidelity simulation.
The key guardrail is the false-negative check. The process should sample rejected regions, test known positives, preserve exception paths, or allow backtracking when fine-stage evidence shows that the coarse pass was wrong. Coarse-to-fine search works only when the early pass is treated as guidance, not infallible truth.
Key Components¶
Coarse-to-Fine Search organizes a two-resolution sweep over a possibility space: cheap broad coverage first, then targeted refinement only where added precision can change the outcome. The Search Space Boundary fixes what the coarse pass is responsible for and what is legitimately out of scope, so the screen cannot look efficient while silently ignoring whole regions. The Coarse Representation is the low-resolution view that scans the bounded space at affordable cost while preserving enough signal to guide attention. The Promising Region Filter then marks which areas warrant escalation, and the Resolution Transition Rule governs when and how the search actually shifts from coarse to fine, preventing arbitrary escalation. The Refinement Step adds detail only inside selected regions and only on the variables that could change the decision.
Three safeguards keep the search from collapsing into a one-shot screen. The False-Negative Check is the central guardrail — sample audits of rejected regions, holdout tests, challenge reviews, or backtracking that catch the cases the coarse pass discarded wrongly. The Coverage Record documents what was scanned, excluded, refined, and why, so exclusions remain inspectable and reversible rather than invisible. The Stop Rule closes the cycle by tying continued refinement to decision impact and remaining uncertainty rather than open-ended curiosity, so detail escalation ends when it stops changing the outcome.
| Component | Description |
|---|---|
| Search Space Boundary ↗ | Defines the option, region, case, diagnosis, design, parameter, or explanation space that the coarse pass must cover. Without a boundary, a coarse screen can look efficient while silently ignoring areas that were never searched. The boundary should state what is in scope, what is excluded, and why the exclusion is legitimate. |
| Coarse Representation ↗ | Represents the whole search space at a lower resolution so broad comparison is cheap enough to perform. This may be a rough grid, sketch, summary metric, symptom group, prototype class, portfolio category, map layer, or simplified feature set. It must preserve signals that are relevant to deciding where detail is worth adding. |
| Promising Region Filter ↗ | Identifies which coarse regions, options, or candidate groups should receive more detailed evaluation. The filter can be a score, threshold, constraint, risk flag, expert triage rule, feasibility test, or comparative ranking. It should be explicit enough that rejected regions can later be audited. |
| Resolution Transition Rule ↗ | Specifies when and how the search shifts from coarse evaluation to finer evaluation. The transition rule prevents arbitrary escalation. It may depend on potential value, uncertainty, risk, representativeness, diversity, or the cost of missing an option. |
| Refinement Step ↗ | Adds detail to selected regions so they can be tested, compared, or optimized at a higher resolution. Refinement should focus only on the variables, tests, views, or evidence that can change the decision. It is not generic improvement; it is targeted detail added after a broad pass. |
| False-Negative Check ↗ | Tests whether the coarse pass wrongly discarded important regions or candidates. This is the main safety guardrail. It can include sample audits of rejected regions, challenge reviews, holdout testing, backtracking, exception channels, or comparison against known positive cases. |
| Coverage Record ↗ | Documents what was scanned, what was excluded, what was refined, and which assumptions justified each transition. A coverage record makes the search inspectable and reversible. It supports later audits, re-opening of excluded regions, and learning about which coarse indicators were reliable. |
| Stop Rule ↗ | Determines when additional refinement is no longer worth the cost or no longer changes the decision. The stop rule keeps coarse-to-fine search from becoming endless detail escalation. It should relate to decision impact, uncertainty reduction, risk, and remaining evaluation budget. |
Common Mechanisms¶
Mechanisms implement the archetype, but they are not the archetype itself. The archetype is the transferable pattern of coarse coverage, selective refinement, and false-negative protection; the mechanisms below are ways to operationalize that pattern.
| Mechanism | Description |
|---|---|
| Multi-Resolution Search (`multi_resolution_search`) ↗ | Implements the archetype by scanning at multiple levels of resolution and escalating detail only where the lower-resolution pass indicates value, uncertainty, or risk. This mechanism should be used only when it preserves the broader coarse-to-fine logic rather than becoming a rote screen or a standalone artifact. |
| Coarse Grid Search (`coarse_grid_search`) ↗ | Implements the archetype in parameter or design spaces by testing a rough grid first, then searching more finely near promising grid regions. This mechanism should be used only when it preserves the broader coarse-to-fine logic rather than becoming a rote screen or a standalone artifact. |
| Funnel Process (`funnel_process`) ↗ | Implements the archetype in review or product workflows by moving many candidates through cheap early screens before detailed evaluation of a smaller set. This mechanism should be used only when it preserves the broader coarse-to-fine logic rather than becoming a rote screen or a standalone artifact. |
| Diagnostic Narrowing (`diagnostic_narrowing`) ↗ | Implements the archetype by starting with broad symptom, signal, or evidence groups and then applying more specific tests to likely diagnostic regions. This mechanism should be used only when it preserves the broader coarse-to-fine logic rather than becoming a rote screen or a standalone artifact. |
| Portfolio Screening (`portfolio_screening`) ↗ | Implements the archetype by using coarse financial, strategic, risk, or feasibility filters before intensive due diligence on selected opportunities. This mechanism should be used only when it preserves the broader coarse-to-fine logic rather than becoming a rote screen or a standalone artifact. |
| Progressive Candidate Review (`progressive_candidate_review`) ↗ | Implements the archetype by reviewing applications, proposals, designs, or options in stages, with deeper review reserved for candidates that pass earlier screens or uncertainty checks. This mechanism should be used only when it preserves the broader coarse-to-fine logic rather than becoming a rote screen or a standalone artifact. |
| Design Downselection (`design_downselection`) ↗ | Implements the archetype in design work by comparing rough concepts first and investing detailed engineering, prototyping, or testing in the most promising concepts. This mechanism should be used only when it preserves the broader coarse-to-fine logic rather than becoming a rote screen or a standalone artifact. |
| Search Tree Pruning with Refinement (`search_tree_pruning_with_refinement`) ↗ | Implements the archetype when a tree or hierarchy is explored shallowly first, then expanded more deeply along selected branches while keeping audit checks for pruned branches. This mechanism should be used only when it preserves the broader coarse-to-fine logic rather than becoming a rote screen or a standalone artifact. |
Parameter / Tuning Dimensions¶
Important tuning dimensions include the size of the coarse regions, the strictness of the promising-region filter, the amount of uncertainty margin around borderline cases, the cost budget for detailed review, the strength of false-negative auditing, the number of refinement levels, and the stop rule for ending further detail.
A conservative configuration uses broad inclusion thresholds, larger uncertainty margins, and stronger audits. This is appropriate when the cost of missing a candidate is high. An aggressive configuration uses tighter filters and fewer audits; it is appropriate only when coarse indicators are well validated and the stakes are low enough to tolerate misses.
Invariants to Preserve¶
The original search-space boundary must remain visible after attention narrows. The coarse representation must correspond to the fine representation closely enough that escalation decisions can be understood. The process must preserve an audit trail of exclusions, assumptions, and exceptions. It must retain a way to recover from false negatives. In high-stakes settings, final decisions should be supported by appropriately detailed evidence rather than by coarse scores alone.
Target Outcomes¶
The intended outcomes are lower search cost, faster identification of promising regions, better allocation of expert attention or computational resources, reduced premature commitment, and improved traceability. A mature implementation also improves learning: results from detailed review update the coarse filter for future searches.
Tradeoffs¶
The central tradeoff is efficiency versus miss risk. Coarse screens make search possible, but they can hide rare, subtle, novel, or underrepresented possibilities. Strong audits reduce this risk but consume resources. Tight filters improve throughput but can encode bias or overfit to what is easy to measure. Loose filters preserve options but may overwhelm the fine-review stage.
Failure Modes¶
The most common failure mode is overaggressive filtering: a candidate is discarded because it does not look promising in the coarse representation even though it would perform well under detailed review. Another failure mode is proxy drift, where the early screen measures convenience rather than value. A third is unreviewable exclusion, where there is no record of what was rejected or why. In organizational settings, a fourth failure is fairness harm: early screens can favor candidates whose value is visible in standardized coarse measures and penalize those whose value requires context.
Mitigation requires uncertainty margins, audit samples, backtracking paths, subgroup checks where people are affected, and stop rules that tie refinement to decision value rather than curiosity alone.
Neighbor Distinctions¶
Coarse-to-Fine Search is distinct from Search Space Pruning because pruning centers removal of regions that are impossible, dominated, or low value, while this archetype centers the resolution shift from coarse evaluation to fine evaluation. It is distinct from Progressive Narrowing because narrowing stages convergence toward fewer options, while coarse-to-fine search changes the level of detail applied to regions of the search space. It is distinct from Progressive Fidelity Increase because that archetype raises the fidelity of a model, prototype, or process; coarse-to-fine search applies detail selectively across alternatives.
It is also distinct from Coarse-Graining. Coarse-graining creates a higher-level representation; coarse-to-fine search may use such a representation, but its intervention is the search strategy that moves from that representation into selected details.
Variants and Near Names¶
Common variants include Multi-Resolution Search, where explicit resolution levels are used; Coarse Grid Search, where parameter spaces are first evaluated on a rough grid; Funnel Screening, where organizational candidates pass through increasingly detailed review stages; and Diagnostic Narrowing, where broad hypothesis categories lead to specific tests.
Near names include broad-then-deep search, coarse-to-detailed search, search by refinement, pyramid search, design downselection, portfolio screening, and progressive candidate review. These names should usually remain aliases, variants, or mechanisms unless future evidence shows distinct cross-domain components and failure modes.
Cross-Domain Examples¶
In engineering design, a team can use rough simulations to screen many design regions, then run expensive high-fidelity tests only on selected configurations. In geospatial planning, a regional map can identify places for detailed local surveys. In incident response, responders can check broad classes of failure before inspecting service-specific logs. In research, cheap screening assays can identify candidate mechanisms for deeper experiments. In portfolio review, many opportunities can be screened by coarse criteria before selected cases receive due diligence.
In each example, the coarse stage should guide attention, not replace the evidence needed for final judgment.
Non-Examples¶
A small three-option decision is not coarse-to-fine search if every option can be evaluated in detail. A legally required inspection is not this archetype if every item must receive the same full review before approval. A dashboard that reports aggregate metrics is not enough unless it triggers selective detailed analysis. Repeatedly improving a single prototype is not this archetype unless the process is also searching across alternative regions.