Search, Model Selection, And Support

Search, selection, and support answer three different questions. Search asks which candidate states were explored. Selection applies a declared rule to those candidates. Support measures stability or relative evidence under a named procedure. A selected topology or model is not self-validating merely because the same run also produced a support number.

flowchart LR
    eligible["Eligible model or<br/>topology population"]
    explored["Evaluated candidates<br/>and failures"]
    rule["Selection rule"]
    selected["Selected result<br/>and ties"]
    support["Independent support<br/>procedure"]
    conclusion["Bounded conclusion"]

    eligible --> explored --> rule --> selected --> conclusion
    selected --> support --> conclusion

Preserve Four Populations

Population	Meaning	Common loss
eligible	every model, tree, start, or configuration admitted by the declared design	alternatives disappear before execution
attempted	candidates for which evaluation began	failures are omitted from the denominator
evaluated	candidates with an admissible objective or score	partial or non-finite results are treated as complete
selected	winner, ties, retained near-optima, or accepted posterior/tree population	only one convenient result survives

Report all four. A candidate that fails to fit is not evidence that the remaining winner is scientifically superior. It is a failed member of the selection experiment until its cause and eligibility are resolved.

Match The Record To The Search Family

Search family	Minimum exploration record	Unsupported shortcut
likelihood parameter optimization	starts, bounds, objective trace, convergence, boundary state	quoting the best scalar without optimizer state
likelihood topology search	starting trees, move set, accepted/evaluated candidates, repeated starts, termination	presenting the best encountered topology as a proven global optimum
Bayesian sampling	initial state, proposals, accepted/rejected moves, retained draws, chain identity	treating requested iterations as effective samples
parsimony search	starting construction, NNI/SPR/ratchet policy, score trace, equal-best set	discarding equal-best trees after the first discovery
distance tree construction	distance model, admitted pairs, correction, construction order and residuals	treating the derived tree as independent of the matrix

Stochastic and heuristic searches need seeds and budgets, but a seed is not an assurance argument. Repeat runs and diagnostics show whether the observed solution depends materially on starts, proposals, or random choices.

Select Only Comparable Candidates

Model ranking requires a common observation population and comparable objectives. Before applying AIC, AICc, BIC, a likelihood-ratio test, posterior model probability, parsimony score, or another rule, align:

• admitted taxa, sites, characters, and missing-data policy;
• tree, branch, root, state, partition, and transformation conventions;
• likelihood or cost definition and all constant terms relevant to the rule;
• parameter-count, constraint, and boundary treatment;
• convergence or completion policy;
• the complete eligible candidate set.

If those identities differ, preserve separate fits and explain the boundary. A ranking over non-comparable objectives is more misleading than no ranking.

Treat Ties And Near-Ties As Results

Exact ties can represent equal-best parsimony trees, identical criteria, or equivalent parameterizations. Near-ties arise under a declared tolerance or decision threshold. Retain every tied candidate, the tie rule, stable candidate identity, and the selection consequence.

Outcome	Correct record
one candidate clearly selected	winner plus every eligible candidate and rule
exact ties	complete tie set and any consensus or ambiguity projection
criterion differences below a declared materiality threshold	near-tie set and bounded interpretation
different starts reach different optima	start-linked results and unstable selection
some candidates fail	failed rows, causes and whether ranking remains admissible

Arbitrary file order, first-completion order, or display sorting is not a scientific tie-breaker.

Keep Support Method-Specific

Support output	Producing experiment	Does not mean
bootstrap clade proportion	resampled-data inference and clade recovery	posterior probability of the clade
SH-like or topology-test quantity	declared site-likelihood comparison	model adequacy or universal topology confidence
posterior clade frequency	retained posterior tree population	independent resampling support
jackknife proportion	character-deletion replicates	bootstrap support under another name
Bremer support	score increase required to lose a clade	probability that the clade is correct
consensus frequency	occurrence in an admitted tree set	evidence independent of how that set was generated

Every value retains its clade/split key, method, scale, requested and completed denominator, failures, and selection policy. Never move a value into a generic support column after dropping its producing method.

Review Selection Before Interpretation

flowchart TD
    identity{"Same observations and<br/>objective convention?"}
    complete{"Eligible candidates and<br/>failures retained?"}
    admissible{"Completion and boundary<br/>policy satisfied?"}
    support{"Support denominator and<br/>method identified?"}
    accept["Interpret bounded selection"]
    stop["Refuse ranking or<br/>qualify the conclusion"]

    identity -->|yes| complete
    identity -->|no| stop
    complete -->|yes| admissible
    complete -->|no| stop
    admissible -->|yes| support
    admissible -->|no| stop
    support -->|yes| accept
    support -->|no| stop

The minimum review packet contains eligible and attempted populations, candidate identities, objectives or scores, failures, search state, selection rule, ties, support rows, warnings, and structured artifacts. A rendered best tree can be included, but cannot replace that packet.