Statistical Genomics Foundry
Molds Mold
eval ✓ · scenarios ✓ family/brole/critiquerole/calibratedomain/phylogeneticsdomain/ancestral-reconstruction
contract

Reference Loading

Typed references describe what casting consumes, and when the generated skill should load each artifact.

Research yang-1995-ancestral

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeOn DemandVerbatimCorpus Observed
Purpose
Ground: the per-site/per-node posterior IS the accuracy measure; a unique best reconstruction can still be low-posterior (0.563, 0.185); long-branch/deep nodes least reliable; NO 0.95 reliability cutoff is prescribed.
Trigger
the analysis reports reconstructed ancestral states and their reliability/posterior is in question, or treats a single best reconstruction as certain
Research hochberg-thornton-2017-resurrection

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeOn DemandVerbatimCorpus Observed
Purpose
Ground: ASR output is an ensemble, MAP is one point estimate; characterize alternate plausible reconstructions to test robustness; qualitative function robust, quantitative values often not.
Trigger
a single ML/MAP ancestor is carried into a functional, structural, or experimental claim
Research szantho-2023-compositional-lba

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeOn DemandVerbatimCorpus Observed
Purpose
Invalidity pattern: site-homogeneous models underestimate convergent substitutions at low-Keff sites → long-branch attraction; simulation failure at p>=0.8; diagnostic = Keff-binned dlogL; remedy = CAT-PMSF.
Trigger
deep-node / long-branch reconstruction run under a site-homogeneous substitution model
Research foley-2022-grasp

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeOn DemandVerbatimCorpus Observed
Purpose
Ground: indels are a first-class inferred process (POG/BE), separate from character states; indel histories can be ambiguous; edge support = proportion of extant seqs; NO numeric indel- or substitution-posterior cutoff is given.
Trigger
the reconstruction involves indels/gaps, or claims indel/hybrid ancestral variants
Research paml-manual

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeOn DemandVerbatimCorpus Observed
Purpose
Ground: RateAncestor=1 -> rst (empirical-Bayes ASR); verbose=1 gives full marginal posterior; marginal (eq.4, per-node) vs joint (eq.2, all-nodes); named misuse = reusing reconstructed ancestral sequences as observed data; NO saturation cutoff stated.
Trigger
the analysis used codeml/baseml (PAML) to reconstruct ancestral states
Research rabosky-goldberg-2015-bisse

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeOn DemandVerbatimCorpus Observed
Purpose
Cautionary: BiSSE Type-I error grossly inflated on real trees (>77%, 61.5%; pure-birth control exactly 5%); required check = simulate neutral traits on THIS tree and recalibrate; a negative BAMM/MEDUSA result does not clear the dataset (25-55% error persists).
Trigger
a reconstructed trait/character state is claimed to affect diversification (BiSSE/SSE), or SSE significance is asserted without a tree-specific null
Research beaulieu-omeara-2016-hisse

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeOn DemandVerbatimHypothesis
Purpose
Offer (unverified) the character-independent / hidden-state model as a candidate null for trait-dependent diversification claims.
Trigger
an SSE analysis needs a character-independent null and HiSSE/hidden-state is proposed as that null
Verify
Only the abstract was readable (body paywalled). The explicit 'use the character-independent model as the null before claiming trait-dependence' recommendation, its parameterization, and any Type-I rates are behind the paywall; the abstract makes NO ancestral-state-accuracy claim. Verify against the OUP body before relying on this as a prescribed null.
Research cunningham-1999-asr-limitations

Background synthesis from a source note, loaded by explicit progressive-disclosure metadata.

RuntimeOn DemandVerbatimHypothesis
Purpose
Ground the thesis-level caution: hypothesis-test characters are usually under selection, so homoplasy corrupts ancestral-state inference more than phylogeny inference; irreversibility hypotheses are the hardest case.
Trigger
reconstructed ancestral states are used to test an evolutionary hypothesis, especially irreversibility (Dollo's law)
Verify
Only p.665 was readable; the detailed detector, the parsimony-vs-ML rate-asymmetry bias analysis, and the recommendations are on pp.666-674 (paywalled). Use only the p.665 thesis; re-summarize from the full text before relying on specifics.

Audit ancestral-reconstruction validity

A Family-B referee. Given an analysis that reconstructs ancestral sequences or ancestral character states on a fixed tree, judge whether its reliability claims are earned. Non-self- certifying: reasoning here (Critique) flags invalidity; where reasoning is not enough, hand off to a Calibrate check. Emit a verdict per axis: pass / flag / fail, each with the note it traces to. Never bless a reconstruction that trips a note-sourced failure mode.

0. Frame the claim

Establish, before auditing:

  • Target type: ancestral sequence (per-site states) vs ancestral character/trait state.
  • Reconstruction type: marginal (one interior node at a time, best = highest-posterior character; Yang eq.4) vs joint (all interior nodes simultaneously; Yang eq.2). In PAML this is model-dependent: one-rate models compute both; gamma models compute marginal only; nonhomogeneous models compute joint only. paml-manual
  • What downstream claim rests on it (a functional/experimental conclusion, a diversification test, a hypothesis test). The stringency below scales with this.

1. Uncertainty must be reported, not assumed

A single best (even unique most-parsimonious) reconstruction is not evidence of certainty: Yang shows unique best reconstructions with posterior only 0.563 (site 2) and 0.185 (site 117). The posterior probability is the accuracy measure. yang-1995-ancestral

  • flag any reconstruction reported as a definite sequence/state without per-site or per-node posterior probabilities. In PAML terms, verbose=1 yields the full marginal posterior distribution; best-character-only is verbose=0. paml-manual
  • Treat ASR output as an ensemble of plausible sequences with the MAP as one point estimate, not the answer. hochberg-thornton-2017-resurrection
  • Threshold discipline: do NOT accept “posterior >= 0.95 = reliable” as a sourced rule. Yang prescribes NO posterior cutoff; the 0.05-0.95 band is a table-display filter, not a reliability threshold. Any such cutoff is downstream convention — label it, don’t cite Yang. yang-1995-ancestral Likewise no cutoff defines an “ambiguous” site. hochberg-thornton-2017-resurrection

2. Deep / long-branch nodes are systematically less reliable

Nodes attached to extant taxa by long branches, or deep in the tree, carry the highest error (Yang: node 10 up to 8.7% per-sequence error; cytochrome-b per-node accuracy floor 0.870).

  • flag confident claims resting on deep/long-branch interior nodes. Joint accuracy of all nodes cannot exceed the marginal accuracy of the least reliable single node. yang-1995-ancestral

3. Model adequacy at depth — compositional LBA (planted-flaw axis)

If deep-node reconstruction runs under a site-homogeneous model (single equilibrium-frequency vector across sites, e.g. LG+F+G4, WAG/JTT-style), it underestimates convergent substitutions at compositionally constrained (low-Keff) sites and is prone to long-branch attraction — simulation fails to recover the true topology at p>=0.8. szantho-2023-compositional-lba

  • fail a deep-node reconstruction that asserts topology/ancestral states from a site-homogeneous model with no compositional-adequacy check.
  • Diagnostic to demand: the compositional constraint analysis — bin sites by Keff (= 1/sum(pi_i^2)) and sum per-site dlogL (log L_B - log L_A) across bins; conflicting signal across Keff bins signals model-driven LBA. Remedy: CAT-PMSF (IQ-TREE2 LG+F+G4 guide -> PhyloBayes CAT posterior mean site frequencies -> IQ-TREE2 PMSF). szantho-2023-compositional-lba
  • Posteriors themselves are unreliable if branch lengths (theta) or the model are wrong — check branch-length estimates and model fit. yang-1995-ancestral

4. Indel / gap handling

If the claim involves indels or gapped regions:

  • Gaps excluded or coded as a single missing state discard the indel process. GRASP models indels as a first-class inferred history (POG / bi-directional-edge encoding), separate from character-state inference. foley-2022-grasp
  • flag any single linearized ancestor presented as unique when indel histories are ambiguous (equally-parsimonious histories, or an edge preferred forward but not reverse); report edge support (proportion of extant sequences carrying the edge). foley-2022-grasp
  • Threshold discipline: GRASP states NO numeric indel-posterior or substitution-posterior cutoff (“high Shannon entropy”, “relatively high posterior” are qualitative). A ”>=0.8 indel-posterior” rule is tool/workflow convention, NOT Foley et al. — label, don’t cite. foley-2022-grasp

5. Downstream reuse of reconstructed sequences (circularity)

Reusing reconstructed ancestral sequences as if they were observed data for a further analysis is a named misuse with systematic-bias risk; prefer full-likelihood methods that integrate over ancestral states. paml-manual

  • fail analyses that plug point ancestral reconstructions back in as fixed data and then test a hypothesis on them without propagating reconstruction uncertainty.

6. Trait-dependent diversification claims (SSE / BiSSE)

If a reconstructed character state is claimed to drive speciation/extinction:

  • fail a BiSSE/SSE significance claim that lacks a tree-specific null. Neutral characters produce grossly inflated false positives on real trees (>77% at p<0.05, 58% at p<0.001 on the cetacean tree; 61.5% pooled across 186 vertebrate subtrees), while a pure-birth control gives exactly 5%. The required check: simulate neutral traits on THIS phylogeny and recalibrate the significance threshold from that null. rabosky-goldberg-2015-bisse
  • A negative BAMM/MEDUSA result does NOT clear the dataset (25-55% error persisted on no-detected-shift subtrees). rabosky-goldberg-2015-bisse
  • A character-independent / hidden-state model (HiSSE) is a candidate null — but this is UNVERIFIED here (abstract-only; see the reference’s verification). Do not present HiSSE as “the required null” on this evidence. beaulieu-omeara-2016-hisse

7. Hypothesis tests on selection-driven characters

When reconstructed states test an evolutionary hypothesis (especially irreversibility / Dollo’s law): the characters chosen are usually under selection, so homoplasy corrupts ancestral-state inference more than it corrupts phylogeny inference — a thesis-level caution. cunningham-1999-asr-limitations

  • flag irreversibility/directional-bias claims built on point ancestral reconstructions.
  • [GAP: the specific detector, the parsimony-vs-ML rate-asymmetry bias direction, and the paper’s recommendations are paywalled (pp.666-674) — cannot supply the concrete decision rule from notes.]

8. Calibrate handoff (the gate)

Critique above is model reasoning; it does not self-certify. Where a claim survives Critique but rests on contested reliability, hand off to a Calibrate deliverable:

Verdict

Emit per-axis pass|flag|fail with the citing note, plus one overall verdict. Any note-sourced failure mode that applies but was not addressed by the analysis -> at least flag; §3, §5, §6 map to fail when their check is entirely absent. A reconstruction reported as certain (no posteriors, or point states carried into a hard claim) must never be returned pass.