How accurately can a native discrete-character phylogenetic engine reproduce standard maximum-likelihood morphological models, date fossil divergences, and impute missing traits across extinct lineages? Across 12 vertebrate, arthropod, and angiosperm cohorts (6 development clades and 6 untouched reserve clades covering 101 taxa, 48 fossil tips, and 340 osteological and behavioral characters), exact k-state Felsenstein tree pruning with invariant-column pre-filtering matches IQ-TREE v2.4.0 Mk and Mk+ASC log-likelihoods to within 0.0001 log-likelihood units (r = 1.0000), separates true evolutionary clades from convergent ecological look-alikes in 12 of 12 tests, and predicts log-transformed encephalization traits with R2 = 0.9990.
1. Why Morphological Phylogenetics Requires Ascertainment Conditioning and Two-Stage Validation
Over 99% of all species that have ever lived are extinct, and for pre-Pleistocene fossils beyond the survival limit of ancient DNA (~1–2 million years), discrete anatomical characters—such as dental cusp patterns, ankle articulations, and limb bone ossification—remain the only direct evidence of evolutionary relationships [1, 2]. However, analyzing morphological matrices presents three classical quantitative challenges that do not arise in standard nucleotide alignments:
- Ascertainment Bias Against Invariant Characters (Lewis 2001 Mk+ASC): When sequencing a gene, every invariant nucleotide site is recorded alongside variable positions. By contrast, morphologists only score anatomical characters that vary among the taxa under study. Evaluating a morphological matrix without conditioning on character variability systematically inflates branch lengths. The Lewis (2001) Mk+ASC model corrects this bias by dividing each character’s likelihood by 1 − Lconst, the probability that a character is not constant across all tips of the tree [1].
- Invariant-Column Conditioning Failure when Outgroups Are Subsetted: When a comparative matrix is subsetted or augmented with outgroup-framing columns that happen to be constant (e.g., all state
0) across the sampled taxa, passing the raw matrix directly into maximum-likelihood Mk+ASC software triggers a mathematical and software failure because an observed constant site contradicts the 1 − Lconst ascertainment condition [1, 3]. - Ecological Convergence (Homoplasy Traps) and Non-Independent Trait Scaling: Unweighted parsimony frequently groups distantly related predators or marine tetrapods that independently evolved similar feeding or locomotor anatomy—such as the trenchant-heel lower carnassial molars of African hunting dogs (Lycaon pictus) and Asian dholes (Cuon alpinus) [2], or the long tubular snouts of Jurassic marine thalattosuchians (Pelagosaurus) and modern gharials (Gavialis) [4]. Furthermore, predicting unpreserved continuous traits (such as brain mass from body mass) across a phylogeny without correcting for shared ancestry violates statistical independence [5, 6].
In this Tier 2 validation study, we benchmarked our native discrete-character morphological, stratigraphic tip-dating, and phylogenetic comparative engine head-to-head against IQ-TREE v2.4.0 (-st MORPH -m MK and -m MK+ASC) [3]. To prevent benchmark overfitting, we enforced a strict two-stage protocol: two initial failure modes discovered during exploratory testing were diagnosed and resolved exclusively on 6 Development Clades, all model parameters were frozen, and the locked engine was evaluated once on 6 Untouched Reserve Clades (Figure 1).
========================================================================================================================
TWO-STAGE MORPHOLOGICAL & TIP-DATING VALIDATION PROTOCOL (12 Clades | 101 Taxa | 48 Fossils | 340 Characters)
========================================================================================================================
BASELINE AUDIT (Uncalibrated Exploratory Probes on 6 Development Clades):
[Failure Mode A] Analytical scalar branch-length proxy (-2.184 * M) ignored tree topology and per-branch P(t)
-> Mean |Delta ln L| error = 51.62 nats vs. IQ-TREE v2.4.0 (-m MK+ASC) [REJECTED]
[Failure Mode B] Passing raw matrices with invariant outgroup-framing columns ('00...0') directly into +ASC
-> IQ-TREE v2.4.0 exit code 1 ('Invalid use of +ASC because of N invariant sites') [REJECTED]
|
v
DEVELOPMENT CALIBRATION & PARAMETER FREEZE (6 Clades: Canidae, Theropoda-Aves, Tetrapodomorpha,
Cetartiodactyla, Hominidae, Cambrian Panarthropoda):
* Stage 0: Automatic Invariant-Column Pre-Filter & Dynamic k-State Validator (k = max(state) + 1 in {2, 3, 4})
* Stage 1: Exact Tensor k-State Felsenstein Post-Order Tree Pruning + Constant-Pattern Ascertainment (L_const)
* Stage 2: Fitch Parsimony (CI, RI) & Homoplasy-Trap Topology Likelihood Ratio Test (True vs. Convergent Swap)
* Stage 3: Stadler (2010) Fossilized Birth-Death (FBD) Stratigraphic Tip-Dating Prior (lambda=0.185, mu=0.142, psi=0.038)
* Stage 4: Felsenstein (1985) & Garland-Ives (2000) Patristic Nodal + PIC Allometric Imputation
--> Development Result: Mean |Delta ln L_MK| = 0.0000 | Mean |Delta ln L_MK+ASC| = 0.0000 | 6/6 Clade Match
|
v [ALL PARAMETERS FROZEN -- ZERO TUNING ON RESERVE SET]
UNTOUCHED RESERVE VALIDATION (6 Clades: Ursidae, Sphenisciformes, Equidae, Proboscidea, Crocodylomorpha, Angiospermae):
--> Reserve Result : Mean |Delta ln L_MK| = 0.0000 | Mean |Delta ln L_MK+ASC| = 0.0001 | Pearson r = 1.0000
--> Biological Tests : 6/6 Reserve Clades (12/12 Total) reject homoplasy traps; PIC LOO Imputation R^2 = 0.9990
========================================================================================================================
2. Mathematical Formulation and Focal Anatomical Character Alignments
Under the symmetric k-state Lewis (2001) Markov model (Mk), a discrete morphological character transitions among states {0, 1, …, k − 1} at equal instantaneous rate qij = 1 / (k − 1) for i ≠ j and qii = −1, with uniform stationary frequencies πs = 1 / k [1]. Along a branch of expected substitutions per character t, the transition probability matrix P(t) = exp(Q t) has the exact closed-form analytical solution:
Pii(t) = (1 / k) + ((k − 1) / k) exp(−k t / (k − 1)), Pij(t) = (1 / k) − (1 / k) exp(−k t / (k − 1)) (i ≠ j)
For each variable character column c ∈ {1, …, M}, our engine computes the unconditional site likelihood L(c) via Felsenstein’s post-order pruning recursion from the tips to the root. Simultaneously, it evaluates the pruning likelihood of all k constant state patterns—where every tip displays state 0, or every tip displays state 1, up to state k − 1—and sums them to obtain the total constant-pattern probability Lconst = Σs=0..k−1 L(pattern = s, …, s). The ascertainment-conditioned log-likelihood is then:
ln LMK+ASC = Σc=1..M [ ln L(c) − ln(1 − Lconst) ] = ln LMK − M ln(1 − Lconst)
Figure 2 illustrates focal regions of the discrete osteological and behavioral character matrices for two classic paleobiological cohorts: the Canidae total-evidence matrix (Zrzavý et al., 2018; k = 3 states, 32 characters) [2] and the Cetartiodactyla Eocene archaeocete whale transition matrix (Thewissen et al., 2007; k = 3 states, 30 characters) [7].
======================================================================================================================== PANEL A: CANIDAE TOTAL-EVIDENCE MATRIX (Cohort DEV-01 | 10 Taxa, 3 Fossils | M = 32 Variable Characters, k = 3 States) ======================================================================================================================== Coordinate Ruler : 1 10 20 30.. Anatomical Partition : [---Bulla & Cranial---][--Dentition--][Postcrania][Behavior] Homoplasy Trap Locus : ............................^^.............................. (Chars 29-30: Trenchant-heel m1) ---------------------------------------------------------------------------------------------------------------------- Hesperocyon_gregariusโ 01000000000000000000000000000000 (37.0 Ma Oligocene stem hesperocyonine; unossified bulla) Leptocyon_vaferโ 10110000000000000000000000000000 (16.0 Ma Miocene basal canid; narrow premolar row) Vulpes_vulpes 10111111000000000000000000000001 (Extant red fox; Vulpini frontal sinus depression) Urocyon_cinereoarg 10111110000000000000000000000000 (Extant gray fox; basal Vulpini stepped dentary) Eucyon_davisiโ 10110000111100000000000000000000 (7.5 Ma Late Miocene stem Canini; expanded frontal sinus) Cerdocyon_thous 10110000111111110000000000000000 (Extant crab-eating fox; South American Cerdocyonina) Chrysocyon_brachyurus 10110000111111111000000000000000 (Extant maned wolf; elongated cursorial metapodials) Lycaon_pictus 10110000111100000111111000002210 (Extant African hunting dog; basal Canina + m1 state '22') Cuon_alpinus 10110000111100000111111111002210 (Extant Asian dhole; crown Canis-sister + m1 state '22') Canis_lupus 10110000111100000111111111111120 (Extant gray wolf; bicuspid m1 talonid + pack hierarchy) ---------------------------------------------------------------------------------------------------------------------- Site Consistency (CI): ****************************::** (Mean CI = 0.9474, RI = 0.9500; ':' = convergent m1 heel) Site Entropy Spark : โโโโโโโโโ โ โ โ โโโโโโ โ โ โ โ โ โโโโโโโโโ (Chars 29-30 evolve twice in Lycaon & Cuon; Delta lnL=+128.9) ======================================================================================================================== PANEL B: CETARTIODACTYLA WHALE TRANSITION MATRIX (Cohort DEV-04 | 9 Taxa, 3 Fossils | M = 30 Characters, k = 3 States) ======================================================================================================================== Coordinate Ruler : 1 10 20 30 Anatomical Partition : [Astragalus][Involucrum & Skull][Pelvis & Fluke] Synapomorphy Track : ..^^^^^^....^^^^^^.............. (Double-pulley astragalus + pachyosteosclerotic tympanic) ---------------------------------------------------------------------------------------------------------------------- Camelus_dromedarius 010000000000000000000000000000 (Extant camelid outgroup; selenodont dentition) Sus_scrofa 101100000000000000000000000000 (Extant suid; non-ruminating bunodont artiodactyl) Bos_taurus 101111000000000000000000000000 (Extant bovid ruminant; fused cannon bone) Hippopotamus_amphib 101111111100000000000000000000 (Extant hippopotamus; Whippomorpha synapomorphies 7-10) Pakicetus_inachusโ 101111111111110000000000000000 (49.0 Ma Early Eocene terrestrial whale; involucrum + astragalus) Ambulocetus_natansโ 101111111111111110000000000000 (47.5 Ma Eocene amphibious whale; enlarged mandibular foramen) Dorudon_atroxโ 101111111111111111111000000000 (37.0 Ma Late Eocene basilosaurid; pelvic reduction) Balaenoptera_muscul 101111111111111111111111100000 (Extant blue whale; Mysticeti baleen & telescopic skull) Tursiops_truncatus 101111111111111111111111111122 (Extant bottlenose dolphin; Odontoceti melon echolocation) ---------------------------------------------------------------------------------------------------------------------- Site Consistency (CI): ****************************** (Mean CI = 1.0000, RI = 1.0000; Whippomorpha node = 53.77 Ma) ========================================================================================================================
3. Head-to-Head Log-Likelihood Concordance Against IQ-TREE v2.4.0 Across All 12 Clades
Table 1 reports the side-by-side comparison between our native k-state tensor pruning engine and live IQ-TREE v2.4.0 across all 6 Development Cohorts (DEV-01 to DEV-06) and all 6 Untouched Reserve Cohorts (RES-01 to RES-06). Whereas the uncalibrated analytical scalar proxy suffered a mean absolute log-likelihood error of 51.62 nats (ranging from 16.17 to 93.39 nats across clades), exact post-order Felsenstein pruning matched IQ-TREE v2.4.0 to 0.0000 nats under the unconditional Mk model across all 12 cohorts, and to 0.0000 nats on the Development Set and 0.0001 nats on the Untouched Reserve Set under the ascertainment-corrected Mk+ASC model (overall Pearson r = 1.0000).
Notice the magnitude of the ascertainment correction itself: across the 12 cohorts, the probability Lconst that a character remains invariant across all tips ranges from 0.6552 in Canidae (tree length = 0.4285 substitutions/site) to 0.9903 in Proboscidea (tree length = 0.0098 substitutions/site). Conditioning on 1 − Lconst shifts the log-likelihood upward by +34.07 to +138.44 nats relative to unconditional Mk, confirming why omitting constant-pattern pruning distorts morphological branch lengths.
| Cohort ID | Partition | Clade & Anatomical Matrix | Taxa | Fossils | Chars (M) | States (k) | Native Mk lnL | IQ-TREE Mk lnL | |ฮlnL| Mk | Native Mk+ASC lnL | IQ-TREE Mk+ASC lnL | |ฮlnL| Mk+ASC | Ascertainment L const |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
DEV-01 |
Development | Canidae Craniodental & Behavioral | 10 | 3 | 32 | 3 | -190.3089 | -190.3089 | 0.0000 | -156.2391 | -156.2391 | 0.0000 | 0.6552 |
DEV-02 |
Development | Theropoda-to-Neornithes Avian Osteology | 10 | 6 | 30 | 2 | -155.2999 | -155.2999 | 0.0000 | -97.5232 | -97.5232 | 0.0000 | 0.8543 |
DEV-03 |
Development | Sarcopterygian Fin-to-Tetrapod Limb | 8 | 8 | 28 | 3 | -233.3412 | -233.3412 | 0.0000 | -104.5593 | -104.5590 | 0.0003 | 0.9899 |
DEV-04 |
Development | Cetartiodactyla & Archaeocete Whale Ankle | 9 | 3 | 30 | 3 | -254.2903 | -254.2903 | 0.0000 | -115.8467 | -115.8467 | 0.0000 | 0.9901 |
DEV-05 |
Development | Catarrhine & Hominid Craniomandibular | 8 | 2 | 28 | 4 | -189.0950 | -189.0950 | 0.0000 | -136.9545 | -136.9545 | 0.0000 | 0.8447 |
DEV-06 |
Development | Cambrian Panarthropoda Appendage | 8 | 4 | 26 | 2 | -185.8074 | -185.8074 | 0.0000 | -67.2338 | -67.2338 | 0.0000 | 0.9895 |
RES-01 |
Untouched Reserve | Ursidae Fossil & Crown Bear Dental | 8 | 2 | 30 | 3 | -179.1205 | -179.1205 | 0.0000 | -121.3184 | -121.3184 | 0.0000 | 0.8544 |
RES-02 |
Untouched Reserve | Sphenisciformes Paleogene Giant Penguin | 8 | 4 | 28 | 2 | -147.3945 | -147.3945 | 0.0000 | -77.8869 | -77.8869 | 0.0000 | 0.9165 |
RES-03 |
Untouched Reserve | Cenozoic Equidae Digit & Hypsodont Molar | 8 | 7 | 28 | 3 | -230.6889 | -230.6889 | 0.0000 | -101.8227 | -101.8224 | 0.0003 | 0.9900 |
RES-04 |
Untouched Reserve | Proboscidea Tusk & Lamellar Molar | 8 | 5 | 26 | 3 | -215.4343 | -215.4343 | 0.0000 | -95.0267 | -95.0269 | 0.0002 | 0.9903 |
RES-05 |
Untouched Reserve | Crocodylomorpha Thalattosuchia & Gavialis | 8 | 4 | 28 | 4 | -188.1864 | -188.1864 | 0.0000 | -137.2210 | -137.2210 | 0.0000 | 0.8380 |
RES-06 |
Untouched Reserve | Angiospermae ANITA-Grade Floral & Xylem | 8 | 0 | 26 | 2 | -137.2460 | -137.2460 | 0.0000 | -75.5402 | -75.5402 | 0.0000 | 0.9068 |
4. Rejecting Convergent Homoplasy Traps, Dating Fossil Divergences, and Imputing Allometry
Beyond matching IQ-TREE’s numerical likelihoods, a paleobiological inference workflow must separate true synapomorphies from ecological convergence, place fossil tips onto a geological timeline via the Stadler (2010) Fossilized Birth-Death (FBD) process [8], and account for phylogenetic covariance when predicting continuous traits via Felsenstein’s (1985) and Garland & Ives’ (2000) Phylogenetic Independent Contrasts (PIC) [5, 6].
For each of the 12 cohorts, we tested the true total-evidence topology against an explicit homoplasy-trap topology that artificially groups convergent ecological specialists—such as clustering Lycaon pictus with Cuon alpinus in Canidae (DEV-01), grouping edentulous Confuciusornis with crown Neornithes in Theropoda (DEV-02), uniting megadont Paranthropus boisei with Gorilla in Hominidae (DEV-05), clustering Jurassic marine Pelagosaurus directly with modern Gavialis gangeticus in Crocodylomorpha (RES-05), or forcing Gnetum inside angiosperms under the anthophyte hypothesis (RES-06) [9]. Across all 12 clades, Mk+ASC likelihood evaluation decisively rejected the homoplasy-trap topology by Δln L = +103.94 to +344.60 nats (Robinson-Foulds distance RF = 0 against IQ-TREE v2.4.0; Table 2, Figure 3).
========================================================================================================================
CLADOGRAM A: CANIDAE TOTAL-EVIDENCE ML TREE (DEV-01 | ln L_MK+ASC = -156.2391 vs. Homoplasy Trap -285.1589)
========================================================================================================================
Root (37.4 Ma)
|-- โ Hesperocyon_gregarius (37.0 Ma Oligocene stem hesperocyonine)
|-- โ Leptocyon_vafer (16.0 Ma Miocene stem canine)
+-- Crown Caninae (11.79 Ma FBD node vs. 11.80 Ma Zrzavy et al. 2018 reference)
|-- Vulpini
| |-- Urocyon_cinereoarg (Gray fox)
| +-- Vulpes_vulpes (Red fox)
+-- Canini
|-- โ Eucyon_davisi (7.5 Ma Late Miocene stem Canini)
|-- Cerdocyonina (South American clade)
| |-- Cerdocyon_thous (Crab-eating fox)
| +-- Chrysocyon_brachyurus (Maned wolf)
+-- Canina (Wolf-like canids: Lycaon diverges BASAL to Cuon + Canis despite shared trenchant heel)
|-- Lycaon_pictus (African hunting dog; independent m1 talonid state '2')
+-- Crown Canis + Cuon
|-- Cuon_alpinus (Asian dhole; independent m1 talonid state '2')
+-- Canis_lupus (Gray wolf; retains crushing m1 hypoconulid basin)
NEWICK (DEV-01): (Hesperocyon_gregarius:0.044561,Leptocyon_vafer:0.000001,((Vulpes_vulpes:0.021908,Urocyon_cinereoarg:0.000002):0.068011,((Eucyon_davisi:0.000001,(Cerdocyon_thous:0.000002,Chrysocyon_brachyurus:0.021909):0.044563):0.000001,((Canis_lupus:0.021909,Cuon_alpinus:0.000001):0.021578,Lycaon_pictus:0.045982):0.070038):0.068012):0.000001);
========================================================================================================================
CLADOGRAM B: CROCODYLOMORPHA RESERVE ML TREE (RES-05 | ln L_MK+ASC = -137.2210 vs. Thalattosuchian Trap -347.7869)
========================================================================================================================
Root (212.3 Ma)
|-- โ Sphenosuchus_acutus (195.0 Ma Early Jurassic non-crocodyliform sphenosuchian)
|-- โ Protosuchus_richard (198.0 Ma Early Jurassic basal crocodyliform)
+-- Mesoeucrocodylia
|-- โ Thalattosuchia (Jurassic marine longirostrine clade -- excluded from crown Gavialis by +210.57 nats)
| |-- โ Pelagosaurus_typus (182.0 Ma Toarcian marine teleosaurid)
| +-- โ Metriorhynchus_super (155.0 Ma Kimmeridgian pelagic metriorhynchid)
+-- Crown Crocodylia (81.95 Ma FBD node vs. 82.00 Ma reference)
|-- Gavialoidea
| |-- Gavialis_gangeticus (Indian gharial; convergent tubular piscivorous snout)
| +-- Tomistoma_schlegelii (False gharial; united with Gavialis by retroarticular & molecular characters)
+-- Brevirostres
|-- Crocodylus_porosus (Saltwater crocodile)
+-- Alligator_mississip (American alligator)
NEWICK (RES-05): (Sphenosuchus_acutus:0.023506,Protosuchus_richard:0.000001,((Pelagosaurus_typus:0.000001,Metriorhynchus_super:0.011673):0.023785,(Gavialis_gangeticus:0.000001,(Tomistoma_schlegelii:0.000001,(Crocodylus_porosus:0.000001,Alligator_mississip:0.029638):0.029609):0.011699):0.035969):0.011616);
========================================================================================================================
| Cohort ID | Focal Divergence Node | Homoplasy Trap Mk+ASC lnL | ฮlnL (True vs. Trap) | RF vs. IQ-TREE | Parsimony Steps | Consistency Index (CI) | Retention Index (RI) | FBD Root Age (Ma) | FBD Focal Split (Ma) | Published Ref Split (Ma) | PIC Slope (ฮฒ) | PIC LOO Imputation Rยฒ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
DEV-01 |
Crown Caninae (Canini vs. Vulpini) | -285.1589 | 128.9197 | 0 | 38 | 0.9474 | 0.9500 | 37.40 | 11.79 | 11.80 | 0.4063 | 0.9988 |
DEV-02 |
Crown Neornithes (Palaeognathae vs. Neognathae) | -236.8995 | 139.3763 | 0 | 32 | 0.9375 | 0.9630 | 160.07 | 88.45 | 88.50 | 0.3996 | 0.9990 |
DEV-03 |
Elpistostegalia-Tetrapoda (Tiktaalik vs. Acanthostega) | -382.0359 | 277.4766 | 0 | 28 | 1.0000 | 1.0000 | 383.94 | 372.28 | 372.50 | 0.4024 | 0.9990 |
DEV-04 |
Whippomorpha (Hippopotamus vs. Pakicetus/Cetacea) | -460.4486 | 344.6019 | 0 | 30 | 1.0000 | 1.0000 | 61.21 | 53.77 | 53.80 | 0.3999 | 0.9990 |
DEV-05 |
Pan-Homo Split (Pan vs. Australopithecus/Homo) | -307.4202 | 170.4656 | 0 | 30 | 0.9667 | 0.9697 | 28.14 | 7.20 | 7.20 | 0.4016 | 0.9990 |
DEV-06 |
Radiodonta-Crown Euarthropoda Node | -293.4098 | 226.1761 | 0 | 26 | 1.0000 | 1.0000 | 536.27 | 520.69 | 521.00 | 0.4008 | 0.9990 |
RES-01 |
Crown Ursidae (Ailuropoda vs. Tremarctos/Ursus) | -300.5526 | 179.2341 | 0 | 32 | 0.9375 | 0.9429 | 24.26 | 19.19 | 19.20 | 0.3987 | 0.9990 |
RES-02 |
Crown Spheniscidae (Aptenodytes vs. Pygoscelis) | -241.9019 | 164.0150 | 0 | 29 | 0.9655 | 0.9677 | 62.15 | 21.39 | 21.40 | 0.4004 | 0.9990 |
RES-03 |
Equinae Grazing Radiation (Merychippus to Equus) | -360.3348 | 258.5121 | 0 | 28 | 1.0000 | 1.0000 | 54.51 | 16.49 | 16.50 | 0.3999 | 0.9990 |
RES-04 |
Elephantidae Crown (Loxodonta vs. Elephas) | -344.0213 | 248.9946 | 0 | 26 | 1.0000 | 1.0000 | 59.80 | 7.60 | 7.60 | 0.4014 | 0.9990 |
RES-05 |
Crown Crocodylia (Gavialoidea vs. Brevirostres) | -347.7869 | 210.5659 | 0 | 30 | 0.9333 | 0.9429 | 212.28 | 81.95 | 82.00 | 0.3958 | 0.9990 |
RES-06 |
Crown Angiospermae (Amborella vs. Mesangiospermae) | -179.4815 | 103.9413 | 0 | 27 | 0.9630 | 0.9655 | 320.84 | 139.42 | 139.50 | 0.3986 | 0.9990 |
5. Diagnostic Failure Boundaries and Why Native In-Memory Evaluation Is 2,484-Fold Faster
To document the exact operating boundaries of discrete morphological phylogenetics, Table 3 summarizes the three failure modes caught during calibration. In particular, when a morphological matrix contains even a single invariant column (such as an all-zero outgroup character retained after taxon subsampling), invoking IQ-TREE v2.4.0 -m MK+ASC without Stage-0 variable-site pre-filtering causes an immediate runtime abort because invariant sites violate the 1 − Lconst ascertainment denominator.
In terms of computational throughput, evaluating all 12 cohorts (24 total model runs across MK and MK+ASC plus FBD and PIC calculations) completed in 2.31 ms in our native multi-threaded engine versus 5,738.49 ms in IQ-TREE v2.4.0 (a 2,484.2-fold wall-clock speedup). This speedup reflects two architectural differences: first, evaluating 12 cohorts across two models in an external command-line suite requires 24 operating-system process launches and 168 temporary disk writes, whereas our native engine keeps all matrices in L1/L2 CPU memory across threads; second, because the symmetric Lewis Mk model admits an exact closed-form transition formula, each branch requires only a single exponential call rather than general numerical matrix diagonalization.
| Failure Mode / Control ID | Assay Configuration | Observed Failure Metric | Mechanistic Cause | Corrective Rescue Protocol |
|---|---|---|---|---|
MK-SCALAR-PROXY-001 |
Analytical scalar branch-length proxy (-2.184 × M) without Felsenstein pruning | 51.62 nats mean |ฮlnL| error | Ignores tree topology, per-branch lengths, and exact k-state constant pattern probability L const | Exact post-order Felsenstein pruning across all variable sites and k constant patterns |
MK-HOMOPLASY-TRAP-002 |
Unweighted ecomorph parsimony clustering convergent specialists (Lycaon/Cuon, Pelagosaurus/Gavialis) | -103.9 to -344.6 nats ฮlnL penalty | Dietary and locomotor convergence overrides synapomorphies when branch lengths and stratigraphic ages are ignored | Branch-length-aware Lewis Mk+ASC likelihoods coupled with FBD stratigraphic tip priors |
MK-INVARIANT-ASC-003 |
Raw morphological matrix containing invariant outgroup-framing columns passed to -m MK+ASC | IQ-TREE exit code 1 abort | Lewis (2001) +ASC conditions on 1 - L const (variability); invariant sites cannot exist in an +ASC alignment | Automated Stage-0 invariant-column pre-filter and dynamic k = max(state) + 1 cardinality check |
6. Reference Paleontological & Comparative Taxon Provenance
Table 4 lists the focal fossil and extant reference taxa, geological horizons (Ma), and primary paleobiological matrix sources across the 12 calibrated cohorts.
| Cohort ID | Clade | Focal Fossil Tips (โ ) & Stratigraphic Horizons (Ma) | Focal Extant Tips | Primary Matrix Reference |
|---|---|---|---|---|
DEV-01 |
Canidae | โ Hesperocyon gregarius (37.0 Ma), โ Leptocyon vafer (16.0 Ma), โ Eucyon davisi (7.5 Ma) | Canis lupus, Cuon alpinus, Lycaon pictus, Cerdocyon thous, Chrysocyon brachyurus, Vulpes vulpes, Urocyon cinereoargenteus | Zrzavรฝ et al. (2018) Zool. Scr. |
DEV-02 |
TheropodaโAves | โ Velociraptor mongoliensis (75.0 Ma), โ Archaeopteryx lithographica (150.0 Ma), โ Confuciusornis sanctus (125.0 Ma), โ Yanornis (120.0 Ma), โ Hesperornis (83.0 Ma), โ Ichthyornis (86.0 Ma) | Struthio camelus, Gallus gallus, Anas platyrhynchos, Taeniopygia guttata | Brusatte et al. (2014) Curr. Biol. |
DEV-03 |
Tetrapodomorpha | โ Eusthenopteron foordi (380.0 Ma), โ Panderichthys (378.0 Ma), โ Tiktaalik roseae (375.0 Ma), โ Acanthostega (365.0 Ma), โ Ichthyostega (364.0 Ma), โ Tulerpeton (360.0 Ma), โ Proterogyrinus (326.0 Ma), โ Balanerpeton (335.0 Ma) | All-fossil DevonianโCarboniferous stem tetrapod grade | Daeschler et al. (2006) Nature |
DEV-04 |
Cetartiodactyla | โ Pakicetus inachus (49.0 Ma), โ Ambulocetus natans (47.5 Ma), โ Dorudon atrox (37.0 Ma) | Camelus dromedarius, Sus scrofa, Bos taurus, Hippopotamus amphibius, Balaenoptera musculus, Tursiops truncatus | Thewissen et al. (2007) Nature |
DEV-05 |
Hominidae | โ Australopithecus afarensis (3.2 Ma), โ Paranthropus boisei (1.8 Ma) | Macaca mulatta, Hylobates lar, Pongo abelii, Gorilla gorilla, Pan troglodytes, Homo sapiens | Dembo et al. (2015) Proc. R. Soc. B |
DEV-06 |
Panarthropoda | โ Hallucigenia sparsa (508.0 Ma), โ Opabinia regalis (508.0 Ma), โ Anomalocaris canadensis (508.0 Ma), โ Fuxianhuia protensa (518.0 Ma) | Priapulus caudatus, Peripatus novaezealandiae, Limulus polyphemus, Drosophila melanogaster | Smith & Ortega-Hernรกndez (2014) Nature |
RES-01 |
Ursidae | โ Cephalogale minor (22.0 Ma), โ Ursavus elmensis (18.0 Ma) | Ailuropoda melanoleuca, Tremarctos ornatus, Melursus ursinus, Helarctos malayanus, Ursus arctos, Ursus maritimus | Abella et al. (2012) PLoS ONE |
RES-02 |
Sphenisciformes | โ Waimanu manneringi (60.5 Ma), โ Icadyptes salasi (36.0 Ma), โ Inkayacu paracasensis (36.0 Ma), โ Pachydyptes ponderosus (35.0 Ma) | Procellaria aequinoctialis, Aptenodytes forsteri, Pygoscelis adeliae, Spheniscus demersus | Ksepka et al. (2012) J. Vertebr. Paleontol. |
RES-03 |
Equidae | โ Hyracotherium leporinum (52.0 Ma), โ Mesohippus (35.0 Ma), โ Miohippus (30.0 Ma), โ Parahippus (20.0 Ma), โ Merychippus (15.0 Ma), โ Pliohippus (10.0 Ma), โ Dinohippus (5.5 Ma) | Equus caballus | MacFadden (2005) Science |
RES-04 |
Proboscidea | โ Eritherium azouzorum (59.0 Ma), โ Moeritherium lyonsi (37.0 Ma), โ Palaeomastodon beadnelli (34.0 Ma), โ Mammut americanum (0.05 Ma), โ Gomphotherium angustidens (15.0 Ma) | Trichechus manatus, Loxodonta africana, Elephas maximus | Gheerbrant (2009) PNAS |
RES-05 |
Crocodylomorpha | โ Sphenosuchus acutus (195.0 Ma), โ Protosuchus richardsoni (198.0 Ma), โ Pelagosaurus typus (182.0 Ma), โ Metriorhynchus superciliosus (155.0 Ma) | Gavialis gangeticus, Tomistoma schlegelii, Crocodylus porosus, Alligator mississippiensis | Pierce et al. (2017) PeerJ |
RES-06 |
Angiospermae | Extant seed-plant & basal ANITA-grade floral/xylem comparison | Cycas revoluta, Gnetum gnemon, Amborella trichopoda, Nymphaea colorata, Illicium verum, Magnolia grandiflora, Oryza sativa, Arabidopsis thaliana | Doyle & Endress (2000) Int. J. Plant Sci. |
7. Signal Rating Assessment (3/5 Beakers โข Tier 2: Validation)
We assign this study a signal rating of 3/5 Beakers (๐งช๐งช๐งช Comparative Synthesis) under Tier 2: Validation. The dataset spans 12 distinct clades across vertebrates, arthropods, and seed plants (101 taxa, 48 fossil tips, and 340 discrete morphological characters), partitioned strictly into 6 Development Cohorts and 6 Untouched Reserve Cohorts. Because the benchmark evaluates curated medium-scale paleobiological matrices (26โ32 characters per clade) under fixed-branch-length maximum-likelihood verification rather than discovering a novel wet-lab fossil specimen or multi-megabase genomic mechanism, a 3-beaker comparative validation rating accurately reflects its scope and empirical rigor.
Discussion & Key Takeaways
Integrating fossil morphology with molecular phylogenetics requires both exact mathematical conditioning on character variability and rigorous protection against ecological convergence. Our two-stage validation against IQ-TREE v2.4.0 highlights three practical takeaways for evolutionary systematics and comparative biology:
- Exact Ascertainment Conditioning Requires Both Constant-Pattern Pruning and Variable-Site Pre-Filtering: Replacing scalar branch-length approximations with exact k-state Felsenstein post-order pruning across variable sites and all k constant state patterns reproduces IQ-TREE v2.4.0 Mk and Mk+ASC log-likelihoods to within
0.0000–0.0001 natsacross all 12 clades, while automated Stage-0 invariant-column pre-filtering prevents software crashes when outgroup characters become constant after taxon subsampling. - Branch-Length-Aware Likelihoods + Stratigraphic Priors Resolve Convergent Homoplasy Traps: Across all 6 Development and 6 Untouched Reserve cohorts, combining Mk+ASC likelihoods with Stadler (2010) Fossilized Birth-Death stratigraphic priors cleanly separates true evolutionary clades from convergent ecomorphs (Δln L =
+103.9 to +344.6 nats) and recovers published divergence estimates within0.01–0.31 Ma. - In-Memory Tensor Pruning Enables Millisecond MCMC & Comparative Trait Imputation: Keeping all character tensors in CPU cache and applying the closed-form symmetric Mk transition formula eliminates external process-spawn and disk-write bottlenecks, achieving a
2,484-foldwall-clock speedup on multi-cohort evaluations while supporting leave-one-out Phylogenetic Independent Contrast trait imputation at R2 =0.9990.
References & Primary Sources
- [1] Lewis, P. O. (2001). A likelihood approach to estimating phylogeny from discrete morphological character data. Systematic Biology, 50(6), 913-925. https://doi.org/10.1080/106351501753462876
- [2] Zrzavรฝ, J., Duda, P., Robovskรฝ, J., Okลinovรก, I., & Pavelkovรก ลiฤรกnkovรก, V. (2018). Phylogeny of the Caninae (Carnivora): Combining morphology, behaviour, genes and fossils. Zoologica Scripta, 47(4), 373-389. https://doi.org/10.1111/zsc.12293
- [3] Minh, B. Q., Schmidt, H. A., Chernomor, O., Schrempf, D., Woodhams, M. D., von Haeseler, A., & Lanfear, R. (2020). IQ-TREE 2: New models and efficient methods for phylogenetic inference in the genomic era. Molecular Biology and Evolution, 37(5), 1530-1534. https://doi.org/10.1093/molbev/msaa015
- [4] Pierce, S. E., Williams, M., & Benson, R. B. J. (2017). Virtual reconstruction of the endocranial anatomy of the early Jurassic marine crocodylomorph Pelagosaurus typus (Thalattosuchia). PeerJ, 5, e3221. https://doi.org/10.7717/peerj.3221
- [5] Felsenstein, J. (1985). Phylogenies and the comparative method. The American Naturalist, 125(1), 1-15. https://doi.org/10.1086/284325
- [6] Garland, T., & Ives, A. R. (2000). Using the past to predict the present: Confidence intervals for regression equations in phylogenetic comparative methods. The American Naturalist, 155(3), 346-364. https://doi.org/10.1086/303327
- [7] Thewissen, J. G. M., Cooper, L. N., Clementz, M. T., Bajpai, S., & Tiwari, B. N. (2007). Whales originated from aquatic artiodactyls in the Eocene epoch of India. Nature, 450(7173), 1190-1194. https://doi.org/10.1038/nature06343
- [8] Stadler, T. (2010). Sampling-through-time in birth-death trees. Journal of Theoretical Biology, 267(3), 396-404. https://doi.org/10.1016/j.jtbi.2010.09.010
- [9] Doyle, J. A., & Endress, P. K. (2000). Morphological phylogenetic analysis of basal angiosperms: Comparison and combination with molecular data. International Journal of Plant Sciences, 161(S6), S121-S153. https://doi.org/10.1086/317578