Biology Buildout — Wave 1 Findings

Wave 1 additions (5 Mns):

• hydra — Cnidaria (Hydrozoa); eumetazoan nerve net; pre-bilaterian. Hypothesis: forms 2-member basal-animal cluster with amphimedon.
• coprinopsis — Basidiomycota; multicellular mushroom-forming fungus. Hypothesis: creates 3-member fungi cluster with yeast + neurospora.
• monosiga — Choanoflagellatea; animal sister group; unicellular with animal-toolkit precursors. Hypothesis: creates protist cluster with tetrahymena + dictyostelium.
• dictyostelium — Amoebozoa; colonial slime mold; cAMP-mediated multicellular transition. Hypothesis: part of protist cluster.
• physcomitrium — Bryophyta (moss). Hypothesis: forms 2-member non-vascular plant cluster with marchantia.

Corpus state: N=26 instance Mns (was 20 + 1 aggregate = 21).

Wave 1 results vs hypotheses

Hypotheses confirmed (1 of 5)

Hypotheses NOT confirmed (4 of 5)

Singletons after Wave 1

At threshold 8/9 (strict): 11 singletons (up from 6 at N=21).

• Pre-existing: amphimedon, halobacterium, marchantia, neurospora, tetrahymena, yeast (vertebrates is aggregate, not counted)
• New: coprinopsis, dictyostelium, hydra, monosiga, physcomitrium

At threshold 6/9 (moderate): 9 singletons (hydra resolved into animals; halobacterium resolved into prokaryotes).

At threshold 5/9 (relaxed): 5 singletons (coprinopsis–neurospora pair forms; halobacterium absorbed; some others marginal).

Methodology lessons (key Stage-2 material)

Lesson 1 — Adding "obvious structural neighbors" does NOT automatically resolve singletons

This is the most important finding from Wave 1. The entity-arrangement Wave 4 lesson ("adding Valkey resolves Redis-as-singleton; adding Cassandra resolves Mongo-as-singleton") was a strong claim about singletons-as-coverage-artifacts. Wave 1 here shows that lesson does NOT transfer cleanly:

• Pair-creation depends on JOINT profile similarity across all chain levels, not just one.
• Two organisms can be intuitively "the same kind" but score substantially differently across primitives (e.g., sponge vs cnidarian — both basal animals but bridge profiles differ).
• The clustering recipe integrates across chain levels and across signatures; what looks like a clean neighborhood at one level often fragments under cross-level integration.

Refined singleton triage (revised from gap-analysis doc):

The original 3-class triage (coverage / structural / authoring) underestimates the role of category coherence. The "outlier of category" class is new — sometimes the category itself isn't a single attractor.

Lesson 2 — Threshold sweeps reveal more than single-threshold reports

Reporting meta-stability at a single threshold (whether 8/9 strict or 5/9 relaxed) misses information. The buildout findings should report:

• Strict threshold spines — what's tightly stable (high anchor-candidate confidence)
• Relaxed threshold spines — what structural relationships exist below strict
• Threshold-where-pair-forms — diagnostic of relationship strength

This is a methodology refinement: anchor authoring should require strict-threshold spine persistence, but corpus-shape understanding requires relaxed-threshold analysis too.

Lesson 3 — Calibration coherence is a real authoring problem

I authored 5 new Mns in one session without re-checking calibration consistency against existing Mns. Result: at least 2 of 5 (physcomitrium, coprinopsis) have calibration suspicions — primitive choices that diverge from existing neighbors and prevent expected co-clustering.

Methodology rule (new): Before authoring a new Mn, sample 3-5 existing structurally-similar Mns and explicitly cross-check the primitive scoring against them. Make the calibration choice consistent unless there's a domain-specific reason to differ.

Probably also: a calibration-spot-check pass after each wave that runs cluster_classical, identifies pairings that DIDN'T form as expected, and recommends re-scoring candidates.

Lesson 4 — "Singletons are coverage artifacts" is a domain-dependent claim

The entity-arrangement Wave 4 lesson was strong because the entity corpus is software systems that share substantial structural similarity at substrate (E1-E3, T0-T2) and differ mostly at deployment/ecosystem layers. Biological organisms differ ACROSS chain levels — adding a single neighbor often doesn't bridge the gap if the new Mn's bridge or surface profile is distinctive.

The general rule should be: singleton resolution depends on the corpus's primitive-profile correlation structure, not on intuitive structural similarity.

In software (entity): high primitive-correlation across systems within a category → adding neighbors resolves singletons reliably. In biology: lower primitive-correlation (especially across substrate-bridge-surface) → adding neighbors creates pairs only when the JOINT profile is close enough.

This is the kind of cross-domain difference that the Stage-2 methodology doc should capture explicitly.

Decisions for Wave 2

Given these findings, the Wave 2 plan needs minor adjustment:

1. Re-score recommendations: Before Wave 2 begins, consider revising:
   • physcomitrium: PF3 → PF1-2 (match marchantia)
   • coprinopsis: Mem4 → Mem3 (match neurospora/yeast)
   • Possibly also: re-examine whether dictyostelium ST4 is correct (or whether it should be ST3 to align with monosiga)
2. Add a calibration-spot-check step to the wave protocol: re-cluster after re-scoring, see if expected pairs emerge.
3. Report meta-stability at multiple thresholds in each wave's findings.
4. Track which singletons are coverage / authoring / structural / outlier-of-category explicitly.

The Wave 2 candidate list (pinus, fern, xenopus, cartilaginous fish, marsupial, sea urchin, planaria) remains appropriate. Wave 2 will test whether phylogenetic gap-filling within already-well-represented clades (vertebrates, plants, bilaterian invertebrates) sharpens or fragments the existing spines.

Key per-level findings (data points for Stage 2)

biology-substrate (silhouettes 0.665–0.785)

Clean substrate-level clusters: {animals + plants together at G4/T4/R4/P4/Reg4/Mem4}, {ecoli + methanococcus + tetrahymena at lower substrate}, {monosiga alone at Mem2}, {fungi at intermediate substrate}. Substrate level CANNOT separate animals from plants (uniform high-substrate). Substrate level CAN separate basal-eukaryote (Mem2 — monosiga) from full-eukaryote (Mem3-4 — most others) from prokaryote (Mem2 — but ecoli at Mem2 too).

biology-to-organism-bridge (silhouettes 0.439–0.594)

Bridge level is the ACTIVE discriminator. Bilaterian invertebrates separate from cnidarians (PF, ECM, ND); hydra clusters closer to bilaterians than to amphimedon at this level. Vascular plants distinguish from bryophytes (VD, ED). Fungi cluster on CDiv-CDif-RD profile. Dictyostelium's ST4 is the dominant feature pulling it out of any cluster.

organism-architecture (silhouettes 0.309–0.498)

Lowest silhouettes — surface-level clustering is fuzziest. This is the dimension where individual-species idiosyncrasy dominates (specific Mo/Sn/Rs/Df values). Probably should be weighted less heavily in anchor-authoring decisions; substrate + bridge carry more structural signal.

Files