Pathophysiology biological_scale: Survey — Final Report

Sample: 91 pathophysiology nodes across 8 dismech disorders, each classified three times under different prompt specifications. Classifier: Sonnet 4.6 via subagents. Raw data: pass1-sonnet-all.csv, pass2-sonnet-all.csv, pass3-sonnet-all.csv.

Naming note. The slot was called kind: throughout the analysis (which is what appears in the classifier prompts, CSV column names, and design doc). The analysis concluded that the four values encode biological scale only — not "kind of thing" in a broader sense — so the slot was renamed biological_scale: on completion. Historical references to "kind" in artifact filenames (primary_kind CSV column, classifier-prompt.md) are intentionally preserved to keep the analysis reproducible. Prose in this report uses biological_scale: where it refers to the current recommendation and "kind" where it describes what the classifier did at time of run.

What this analysis answered

The proposed biological_scale: slot on Pathophysiology (an additive incremental improvement to the dismech pathograph model, worked on under the working name kind:) needed two design questions answered before implementation:

1. Does a small enum cover real pathophysiology nodes? And if so, how many values?
2. Does forcing a single-kind assignment usefully surface bundled nodes that should be split?

The three-pass survey answered both: yes, four scale-only kinds suffice (MOLECULAR / CELLULAR / TISSUE / ORGANISM), and yes, ~41% of nodes in the sample are bundle candidates that the slot would surface.

Three passes, three prompts

Three-pass headline

Two cleanly separated findings:

• Renaming the enum to be scale-only (v1 → v2) eliminated the state-vs-process taxonomy problem. New-kind proposals went from 4 to 0 and stayed at 0 across passes 2 and 3. State-shaped nodes (Ectopic Endometrial Tissue, Persistent Blastemal Progenitor State, Hyperphenylalaninemia, Elevated Circulating LDL) all landed cleanly when the suffixes that pre-judged process-vs-state were removed.

• Sharpening the prompt to treat split detection as a co-equal task (v2 → v3) recovered the bundle-surfacing signal. Split rate returned to 41% — identical to pass 1's number — confirming that v2's 4% was a prompt-attention artifact (agents anchored on "this is clean now"). The split-detection value of the biological_scale: slot is real and substantial.

Per-disorder breakdown (Pass 3, final)

Histogram is balanced — every kind has population, none dominates. Split rate varies widely by disorder (24-80%) — curators differ substantially in how atomically they decompose pathology.

The state-vs-process finding (resolved)

Pass 1 flagged 4 explicit new-kind proposals, all variations on *_STATE:

Plus agent-narrated state nodes (METABOLIC_STATE for Hyperphenylalaninemia, MOLECULAR_STATE for elevated LDL, ECOLOGICAL_STATE for gut dysbiosis) all became clean ORGANISM calls under the scale-only enum.

The insight: the v1 names suggested processes were the noun. Renaming made the substrate's biological scale the noun, and "process vs. state" became a verb-level detail — handled by the node's description and the existing modifier: slot.

The bundle/split finding (~41% of nodes)

Pass 3 confirmed pass 1's 41% rate (37/91 nodes flagged as split candidates) on the same 91-node sample, with strong but not identical overlap (24 nodes flagged by both passes, 13 unique to each).

Bundle patterns the agents named, by frequency in pass 3:

Per-disorder split rates from pass 3 vary from 24% (PKU, FH) to 80% (Endometriosis). Curators differ substantially in atomicity — the biological_scale: slot's split-surfacing value will be higher for some disorders than others.

Representative examples

Clean exemplars (high-confidence, atomic) per kind

Splits where pieces land cleanly (validation that splits don't just shift the problem)

11 atomic pieces from 4 batch-1 bundles were independently classified (split-pieces.md). 7 of 11 (64%) landed at confidence ≥ 4 in a single kind. The 2 pieces still proposing new kinds were both *_STATE variants that the v2 enum subsequently absorbed:

Deep-split (bundles whose pieces themselves bundle)

Edge cases — genuine substrate-boundary ambiguity

These are nodes any 4-bucket enum would find hard, and the survey is honest about them:

• Competitive LNAA Transport at the Blood-Brain Barrier (PKU) — molecular kinetics at an anatomical structure
• Vagal-Brainstem Alpha-Synuclein Propagation (Parkinson) — prion-like cell-to-cell mechanism routed through anatomical conduits
• Canonical Wnt Signaling Hyperactivation (Wilms) — pathway-level signaling, MOLECULAR vs CELLULAR genuine tie
• PCSK9 Gain-of-Function vs PCSK9-Mediated LDLR Degradation (FH) — adjacent nodes covering one gene at different scales (variant vs. consequence)

Conclusion and recommendation

Land the 4-value scale-only optional enum on Pathophysiology:

BiologicalScaleEnum:
  permissible_values:
    MOLECULAR: "Pathology at molecular scale — activities or states of molecules,
                complexes, or genetic elements"
    CELLULAR:  "Pathology at cellular scale — processes or states of a cell type
                or single cell"
    TISSUE:    "Pathology at tissue/organ scale — processes or states of a tissue,
                organ, or anatomical structure"
    ORGANISM:  "Pathology at organism scale — systemic / multi-organ / whole-body
                processes or states"

Why this works

1. Coverage — 91/91 nodes in the sample fit one of these 4 kinds with zero new-kind proposals across two passes.
2. Confidence — average classifier confidence 4.0+, with most residual ambiguity at genuine substrate-boundary cases that no enum would resolve.
3. No structural commitment — slot is optional; legacy 958 disorders work unchanged.
4. Heuristic backfill — for legacy nodes, a deterministic script can infer biological_scale: from existing descriptor slot usage (presence of molecular_functions → MOLECULAR; cell_types + biological_processes without anatomy → CELLULAR; etc.) for the confident cases.
5. Side benefit: node hygiene — 41% of sampled nodes are bundle candidates. The slot's forced single-value assignment, combined with curator awareness of split patterns (etiology+consequence, multi-organ enumeration, conjunctions), creates pressure toward atomic decomposition that no other slot currently provides.

What we did not recommend

• Adding STATE kinds. Pass 1 implied this was needed; passes 2/3 showed it was a v1 naming artifact.
• Splitting Pathophysiology into multiple classes. A single optional tag is far less disruptive than class-level decomposition.
• Multi-value (multivalued biological_scale:). Single-value assignment is the friction that surfaces bundles. Multi-value would lose this.

Caveats

• Sample is 8 disorders / 91 nodes, biased toward recently curated and toward mature/well-decomposed pathographs. A larger or differently-sampled corpus (older curations; thin pathographs; non-Mendelian conditions) may surface different patterns.
• Classifier is LLM (Sonnet 4.6). Calibration check with Opus 4.7 on the batch-1 16 nodes showed 94% primary-kind agreement after rename mapping; the methodology is not model-fragile but the absolute numbers carry the usual single-classifier caveats.
• The 41% split rate is what the classifier surfaced, not what a human reviewer would necessarily agree with. The batch-1 split-piece validation (where 7 of 11 hypothesized pieces landed cleanly) is the strongest evidence the splits aren't classifier hallucination, but a human-curator review pass would tighten this.

Files

• classifier-prompt.md — v1 spec (process-named enum, calibrated for confidence/split independence)
• classifier-prompt-v2.md — v2 spec (scale-only enum)
• classifier-prompt-v3.md — v3 spec (scale-only + sharpened split detection)
• pass1-sonnet-all.csv — v1 results (91 nodes)
• pass2-sonnet-all.csv — v2 results (91 nodes)
• pass3-sonnet-all.csv — v3 results (91 nodes, the authoritative pass)
• Per-disorder per-pass CSVs in same directory
• split-pieces.md — split-piece validation from batch 1 (11 atomic pieces from 4 bundles)
• pass1-results.csv — initial Opus inline pass (kept for record, superseded)