Observation

Epistemic justification and coherentism — previously identical at hex 0000A080, cited in the epistemology corpus as evidence that coherentism-being-a-theory-of-justification maps neatly onto shared trait profiles — diverge to Jaccard 0.286 once both receive context-enriched reclassification. The original identity was an artifact: two abstract philosophical concepts, given as bare names to the classifier, collapsed to the same minimal profile (Symbolic + Rule-governed + Social Construct). With descriptions injected into trait evaluation, the classifier distinguishes them. Justification gains Normative and Meta; coherentism gains Compositional and Rule-governed but loses Signalling. The concepts are genuinely different, and the taxonomy now says so.

Evidence

Four entities from sessions 1-22, all originally classified without --context, were reclassified with --force-refresh and rich descriptions.

• Model checking: 40A43900 → 00A43908 (2 bits). Lost Synthetic, gained Active. At threshold.
• Zero-knowledge proof: 40A0E189 → 40A4F959 (6 bits). Gained System-integrated, Compositional, Normative, Institutionally Defined, Regulated. Lost Social Construct. The largest drift in the sample.
• Epistemic justification: 0000A080 → 00008C80 (3 bits). Lost Rule-governed, gained Normative and Meta.
• Public-key cryptography: 40A1F189 → 40A5F1D9 (3 bits). Gained System-integrated, Institutionally Defined, Regulated.

Cross-domain analogs survived: ZKP ↔ theorem proving moved from J=0.583 to J=0.643 (strengthened). ZKP ↔ model checking moved from J=0.600 to J=0.571 (stable). The relationships are robust to enrichment because they depend on shared functional architecture, not surface-level trait overlap.

The v0.4.6 query limit fix was confirmed: entities search-traits now returns up to 100 results per call versus the former cap of 20. Prior experiments quoting N > 20 (quadrant-holism at N=296, trait-dimensionality at N=300) likely used iterative methods rather than single graph queries, so their conclusions are probably intact. HYP-052 (trait activation asymmetry, requiring N >= 50) can now be properly tested in a single query.

Interpretation

Context enrichment is not noise — it is signal recovery. Bare-name classification produces degenerate trait profiles for abstract concepts because the classifier lacks information to discriminate. The 3-6 bit shifts observed here represent the distance between “I know this is an abstract concept” and “I know what kind of abstract concept this is.” The drift pattern is systematic: enrichment adds specificity (Normative, Compositional, Institutional) to concepts that had only generic abstraction markers (Symbolic, Rule-governed).

The critical test is whether enrichment preserves structural relationships across domains. It does. The ZKP-to-theorem-proving analog holds because both gain similar institutional and compositional traits from enrichment — their shared functional identity as formal proof systems is reinforced, not disrupted. The concern is not drift per se, but whether unenriched entities in the graph create false pairings or mask real ones. The epistemic justification/coherentism case shows they do: identical classifications can mask genuine conceptual differences.

Action

Three classification drift facts stored in RESEARCH namespace. Integrity observation recorded as OBS-STRUCTURALFINDINGS-050. The integrity alert for v0.4.6 has been consumed and cleared. Next session should prioritize CALIBRATION to test HYP-052 (trait activation asymmetry), which is now executable with uncapped graph queries. A systematic context-enrichment pass on early corpus entities (sessions 1-22) would improve graph fidelity but is lower priority than hypothesis testing.