Loading ToolBench
The triage engine for non-coding DMG variants, as a tool. NCypher couples three evidence axes for a somatic non-coding regulatory variant in paediatric diffuse midline glioma (DMG): chromatin accessibility (ChromBPNet), evolutionary constraint (Zoonomia phyloP), and measured function (lentiMPRA). Surfaces informative disagreements and is honest when a call sits out of domain.
NCypher defines 3 tools with clear, domain-specific descriptions (avg 180 chars, within baseline 194). All tools have input schemas with type definitions. However, parameter descriptions are sparse or missing entirely. score_variant accepts flexible variant formats but the description doesn't document accepted formats (chr14-33788719-A-G vs chr14:33788719:A:G). top_candidates has 3 params with descriptions, but ranked_by lacks enum constraints, LLM must guess valid values. Output schemas are declared but verbose (include unnecessary fields like provenance in every response). Error handling is absent: no guidance on invalid variant IDs, malformed formats, or missing data. Tool names follow verb_noun pattern (score_, top_, cohort_) but lack action clarity, 'score_variant' could mean validate, rank, or assess. No tool annotations (readOnlyHint, idempotentHint) despite all being read-only operations.
Honest headline numbers from the cached discovery sweep: total variants scored, high-impact chromatin count, constrained phyloP count, converged two-axis count, recurrent among converged, and in canonical driver genes. Includes caveats about domain and recurrence.
Three-axis convergence call + go/no-go memo for one variant. Scores a single non-coding regulatory variant by coupling chromatin accessibility (ChromBPNet fetal-OPC model), evolutionary constraint (Zoonomia phyloP), and measured function (lentiMPRA where available). Returns a convergence verdict, confidence tier, and actionable memo.
Ranked list of top N converged variants from the discovery sweep, optionally filtered to converged-only. Returns the highest-confidence candidates for experimental validation, ranked by agreement score or chromatin effect size.
Parameter constraints missing: ranked_by accepts 'agreement', 'chromatin', 'constraint' but no enum declared; LLM must guess valid values.
Variant ID format documentation incomplete: description mentions 'flexible formats' (chr14-33788719-A-G, chr14:33788719:A:G, etc.) but does not specify which are preferred or how parser handles ambiguity.
No error handling guidance: tools do not document what happens on invalid variant ID, missing cache entry, or malformed input. LLM has no recovery path.
Tool annotations absent: all 3 tools are read-only but lack readOnlyHint annotation. Agents cannot infer safety without explicit metadata.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
Output schemas include unnecessary provenance field in every response (SCORE_VARIANT_OUTPUT_SCHEMA, TOP_CANDIDATES_OUTPUT_SCHEMA). Provenance is static metadata, return once at server init or via resource, not per-call.