biolink: migrate the AI-curated corpus to Biolink 4.4.4 (qualifier model)

[andrewsu]

Why

docs/biolink_version_decision.md recorded the strategy as "re-curated from scratch in the latest
Biolink version… storage is the current vocabulary," and closed #19 and #54 as moot on that
basis. The premise did not hold. The curator was handed the legacy 67-predicate enum
(AGENTS.md §3, .claude/commands/curate.md) and QC Layer 3 enforced it, so forwardfilled records
came out in the legacy vocabulary.

Measured against Biolink 4.4.4 (latest release, 2026-08-10) with the repo's own drift monitor:

	removed predicates in use	affected edges
AI-curated (135 records, 4 runs)	7 distinct	323 / 558 = 58%
Legacy (kb/paths, 4,846 records)	23 distinct	16,873 / 32,641 = 52%

#69 built a drift monitor — read-only, dormant, human-PR-only. It detects this; it does not fix it.

What this does

Migrates the AI-curated corpus to Biolink 4.4.4 and makes the gate hold new curation there.
kb/paths/ is untouched — it keeps its vocabulary until forwardfill replaces it.

Pin + status. New src/drugmechdb/schema/biolink_predicate_status.yaml: the pinned release
(4.4.4), each enum value's status (current / legacy_only / deprecated), and the replacement
each removed predicate maps to. One source of truth, read by both the gate and the migration.

Schema (additive). PathEdge gains three optional slots — object_aspect_qualifier,
object_direction_qualifier, qualified_predicate — plus ObjectAspectEnum / ObjectDirectionEnum
mirroring Biolink's. Legacy records carry none of them and are unaffected.

Layer 3 is now era-aware. The enum deliberately spans both eras (legacy records still validate).
Which era a record may use is profile-dependent, mirroring qc.py's own split: legacy accepts any
enum member, ai_curated accepts only status: current — and affects/regulates must carry the
qualifiers that hold their polarity. Failures name the exact replacement.

Records migrated, not re-curated: scripts/migrate_predicates.py applied 324 rewrites across
132 files (line-based, so formatting/comments survive and the diff is only the changed edges).

-  - key: decreases activity of      +  - key: affects
                                     +    object_aspect_qualifier: activity
                                     +    object_direction_qualifier: decreased

Polarity follows the qualifier. structural_quality.py reads the sign from
object_direction_qualifier (synthesizing a <predicate>|<direction> lexicon entry per record), and
domain/range constraints key on affects|<aspect>. Net polarity on the migrated corpus:
coherent=124 / 135, unchanged in character — and type_violation coverage went 9 → 11, so the
checks got slightly sharper rather than silently lapsing.

Two defects found on the way

• sonnet/outputs/P13.yaml used binds, which was not in the enum — it fails Layer 3 today, on
  main. Cause: AGENTS.md advertised binds and affects in its example list while the enum
  contained neither. binds turns out to be deprecated in 4.4.4, so it is mapped to
  directly physically interacts with; the contract list is corrected.
• The status file's replacements initially pointed at predicates not in the enum
  (preventative for condition, contraindicated in, …), which would have migrated a legacy record
  straight into a Layer 3 failure. Those five are now enum members with polarity and type
  constraints, and test_every_replacement_points_at_a_current_predicate keeps the mapping closed.

Left for a human (never auto-rewritten)

Three predicates carry needs_human_decision: true — affects risk for (4.4.4 splits it by
direction; needs a per-edge reading) and increases/decreases response to (no response aspect
exists). All three are legacy-corpus-only.

Issue status

• [P2] Implement the predicate to qualifier remap rules #54 is not moot — the remap rules are implemented here and were used.
• [P1] Publish-time translation to current Biolink (qualifier remap) #19 is live again for the legacy corpus — publishing kb/paths/ as current Biolink still needs
  translation.
• Biolink predicate-vocabulary drift monitor (human-PR only) #69 stays open as the monitor; its report text no longer describes the superseded
  translate-at-publish policy.

Verification

• just qc on kb/paths/ (4,846): Layers 1 + 3 PASS. Layer 2's 230 failures are pre-existing —
  confirmed by stashing this branch and re-running on unmodified main.
• QC Layers 1–3 on all 135 AI-curated records: PASS (profile ai_curated).
• scripts/migrate_predicates.py --check: 0 edges remaining.
• Drift monitor vs 4.4.4: polarity lexicon 0 missing / 0 stale entries.
• Full test suite: 188 passed, 9 skipped, 0 failed (baseline on main: 182 passed, 5 failed).

🤖 Generated with Claude Code

---

Second commit (site:) is the browse layer — it builds the site JSON from the run outputs and
teaches the record page and graph renderer to read polarity from the qualifier, which they otherwise
lose after this migration. Happy to split it into its own PR if you'd rather review them apart.