Add TypeSafe / Jev System One pass for U-NSGA-III Pareto scoring

[jkbennitt]

Summary

Adds an optional TypeSafe / Jev System One helper beside classical Pareto fronts so a small candidate sample can get calibrated Score and Choice judgments (constraint satisfaction, diversity, exploit-vs-explore, keep/drop/review). This is an additive semantic layer — it does not replace NSGA-III / U-NSGA-III objectives.

Python under tools/typesafe-pareto/ (same pattern as tools/oracle). One POST https://api.typesafe.ai/v1/systemone fans out all questions. Official typesafe-sdk is optional; stdlib HTTP matches the documented API. Live calls only when TYPESAFE_API_KEY is set. No keys in the repo.

Checklist

• dotnet test Unsga3.slnx -c Release not required (library/algorithm unchanged)
• CHANGELOG.md updated under Unreleased
• Public API unchanged
• Survival / normalization / metrics (IGD) untouched — no oracle impact

Test plan

• python -m unittest tools/typesafe-pareto/test_score_pareto.py -v (16 mock tests)
• python tools/typesafe-pareto/score_pareto.py --smoke --force-mock
• New CI job typesafe-pareto (Python 3.12, no secrets)
• Metrics append {ts, experiment:"unsga3_pareto_score", repo:"AppSprout-dev/Unsga3", model, latency_ms, usage, candidate_count, answers, notes} to metrics/typesafe-runs.jsonl (gitignored)

How to run

python tools/typesafe-pareto/score_pareto.py --smoke --force-mock
# live: export TYPESAFE_API_KEY=...  (never commit)

Docs used: HTTP API, Python SDK, fan-out.

  

[jkbennitt]

Live metrics (post-merge smoke)

Ran against tip after CI green / merge path. TYPESAFE_API_KEY via env only — not in repo.

unsga3_pareto_score (this PR)

• Model: jev-1.13.0
• Mode: live HTTP POST /v1/systemone
• Fixture: 8 synthetic ZDT1-like candidates (tools/typesafe-pareto/fixtures/zdt1_candidates.json)
• Questions: 32 in one fan-out (3× Score + 1× Choice per candidate)
• Latency: 529 ms wall
• Usage: 4661 input / 771 output tokens

id	f1	f2	feas	constraint	diversity	exploit	disposition	composite
c0	0.00	1.00	Y	2.91	1.57	2.00	keep	0.911
c5	1.00	0.00	Y	2.83	1.64	1.99	keep	0.911
c4	0.75	0.13	Y	2.97	1.36	1.66	keep	0.850
c1	0.10	0.68	Y	2.97	1.17	1.63	keep	0.813
c2	0.25	0.50	Y	2.98	1.10	1.00	keep	0.739
c3	0.50	0.29	Y	2.93	1.06	0.65	keep	0.690
c6	0.50	0.60	Y	2.97	0.90	0.39	drop	0.642
c7	0.40	0.37	N	0.03	0.91	0.74	drop	0.238

Read: extremes kept; crowded mid-front dropped; infeasible c7 correctly dropped on constraint score.

Companion baseline (Petra stage-gate smoke, same day)

Separate @typesafe-ai/sdk / JS smoke on a SMARTS-only Challenge 0 row (not this repo):

• 543 ms · 710 in / 114 out
• evidence_class = computational_proxy (conf 1.0)
• may_crown_stage_e noul = 0.06
• slate_action = demote (0.88)

Logged for X/metrics artifact; NSGA objectives still authoritative here — Jev is additive only.