UX5 — Instance-vs-historical compare view (E3)

[jack-champagne]

Part of the seam UX-prototype track (Krishna interview). Priority: later — after the basics (UX1/UX2) and comprehensive (UX3/UX4) surfaces.

What

An instance-vs-historical comparison view: on the instance side, trial-to-trial changes in the pulse plots; on the historical side, the ILC objective over iterations — plus comparison between runs.

Krishna's signal (interview 2026-06-30, p2)

• "instance artifacts vs. historical data plots → comparison between runs?"
• (instance) trial changes in pulse plots; (historical) ILC objective over iterations

Seam-prototype approach

Side-by-side / two-run webview fed by fake run.toml / result.toml fixtures + a synthetic ILC-iteration series (grounded in @amicode/schema). No RunsManager. Bridges the inspector (near-term) and the catalog (later).

Feedback questions for Krishna / Andrew

1. Which comparison matters most — the same pulse across ILC iterations, or two different pulses head-to-head?
2. What does "ILC objective over iterations" need to show to be trustworthy (per-iter fidelity, convergence, the changing pulse)?

Lanes

• Kate — comparison layout · Raghav — shell + contract test + 2-run fixtures · Jack — which comparisons matter

Design freedom

Exploration prompt, not a fixed spec — reshape / combine / drop as feedback shows what works.

Refs: spec-20260529-amicode-architecture (RunInspector multi-run). Maps to build Phase 1.3 + catalog.

[jack-champagne]

Fold-in from the workflow-frames spec §6.2 — this is the design for the compare view. A sweep family = runs sharing one hash while the other varies: fix formulation_hash vary system_hash (robustness across calibration drift) or fix system_hash vary formulation_hash (the reformulation study). Multi-select on the Run/Pulse node → overlay Ω(t)/φ(t) (≤~5), a small-multiples grid, or a param→metric scatter (likely most-wanted for 'diff params'). Comparison must span sim↔device (overlay optimization fidelity with on-device data_hash/wigner_hash outcomes, not just waveforms). Depends on the hashing model → #64.