[codethor0]

Summary

Model transparency, part 2 of 3. Makes the difference between a real measurement and the normalized model response obvious, and makes it easy to see why a tile moved and how much that depends on editorial assumptions.

• Baseline versus response. Each tile shows its stored real-world baseline, with its period, under the title. The 0 to 100 readout is labeled MODEL RESPONSE ("normalized, automatic") and is never called a result, forecast or prediction.
• Paths. Why & source lists every path strongest first. Each step shows its direction and the relationship's kind, citation status and horizon (for example "Causal | modeled | long run"). Every relationship is labeled the same way, and a dead end states why it drives nothing.
• Sensitivity. src/model/analysis.ts reruns the scenario with the same engine under nine nearby settings: decay 0.5, 0.7 and 0.9, and up to 2, 3 or 4 relationships. It classifies each tile as same direction under all settings, moves under some settings only, or direction depends on the settings. It is explanatory, not a second prediction engine. Hand-worked tests cover every class, clamping, noise and sign symmetry.
• Category filter. A radio group by category. It changes visibility only; hidden tiles still take part in the calculation, and the address bar does not change.
• Cited relationships only. The same engine on empirical edges alone. None exist yet, so the board says that no input reaches another tile in this view, and that this is a gap in citations, not evidence of no effect. The mode is deliberately not part of Share links: links carry inputs only and open with every relationship, and the page states this.
• Scenario A. Save the current results in memory for this tab only. Each tile then shows A beside the current position. Reset and history change the inputs, not A. Nothing is written to the address bar or to storage.

No new dependency and no new color token. Built JS goes from 266.5 kB to 275.5 kB (gzip 84.2 to 86.8 kB); CSS from 10.1 to 11.3 kB, including all three parts.

Validation

• npm run verify passes: 754 tests, 100% coverage on the model and display helpers, and the numerical audit passes.
• npm run test:e2e passes. It now drives the filter, cited-only mode and scenario A at 1440, 390 and 320 px: hidden tiles, no modeled edge used, the comparison text, no horizontal overflow, an unchanged address. It also runs axe WCAG 2.2 A/AA with all three active.
• npm audit --audit-level=low: 0 vulnerabilities. No dependency changes.
• No secrets, credentials, private logs, local machine paths, or generated build output are included.
• The pull request contains no unrelated changes.

Data or model changes

Not applicable. No data or relationship changes. The propagation engine is unchanged; analysis reuses combineEffects.

Security

None. No new input source. The view state (filter, cited-only mode, scenario A) is in memory only and never read from or written to the URL. CSP and the analytics loader are unchanged.