Phase 3 — Analysis + artifact

[Tejas-JB]

PRD §6 Phase 3. Blocked by Phase 2.

Deliverables

• Compute the four PRD §3.4 metrics with mean ± std across seeds:
  • Flagged retention (primary)
  • Selective retention gap = flagged_retention_cued − unflagged_retention_cued
  • Collateral cost = unflagged_cued − unflagged_uniform
  • No-confound check from Phase 1 (learn-phase indistinguishability)
• results/retention-curve.png — 3 arms × {flagged, unflagged} = 6 lines, x=interference step, y=probe accuracy. Mean line + shaded std band.
• results/summary-table.csv — arm × {flagged_final, unflagged_final, selective_gap}.
• RESULTS.md — headline number, plot embed, table, honest caveats.
• Tag v1.0.

Artifact scope (owner-confirmed)

• (a) retention-curve figure
• (b) summary table
• (c) RESULTS.md with headline + selective gap + collateral cost + honest caveats
• (d) reproducible repo

Not in scope here: the reel itself. (B-roll is being captured in notes/reel-broll.md throughout.)