First-token identity accounting for the #1415 steering shift (parked: duplicates the 9a-ter cap-parked item)
Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.
No comments yet.
Goal
[parent Goal verbatim — identical to proposal 1's Goal field; same-question round]
2. First-token identity accounting (EXISTING — cap-parked, epm:progress v73; ranked, NOT re-filed) — Type: Diagnostic
Parent: #1415
question_relation: same
followup_label: first-token-identity-accounting
Goal: [parent Goal verbatim — identical to proposal 1's Goal field; same-question round]
Hypothesis: A large share of the first-bin norm spike is discrete token-identity change: cells with a higher fraction of steered draws whose first token differs from the baseline modal first token show proportionally larger first-bin shift norms (positive rank correlation across the 28×4 cells).
Falsification: No correlation between changed-first-token fraction and first-bin norm, or high first-bin norms at near-zero identity-change fractions — the spike would then be a same-token representational shift, strengthening (not just qualifying) the early-kick reading.
Differs from parent (round): Exactly ONE thing — a new pure-text reduction over the SAME stored completions (raw_completions/issue_1415/ + per_pair_profiles.json @ 6ceda65f66); zero new forwards.
Pre-filled spec (from parent): all fields same; analysis-only over existing artifacts (both Hub/git-verified above).
Estimated cost: 0 GPU-h (free analysis) If it works: Quantifies the token-identity share of the first-bin spike (currently carried only as caveat/limitation 4), tightening the "first ~5 tokens" Takeaways clause. If it fails: See falsification — either outcome refines the same clause; no wasted compute.
auto_run: no
auto_run_reason: The Step 9a-ter free-analysis round is already consumed on this task (epm:progress v73 cap-park; epm:free-analysis-followup-run v1 exists for the layer-sweep judging round) — this stays parked for manual pick-up post-promotion or a deliberate cap decision; do NOT double-file.
cost_class: free-analysis headline_affecting: no est_gpu_hours: 0
Value critique
Screened redundant by the follow-up-critic ensemble (epm:followup-value-critique v2 on #1415, 2026-07-23; single-Claude, Codex quota outage): duplicates the cap-parked free-analysis item recorded on #1415 (epm:progress followup-parked-by-cap followup_ref=first-token identity accounting, 2026-07-23T00:05:37Z — parked by the 9a-ter one-round cap, not by lack of value). Redundancy is PROPOSAL-LEVEL: the analysis was never run and the question is open. Revive via task.py set-status <this task> proposed OR run it as the parked 9a-ter pick-up post-promotion — do NOT double-file.