Codex agentic re-executions — 2026-09-07

The same 16 blast-radius tasks were executed twice more on gpt-5.6-luna, at high reasoning effort on Codex CLI 0.153.4, 48/48 cells persisted in each:

• Luna⁺ — results/codex-agentic-20260907T065010Z, launched --agentic --isolated, which builds a private run worktree and relocates the locked index into it (e0ce11d9…, the same index content at a different path). Started 2026-09-07T06:50:13Z.
• Luna⁺⁺ — results/codex-combined-20260907T055156Z/agentic, the agentic stage of the combined launch whose structural half ran terra, against the shared clone and the locked index 3c584089…. Started 2026-09-07T09:51:11Z.

Both ran the repaired agentic_contracts prompt; the 2026-09-06 run did not, so these two are executions of one contract and the older run is not a third repetition of it. The two 09-07 manifests differ only in the structural-manifest and launcher hashes they carry, not in any agentic material.

Luna⁺⁺ held the Luna row of the cross-provider agentic table until the balanced-order sweep replaced it; the row it held read 16 pairs, 11 → 13 correct at −45% tokens and −49% time. All three Luna executions are now reported here and nowhere else: one model with several studies published as adjacent rows reads as a repetition design, which it is not.

Run	Pair	Cells	Control score	Treatment score	Gain	Cells correct	Tokens	Time
Luna⁺	A_plain vs C_strict	15	0.959	0.971	+0.012	11 → 11	−65%	−62%
Luna⁺	A_plain vs B_auto	15	0.959	0.844	−0.115	11 → 5	−32%	−43%
Luna⁺⁺	A_plain vs C_strict	16	0.959	0.985	+0.026	11 → 13	−51%	−46%
Luna⁺⁺	A_plain vs B_auto	16	0.959	0.851	−0.108	11 → 7	−6%	−28%

Score is the mean semantic quality over the scored cells; token and time are cohort totals against A_plain, the same estimator the 09-06 table uses. Luna⁺ reports 15 cells rather than 16 because its A_plain BA-06 cell exited without an answer (incomplete) — an absent answer, not an envelope loss.

Luna⁺'s C_strict row is the one place the two views disagree in sign, and the reason is the admission rule rather than the model: over its 15 scored cells the strict arm converts BA-13 and BA-15 and loses BA-07 and BA-11, for 11 correct against 11. The binary cross-provider row drops BA-13 and BA-14 for non-adherence — BA-13 being one of the two cells the arm converted — which is what turns 11 → 11 into 10 → 9. Read the −7.7 pp there as the cost of removing a converted cell, not as a regression the arm produced.

The direction reproduces; the size does not. Across the three executions C_strict scores +0.031, +0.012, and +0.026 against its control while reading 48%, 65%, and 51% fewer input tokens, and B_auto scores −0.069, −0.115, and −0.108 while reading 21%, 32%, and 6% fewer. Every run has the strict arm up and the optional arm down; no run has them within noise of each other. Correct-cell counts move the same way — B_auto ends below its own control in all three (9 → 6, 11 → 5, 11 → 7) — and the same task, BA-07, is lost by B_auto in all three and by C_strict in all three, which makes it the clearest single-task counterexample in this lane.

The envelope repair holds. The 09-06 run lost 12 of 48 cells to the missing wrap instruction; Luna⁺ and Luna⁺⁺ lose zero, and Luna⁺⁺ is the first Codex agentic run in this record with no non-poolable cell and no non-adherent cell — 48 admissible cells out of 48. (Claude's Haiku and Opus tiers already pair 16 of 16 in both arms on their own 2026-09-06 artifact.) Voluntary Codemap use in B_auto was 16/16, 14/16, and 16/16, and A_plain used it in none, so no control is contaminated in any run.

Expected-importer recall is 0.99 for A_plain and 1.00 for both treatment arms in each 09-07 run, with exposure hits per command (DEFF) 1.84 / 1.76 / 2.60 on Luna⁺ and 2.56 / 1.75 / 2.53 on Luna⁺⁺. Both absolutes carry the recall and floor caveats below.

These runs share the arm-order confound with every other agentic table: arms ran A_plain → B_auto → C_strict per task with no provider cache reset, so the elapsed reductions overstate the treatment effect.