Codex agentic strata — 2026-09-07

The same 16 blast-radius tasks were then executed on the other two declared strata, at high reasoning effort on Codex CLI 0.153.4, 48/48 cells persisted in each, under the same repaired agentic_contracts prompt as Luna⁺ and Luna⁺⁺:

• Terra — results/codex-agentic-20260907T140422Z, gpt-5.6-terra against the shared clone and the locked index 3c584089…. Started 2026-09-07T14:04:22Z.
• Sol — results/codex-agentic-20260907T141122Z, gpt-5.6-sol, launched --agentic --isolated against a private run worktree holding a relocated copy of that index (6f2d3cd7…, same graph content at a different path). Started 2026-09-07T14:11:22Z.

These are the first agentic studies of either stratum, and they exist because of a defect rather than a plan: until 2026-09-07 the launcher's --models selection reached the structural lane only, and the agentic lane silently ran the manifest's default stratum. Every earlier "sol" or "terra" agentic invocation therefore produced a Luna study. A third artifact from the same afternoon, results/codex-agentic-20260907T140408Z, holds zero cells and a KeyboardInterrupt status: it is an interrupted launch, not a study, and nothing in this file reads from it.

Run	Pair	Cells	Control score	Treatment score	Gain	Cells correct	Tokens	Time
Terra	A_plain vs C_strict	16	0.932	0.970	+0.038	11 → 13	−3%	−31%
Terra	A_plain vs B_auto	16	0.932	0.826	−0.106	11 → 5	−1%	−33%
Sol	A_plain vs C_strict	16	0.967	0.991	+0.023	12 → 15	−21%	−40%
Sol	A_plain vs B_auto	16	0.967	0.855	−0.112	12 → 6	+54%	−14%

Score is the mean semantic quality over the scored cells; token and time are cohort totals against A_plain, the same estimator the two tables above use. Both runs scored all 16 cells in all three arms — no envelope loss, no incomplete cell, no contamination — so these are the cleanest Codex agentic artifacts in the record.

Sol is the strongest strict-arm result this lane has produced, and it is also the cleanest. Fifteen of sixteen cells correct against twelve, every C_strict cell adherent, 21% fewer input tokens, 43% less wall-clock at the paired median. It converts BA-01, BA-12, BA-13, and BA-15 and loses only BA-08, where both treatment arms land on 0.850 against a perfect control.

Terra's strict arm gains on the same tasks but pays nothing back in tokens. Its cohort totals are almost flat across the three arms — 3.12M, 3.08M, and 3.04M gross input for A_plain, B_auto, and C_strict — so the saving that every other stratum shows is absent here, and the paired median over its ten admissible pairs is only −14%. Frozen native output records a query in every strict cell. The six historical nonadherence records (BA-01, BA-04, BA-05, BA-12, BA-14, BA-16) instead reflect an observation mismatch: those cells used per-cell absolute launchers while the original rule credited only the literal variable form. Absolute paths need reviewed per-coordinate provenance in replay and are never inferred from a basename. The flat token totals are descriptive, not evidence that the tool was skipped. BA-12 is the extreme case at 574.0k input tokens against its control's 267.7k.

The optional arm now fails on every Codex stratum, not just on Luna. B_auto ends below its own control in all five Codex agentic executions up to this point — eight of eight once the balanced-order sweep is counted — and the two strata here are the worst of the five: 11 → 5 on terra and 12 → 6 on sol, both −37.5 points. Sol's optional arm also reads more input than its control — 3.90M against 2.53M gross, a +45% paired median — so on that stratum the available-but-optional integration costs both accuracy and tokens. Uptake does not explain it: B_auto reached for Codemap on 14 of 16 terra cells and 16 of 16 sol cells. Those counts are codemap_used, the observational signal, which credits a query issued inside a compound shell command; the stricter codemap_calls, which demands one standalone canonical query, reads 5 of 16 on terra and 14 of 16 on sol. Terra habitually chained its queries behind && or ;, so the gap between the two fields is a shell-style difference and not a difference in whether the tool ran. A_plain used it on none under either field, in both runs.

Expected-importer recall is 0.990 for A_plain and B_auto and 1.000 for C_strict in both runs; the single cell below full recall in each is BA-15, which the strict arm alone resolves completely. Discovery efficiency (DEFF) means run 3.56 / 2.71 / 2.47 on terra and 3.45 / 1.73 / 3.29 on sol — the one metric where the unaided control leads, because a grep-driven arm issues many cheap commands that each touch an expected importer.

Cached input is 86% of gross on terra's control and 84% on sol's, so the token columns here are gross-input claims with the same caveat the structural tables carry: sol's 21% gross reduction is a 21% fresh reduction (410.6k → 322.7k), which is the one place gross and fresh happen to agree.

These runs share the arm-order confound with every other table in this lane, and each is one execution at one repetition per cell. Two strata agreeing on the direction of B_auto is worth more than either alone; neither is a replication of the other's magnitude.

These two executions, together with Luna⁺⁺ above, held the Codex rows of the cross-provider table until the balanced-order sweep replaced them. Their binary paired rows are kept here so nothing published is lost when the table moves:

Run	Arm	Paired n	Control accuracy	Treatment accuracy	Gain	Tokens	Time
Luna⁺⁺	C_strict	16	68.8% (11/16)	81.2% (13/16)	+12.5 pp (+2)	−45%	−49%
Luna⁺⁺	B_auto	16	68.8% (11/16)	43.8% (7/16)	−25.0 pp (−4)	−15%	−35%
Terra	C_strict	10	70.0% (7/10)	80.0% (8/10)	+10.0 pp (+1)	−14%	−27%
Terra	B_auto	16	68.8% (11/16)	31.2% (5/16)	−37.5 pp (−6)	+7%	−32%
Sol	C_strict	16	75.0% (12/16)	93.8% (15/16)	+18.8 pp (+3)	−21%	−43%
Sol	B_auto	16	75.0% (12/16)	37.5% (6/16)	−37.5 pp (−6)	+45%	−15%