22 lines
2.3 KiB
Markdown
22 lines
2.3 KiB
Markdown
# Conditional H2 phase3 plan — record only
|
||
|
||
Source: Claude's actual Slack ruling1794928372.222109, independently read on2026-11-18: https://arcticnimbus.slack.com/archives/C0C78UFKGRM/p1794928372222109 . This records the ruling, not a scientific launch or an independent review of the referenced specification bodies.
|
||
|
||
The three MoldAstra specifications are accepted by Claude as contract inputs: a=`5dce5d0f`, b=`de120e9d`, c=`9974ae27`. His amendment and execution order supersede their earlier sequence. **All items wait for the S2 pooled result CONFIRMED. If S2 does not confirm, none runs; Claude replans from the data-sufficiency question.** A failed S2 result does not itself authorize the data-sufficiency test.
|
||
|
||
| Conditional order | Item | Planned GPU hours |
|
||
|---|---|---:|
|
||
| 1 | Data-sufficiency test, reference `e24fff7c` | 24 |
|
||
| 2 | Spec a: seed1, five folds; paired seed1−seed0 training-randomness reference | about33 |
|
||
| 3 | Spec c: last-fill objective inside H2, using its original stage design | 34–38 |
|
||
| 4 | Spec b: H2+B13 context | about33 |
|
||
| 5 | Remaining spec a: seed2 and three-model ensemble, if budget remains | about33 |
|
||
|
||
The new aggregate phase3 ceiling is170 GPU-hours. Each item requires its own first-epoch cost projection; an overrun stops and is reported. These are conditional budgets, not measured execution costs or authorizations to bypass the S2 gate or item-specific contracts.
|
||
|
||
Seed1 precedes b/c because a component bootstrap captures variation across parts, not training randomness. In addition to each original CI gate, b and c must each have a point effect whose absolute magnitude exceeds the seed1−seed0 difference on the same endpoint; report both. Seed2 and the ensemble come last.
|
||
|
||
Spec c's fresh-family confirmation is allowed only after gate calibration v2 passes. The actual current result is FAIL_OUTCOME_EQUIVALENCE, so that condition is unsatisfied; this record does not open fresh-family work or silently revise the stage design. Claude's subsequent interpretation remains pending. See `docs/phase2-gate-outcome-results-2026-11-18.md`.
|
||
|
||
No phase3 code, test, training, prediction or experiment is implemented or launched by this record. Existing S2 training continues unchanged. The current gate result is archived and submitted for the standing collaborative stage review.
|