Second-Stage Planning for Multi-Step Questions / V122–V122.2
Released baseline language models used on their own as planners for multi-step calculation questions committed 6 to 10 wrong answers each on familiar question kinds. Used only as a second stage where EMMA’s own planner withholds, and held to the same checks, every combination committed no wrong answer; the selected one, used without training, answered 19 of 32 held-out question kinds. Two whole-demonstration rehearsals then adopted it.
SCOPE + LIMITATIONS
- One family of calculation questions and one sealed set. Second-stage answers take about 12 seconds, and about 30 seconds in the demonstration while the stage loads per use. Development data revealed two gaps in the checks, both fixed before the sealed run.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
A sealed set of 109 questions: 52 familiar multi-step and direct questions, 15 that should be declined, 32 of question kinds held out from training, and 10 outside the supported grammar. Four released reference models are scored alone, one also with a small trained adapter, and each as a second stage behind EMMA’s own planner. Every committed answer must pass the same step and coverage checks. The selection rule, fixed before any result, is fewest wrong commits, then most right.
Every reference model failed alone; every cascade passed all eight gates. The selected cascade keeps the precision of EMMA’s own planner and adds held-out kinds with no training. The first rehearsal failed its memory gate because the second stage stayed loaded beside the document reader; releasing it after each use fixed this in V122.1. The rehearsals’ learning gates failed as in V120.2 and were declared non-deciding before the run; that question is answered separately (V123–V124.1).
- EMMA’s own planner alone
- 0 wrong; 52 of 52 familiar right; held-out kinds withheld
- Reference models alone
- 6 to 10 wrong commits each on familiar kinds
- Selected cascade · familiar
- 0 wrong; 100% multi-step; 10 of 10 direct
- Selected cascade · held-out kinds
- 19 of 32 right; 0 wrong
- Every cascade · decline questions
- 15 of 15 declined; none committed
- V122.1 rehearsal
- Adoption gates passed; peak memory 13.3 GB; 6 of 6 reasoning questions right
- V122.2 rehearsal
- Exact what-if questions routed to the planner; adoption gates passed
Question and method
A sealed set of 109 questions: 52 familiar multi-step and direct questions, 15 that should be declined, 32 of question kinds held out from training, and 10 outside the supported grammar. Four released reference models are scored alone, one also with a small trained adapter, and each as a second stage behind EMMA’s own planner. Every committed answer must pass the same step and coverage checks. The selection rule, fixed before any result, is fewest wrong commits, then most right.
Recorded results
| Condition | Recorded result |
|---|---|
| EMMA’s own planner alone | 0 wrong; 52 of 52 familiar right; held-out kinds withheld |
| Reference models alone | 6 to 10 wrong commits each on familiar kinds |
| Selected cascade · familiar | 0 wrong; 100% multi-step; 10 of 10 direct |
| Selected cascade · held-out kinds | 19 of 32 right; 0 wrong |
| Every cascade · decline questions | 15 of 15 declined; none committed |
| V122.1 rehearsal | Adoption gates passed; peak memory 13.3 GB; 6 of 6 reasoning questions right |
| V122.2 rehearsal | Exact what-if questions routed to the planner; adoption gates passed |
Comparative standing and locked gates
Every reference model failed alone; every cascade passed all eight gates. The selected cascade keeps the precision of EMMA’s own planner and adds held-out kinds with no training. The first rehearsal failed its memory gate because the second stage stayed loaded beside the document reader; releasing it after each use fixed this in V122.1. The rehearsals’ learning gates failed as in V120.2 and were declared non-deciding before the run; that question is answered separately (V123–V124.1).
Interpretation and limitations
One family of calculation questions and one sealed set. Second-stage answers take about 12 seconds, and about 30 seconds in the demonstration while the stage loads per use. Development data revealed two gaps in the checks, both fixed before the sealed run.
SOURCE PROVENANCE
Second-Stage Planning for Multi-Step Questions / V122–V122.2
LABORATORY REPORT / 2026-10-27SOURCE CHECKSUM / SHA-256
3ce565e86df03547a81d91d6459ae69dedfc9d5b2fa77b28bd6a2d9840786eadPublic journal edition reviewed 2026-10-28. Source documents and saved evidence were inspected; experiments were not rerun for this edition. Proprietary implementation code, model binaries, private infrastructure, and detailed machine records are not published here. Journal identifiers are editorial references. Catalog inclusion does not imply qualification or runtime promotion.