Round r34
r34
Source-reported narrativeRound README and results text are source material. Numeric values that are not tied to a claim ledger record remain source-reported.
Round source README · rounds/r34/README.md
Round date: HISTORICAL (before 2026-10-09) Publication badge: DESCRIPTIVE; KEPT FOR AUDIT Why not VALID: The historical inventory retains DESCRIPTIVE; the available public record does not support upgrading that classification. Recomputation status: OBSERVED SOURCE ONLY
Recomputation update: The mined raw records now provide per-cell outcome, manifest token-counter, and wall-time recomputation for the observed cohort. This does not reconstruct missing planned cells, a full intention-to-treat denominator, or model execution.
Status
DESCRIPTIVE
- NO_PREREG — The README says this round was not pre-registered, and its files show no timestamped rule predating the first result.
- NO_DECISION_RULE — The README says no public predeclared decision rule was recovered.
- RAW_BUNDLE_INCOMPLETE — The README says the public sources do not provide a complete round-specific request and grade bundle.
- NO_RESEARCH_QUESTION — The README says no testable research question is recoverable from the committed public record.
- EVIDENCE_GAPS — MEASURED.md and MISSING.md document missing per-cell execution, environment, or timing evidence that prevents full reproduction and audit.
Required reproduction metadata
- Kogen commit: not recorded in the cited round source.
- Harness commit: not recorded in the cited round source as a Git commit; any adapter or runtime fingerprints are not Git commits.
- Model and effort: a complete per-cell mapping is not recorded in the cited round source; values present in the source summary remain as stated above.
- Historical execution command: not recorded in the cited round source.
- Raw records: the cited sources do not provide a complete public round-specific request and grade bundle; public delivery and outcome exports are linked above where applicable.
Terminology: public round glossary.
STATUS: DESCRIPTIVE
Data completeness: 39.6% (15 deliveries; PASS WITH DECLARED GAPS). Measurement contract; declared gaps.
Pre-registered: no
Lifecycle: Historical capture; closure is not independently reconciled from the public record.
Question: Can a multi-stage Kogen build with a Luna low builder improve on direct Luna max outcomes for selected tasks? Design: Historical design note: Luna max investigator and scoped review with a Luna low builder; three selected tasks. The earlier direct Luna max public export in r27 recorded 0/3 passes for rails-as-variant-processed-once, 1/3 for elx-07-cli-stats and 1/3 for elx-12-retry-api-deprecation. These are small baseline samples, not a model ceiling. Five repetitions per task were planned across Studio for Rails, kogen-bench-us for elx-07 and kogen-bench-eu for elx-12.
Venue: Public host identifiers are listed in the reconciliation table. The surviving record does not establish every planned task-to-host assignment or the execution-environment details.
Decision rule: No public predeclared decision rule was recovered; outcome counts below are descriptive.
Arms: Arm definitions are not re-derivable from the public record; exported labels appear in the reconciliation table below.
Planned tasks in the surviving round note: rails-as-variant-processed-once, elx-07-cli-stats, elx-12-retry-api-deprecation
Verdict: The public run-record export records five ungraded Rails deliveries for rails-as-variant-processed-once. The official cells export contains ten graded Elixir rows, with 1/5 passes for each of elx-07-cli-stats and elx-12-retry-api-deprecation. The small r27 direct Luna max rates are bounded historical observations, not evidence of a model ceiling. No test or winner is inferred.
Public delivery and outcome reconciliation
This table is grouped from the public run-record export by audit round, task ID, public host ID, and the exact exported arm label. Captured and graded counts describe that ledger only; pass, fail, and other are its outcome fields. Official outcome states are published separately in cells.jsonl.
| Task ID | Public host ID | Exact arm label in run-record export | Captured | Graded | Pass | Fail | Other graded outcome |
|---|---|---|---|---|---|---|---|
| elx-07-cli-stats | kogen-bench-us | r34-elx07:agentic-lmax-review | 5 | 5 | 1 | 4 | 0 |
| elx-12-retry-api-deprecation | kogen-bench-eu | r34-elx12:agentic-lmax-review | 5 | 5 | 1 | 4 | 0 |
| rails-as-variant-processed-once | studio | agentic-lmax-review | 5 | 0 | 0 | 0 | 0 |
Recomputed from raw records
`recomputed.json`, mined cell records, and the token audit back observed official outcomes, per-arm pass counts/rates, manifest token counters, and wall medians. The source contains 10 rows (fail 8, pass 2); overall pass rate is 20.0% (2/10) among pass/fail grades. Per-arm detail and coverage counts are in the JSON.
Published-export outcome cross-check: 10/10 shared cell IDs match; 0/10 public-export rows have no mined source record. Exact IDs and results for unpaired rows are in recomputed.json.
No per-cell outcome mismatches were found in the shared IDs.
Token reconciliation under METHOD.md: 2/2 published metadata totals equal the manifest input + cached input + output sum; 0 differ (no mismatches). Reasoning remains a separate reported component and is not added to output.
The recompute keeps both numbers visible. 0 unexplained single-response numeric errors were identified; multi-response rows retain the published full-cell total and the manifest counter separately because per-response usage records are excluded.
Captured run records
Standard-schema records assigned to this round in the indexed snapshot.
Reproduction inventory
Published analysis replay and full experiment replay are reported separately.
- Replay type
unavailable- Script
- No per-round analysis script is published.
- Inputs
- No identified public input bundle.
- Full execution replay
- Unavailable
- Scope
- No public per-round analysis script with a complete, identified input bundle is published. See the round README, MEASURED.md, and MISSING.md for recorded scope and gaps; historical execution inputs/toolchains/grading are not bundled.
View the complete per-round reproduction inventory · Repository reproduction guide