Complete nine-model text baseline: result-to-claim review
PARTIAL — a complete descriptive baseline and a controlled label-dependent tradeoff; no established method win. Fresh-context delegated review; same-family, provisional, [pending external review]. All nine fits are retained with their registered seeds. No method, paper or test gate advances.
On the reused 40-task/160-attempt development cohort, changing the target of the fixed single-head text baseline from final to current correctness changed its positive-score stopping tradeoff. Across all three registered fits, deterministic resolve-loss calibration increased mean generator-input saving by 10.8117 percentage points (conditional paired task-bootstrap 95% interval 1.8568 to 18.1361), while mean quality was 2.2917 points lower and harm 2.5000 points higher. The current-label policy had mean harm 3.1250% and net quality loss 2.9167 points relative to native. This supports a bounded label-ablation observation, not risk-qualified method superiority, a stable increment over counts, population safety, or net compute savings.
What the completion establishes
The record contains joint-now/final, independent-now and independent-final fits for seeds 0/1/2: 28 score/policy/calibration groups, 84 seed-policy points and 20 paired contrasts. Every fit completed with exit 0 and a closed allocation. They share the fixed configuration, dataset path, ordered token caches, 296 extension training tasks, 13,178 decisions and 3,296 reported updates. Evaluation uses the same 40 development tasks and 160 attempts. Native quality is 77/160 = 48.125%.
Independent arithmetic recomputation matched all 4,844 checked numeric values exactly, including every aggregate mean, sample SD, interval endpoint, paired contrast and logical classifier-cost aggregate. All 18 summary source hashes matched. The deterministic evidence precheck verifies existence of 80 cited contrast values; existence does not establish semantic support. This review read completed summaries/reports, source code and prior review receipts; it did not read the raw dataset, test labels, active unknown outputs, predictions or tensors, run model forwards, or refit anything.
The prior per-seed reviews reconstructed policy arithmetic, checked all saved-draw quantiles and checked model artifacts; their bounded recalibration replay used the first 64 of 2,000 draws. This aggregate review does not claim to regenerate all 2,000 calibrations or establish numerical prediction equivalence from weights. The root audit remains [INTEGRITY: WARN] and concerns repairedP2/development/no-online scope, rather than being a dedicated full integrity audit of the nine fits.
Final seed1 is the explicit recovery4 successor after the original final1 stopped at update 2,160 with a recorded container exit137 and no completed effect outputs. Earlier recovery preflight failures are retained. Final seed2 is the explicit successor after the original queue was lost; its original final2 launch was not recorded. Both successors retain the original seed/config/data/cache identities and count once. They are neither additional seeds nor selected replacement outcomes. Recorded provenance is consistent with infrastructure recovery; it cannot exclude unrecorded attempts and does not establish what caused exit137. Successful-run training hours omit predecessor/recovery overhead.
Pure label contrast
The independent-now and independent-final paths use the same single-head architecture, training examples/weights, optimization recipe, seeded initialization path, score orientation, first-crossing action, threshold menu and calibration protocol. The changed supervision target is therefore isolated at the recipe level. It does not hold learned score scale, calibrated thresholds, stop rate or achieved risk constant. Positive final confidence and the separate 1−final failure rule answer different questions.
All differences below are current-label minus final-label in percentage points. Intervals are the saved conditional paired task-bootstrap 95% intervals.
| Policy / calibration | Δ input saving [95%] | Δ output saving [95%] | Δ quality [95%] | Δ harm [95%] |
|---|---|---|---|---|
| families.harm | 6.7100 [-0.5522, 15.0694] | 3.8309 [-1.7588, 9.6058] | -0.4167 [-3.5417, 2.0833] | 0.6250 [-1.6667, 3.5417] |
| families.resolve_loss | 10.8117 [1.8568, 18.1361] | 6.5828 [-0.1447, 11.7056] | -2.2917 [-4.3750, 1.2500] | 2.5000 [-0.8333, 4.3750] |
| mixtures.harm | 6.1254 [-0.1140, 11.9726] | 3.0252 [-1.6676, 7.0710] | 0.2937 [-1.5348, 1.7878] | -0.0204 [-1.2835, 1.7863] |
| mixtures.resolve_loss | 9.4318 [2.0145, 16.3180] | 5.2431 [-0.4203, 10.0935] | -1.3188 [-2.8648, 0.9573] | 1.5966 [-0.6352, 3.2638] |
The deterministic resolve-loss input advantage is positive for all three paired seed points: 14.0071, 11.8790 and 6.5492 points. Its quality difference is negative for all three: −3.7500, −2.5000 and −0.6250 points. Thus even the within-seed pattern is a tradeoff. The mixture resolve-loss input interval is also above zero, but its current-label mean H is 3.6833% and net loss is 3.4055 points. Both exceed the registered 1%/1-point observed limits. The deterministic current-label resolve-loss point exceeds both limits in every seed.
The deterministic harm-calibrated current-label mean H is 1.25% and mean net loss is 1.0417 points; its input-saving interval crosses zero. Neither harm-calibrated policy family establishes an input/output-saving increment. Quality and harm difference intervals crossing zero do not establish risk matching, equivalence or noninferiority. Output-saving intervals cross zero in all four label cells. The allowed wording is “changing supervision changes this calibrated stopping tradeoff,” with the adverse risk/quality points included. The wording “current labels improve low-risk efficiency,” “final labels are useless,” or “the method succeeds” is not supported.
Three-seed description and deployment limits
The following table includes all 28 groups. Savings are percentages; SD is the three-seed sample SD, in percentage points. Q and H are percentage rates and ΔQ is relative to native, in points. Means describe three fixed fits; they are not one deployed policy.
| Group | Input mean ± SD | Output mean ± SD | Q | H | ΔQ |
|---|---|---|---|---|---|
| joint.families.now.harm | 5.1921 ± 3.0883 | 3.4766 ± 1.7051 | 46.8750 | 1.2500 | -1.2500 |
| joint.families.now.resolve_loss | 7.5690 ± 7.0822 | 5.0359 ± 4.3412 | 46.0417 | 2.0833 | -2.0833 |
| joint.families.final.harm | 4.0452 ± 3.6561 | 2.8894 ± 2.5951 | 47.0833 | 1.0417 | -1.0417 |
| joint.families.final.resolve_loss | 4.0452 ± 3.6561 | 2.8894 ± 2.5951 | 47.0833 | 1.0417 | -1.0417 |
| joint.families.gain.harm | 9.7134 ± 1.8596 | 6.0791 ± 1.5572 | 47.9167 | 0.8333 | -0.2083 |
| joint.families.gain.resolve_loss | 11.2660 ± 1.9425 | 7.1994 ± 1.8787 | 47.5000 | 1.4583 | -0.6250 |
| joint.families.failure_termination.harm | 7.2786 ± 6.2242 | 5.4264 ± 4.6181 | 46.8750 | 1.2500 | -1.2500 |
| joint.families.failure_termination.resolve_loss | 7.2786 ± 6.2242 | 5.4264 ± 4.6181 | 46.8750 | 1.2500 | -1.2500 |
| joint.mixtures.now.harm | 10.4448 ± 2.5116 | 7.0329 ± 1.3947 | 45.8467 | 2.3511 | -2.2783 |
| joint.mixtures.now.resolve_loss | 12.4402 ± 3.9949 | 8.3774 ± 2.2589 | 45.2258 | 2.9719 | -2.8992 |
| joint.mixtures.final.harm | 6.3986 ± 3.0439 | 4.8908 ± 2.1127 | 46.0100 | 2.1150 | -2.1150 |
| joint.mixtures.final.resolve_loss | 6.3986 ± 3.0439 | 4.8908 ± 2.1127 | 46.0100 | 2.1150 | -2.1150 |
| joint.mixtures.gain.harm | 11.8534 ± 0.5793 | 7.9313 ± 0.2854 | 47.0180 | 1.9444 | -1.1070 |
| joint.mixtures.gain.resolve_loss | 16.0801 ± 2.8003 | 11.1719 ± 2.3618 | 46.0975 | 2.9722 | -2.0275 |
| joint.mixtures.failure_termination.harm | 10.2823 ± 5.6680 | 8.1247 ± 3.8989 | 45.9495 | 2.1755 | -2.1755 |
| joint.mixtures.failure_termination.resolve_loss | 10.3723 ± 5.5123 | 8.2118 ± 3.7484 | 45.9032 | 2.2218 | -2.2218 |
| now.families.now.harm | 9.1651 ± 5.9975 | 5.8916 ± 3.7028 | 47.0833 | 1.2500 | -1.0417 |
| now.families.now.resolve_loss | 13.2669 ± 3.8725 | 8.6434 ± 2.3962 | 45.2083 | 3.1250 | -2.9167 |
| now.mixtures.now.harm | 11.8188 ± 3.5198 | 7.7588 ± 1.9323 | 46.3320 | 2.0663 | -1.7930 |
| now.mixtures.now.resolve_loss | 15.1253 ± 2.2147 | 9.9767 ± 1.1925 | 44.7195 | 3.6833 | -3.4055 |
| final.families.final.harm | 2.4551 ± 1.7300 | 2.0607 ± 1.6167 | 47.5000 | 0.6250 | -0.6250 |
| final.families.final.resolve_loss | 2.4551 ± 1.7300 | 2.0607 ± 1.6167 | 47.5000 | 0.6250 | -0.6250 |
| final.families.failure_termination.harm | 10.1253 ± 6.5850 | 7.6357 ± 5.2549 | 47.7083 | 0.4167 | -0.4167 |
| final.families.failure_termination.resolve_loss | 11.2239 ± 4.9677 | 8.3593 ± 4.1699 | 47.0833 | 1.0417 | -1.0417 |
| final.mixtures.final.harm | 5.6935 ± 0.4407 | 4.7336 ± 0.1834 | 46.0383 | 2.0867 | -2.0867 |
| final.mixtures.final.resolve_loss | 5.6935 ± 0.4407 | 4.7336 ± 0.1834 | 46.0383 | 2.0867 | -2.0867 |
| final.mixtures.failure_termination.harm | 12.7525 ± 4.6955 | 9.8053 ± 4.1496 | 46.5606 | 1.7067 | -1.5644 |
| final.mixtures.failure_termination.resolve_loss | 13.8409 ± 3.7133 | 10.6272 ± 3.6088 | 46.3589 | 1.9083 | -1.7661 |
Completeness supports honest variance reporting, not blanket stability. For independent-now/harm, input saving ranges from 2.8421% to 14.7730%, and harmful-stop counts range from zero to four. Joint gain/harm has mean H=0.8333%, yet seed0 has H=1.25%; averaging cannot make all seeds risk-qualified. Its input saving mean is 9.7134% with SD 1.8596 points, but this alone establishes neither a new method nor a counts increment.
Independent-final failure/harm is descriptively within both observed risk limits in all three seeds. Input savings are 3.3577%, 10.5071% and 16.5111% (mean 10.1253%, SD 6.5850 points); quality is 77/160, 76/160 and 76/160. Its descriptive mean H is 0.4167%, with conditional interval [0.4167%, 3.1250%]. The varying magnitude, lost native successes, unestablished counts increment and unmeasured online overhead prevent a deployment recommendation. Final-positive/harm itself fails the observed limits in seed2. A low mean risk is not certification of a chosen fit.
All registered contrast cells and the counts boundary
| Contrast | Δ input saving [95%], pp | Δ output saving [95%], pp |
|---|---|---|
| sharing_now.families.harm | 3.9730 [-3.7239, 5.8185] | 2.4150 [-2.5957, 3.7623] |
| sharing_now.families.resolve_loss | 5.6978 [-6.7105, 7.1763] | 3.6075 [-4.8503, 4.6860] |
| sharing_now.mixtures.harm | 1.3740 [-4.0783, 3.6638] | 0.7260 [-2.9668, 2.4352] |
| sharing_now.mixtures.resolve_loss | 2.6850 [-6.1284, 4.6263] | 1.5993 [-4.6761, 2.8800] |
| sharing_final.families.harm | -1.5901 [-4.0423, 4.6497] | -0.8288 [-2.7103, 4.2342] |
| sharing_final.families.resolve_loss | -1.5901 [-4.1407, 4.6497] | -0.8288 [-2.7869, 4.2342] |
| sharing_final.mixtures.harm | -0.7051 [-2.5062, 3.6443] | -0.1572 [-1.5147, 3.4385] |
| sharing_final.mixtures.resolve_loss | -0.7051 [-2.6238, 3.6202] | -0.1572 [-1.5749, 3.4295] |
| sharing_failure.families.harm | 2.8467 [-1.7494, 10.6280] | 2.2093 [-2.1739, 8.3508] |
| sharing_failure.families.resolve_loss | 3.9453 [-1.0495, 11.5788] | 2.9329 [-1.5088, 8.9885] |
| sharing_failure.mixtures.harm | 2.4702 [-2.1251, 8.4144] | 1.6806 [-2.4654, 6.4158] |
| sharing_failure.mixtures.resolve_loss | 3.4686 [-1.5613, 9.4314] | 2.4154 [-1.9521, 7.1033] |
| label_positive_stop.families.harm | 6.7100 [-0.5522, 15.0694] | 3.8309 [-1.7588, 9.6058] |
| label_positive_stop.families.resolve_loss | 10.8117 [1.8568, 18.1361] | 6.5828 [-0.1447, 11.7056] |
| label_positive_stop.mixtures.harm | 6.1254 [-0.1140, 11.9726] | 3.0252 [-1.6676, 7.0710] |
| label_positive_stop.mixtures.resolve_loss | 9.4318 [2.0145, 16.3180] | 5.2431 [-0.4203, 10.0935] |
| against_final_failure_stop.families.harm | -0.9602 [-14.6577, 11.1409] | -1.7442 [-12.1080, 6.7664] |
| against_final_failure_stop.families.resolve_loss | 2.0429 [-14.2865, 13.3473] | 0.2841 [-11.9235, 8.4861] |
| against_final_failure_stop.mixtures.harm | -0.9336 [-13.7738, 10.0394] | -2.0465 [-12.0986, 5.5731] |
| against_final_failure_stop.mixtures.resolve_loss | 1.2844 [-13.5803, 12.9968] | -0.6505 [-12.1756, 7.2321] |
All four sharing-now input-saving intervals contain zero. Independent-vs-joint training changes loss/head structure and does not establish negative transfer. The now-versus-final-failure contrasts compare different stopping rules and do not support a general rejection of final supervision. The full structured report retains every metric and interval for every contrast, not only favorable input cells.
The nine-model summary contains no seed-aggregated counts contrasts. Prior per-seed reviews do not establish a stable practical text increment over the fixed counts controls. Counts failure/harm already gives 77/160 observed successes with 17.7944% input and 12.4584% output saving, one harm and one rescue. That is a stronger descriptive control than positive-final stopping, but not itself a population safety guarantee. Per-seed intervals cannot be averaged, and per-seed non-significance does not prove an aggregate null. If aggregate counts inference is wanted, compute all existing matched contrasts from aligned saved draws before claiming it; this review has not computed those intervals. Neither universal text inferiority nor equivalence follows.
Uncertainty and cost interpretation
Each draw resamples tasks jointly across all three fixed fits, recalibrates thresholds/mixtures within the existing five folds, averages seed metrics in the draw, then takes quantiles. Paired contrasts subtract the aligned draw-level averages. There are 40 task clusters, not 120 independent tasks; the 160 attempts also are not treated as independent clusters. Model weights remain fixed. The procedure has no training-randomness interval, no correction for repeated development selection or the family of reported contrasts, and no population H/noninferiority guarantee. Three-seed SD is descriptive. Mixtures are analytic expectations under a threshold drawn at episode start, not integer observed outcomes. Cross-fitted thresholds are not the final deployment policy.
For deterministic resolve-loss label comparison, current/final mean generator input is 2,801,272.8271 / 3,150,465.8083 tokens per planned attempt; output is 20,333.9771 / 21,799.1542. Additional logical 0.6B classifier input is 49,396.8104 / 40,402.8042 tokens with 21.8479 / 18.1146 queries. Classifier generated tokens are zero, but inference cost is not. Generator input, generator output and classifier input are separate model-specific accounts. No logical-cost interval, online latency, padded FLOPs, energy or net GPU saving was measured. Training/recovery expenditure is separate again.
Claim decision and next action
- A: yes for complete descriptive means/variance; no blanket stable-performance claim.
- B: partial for the controlled label-dependent tradeoff; no risk-qualified method superiority.
- C: no established text increment over counts; aggregate counts intervals are absent.
- D: no deployment conclusion. Retain low observed-harm failure points without promoting them to population safety, native-quality preservation or net efficiency.
No further label-only fits are needed to report this completed narrow ablation. The next scientific step is method refinement under the explicit constraint that a new mechanism must improve on counts at the registered risk/quality standard. The positive resolve-loss input contrast is insufficient to advance directly to submission-oriented main-method ablations. A subsequent candidate should have a predeclared primary risk/cost endpoint and minimal mechanism ablation against the counts and text controls, with all registered seeds retained. Any negative-transfer mechanism would need controlled loss/head/optimization evidence beyond these sharing contrasts.
Only a candidate that survives development review should proceed to a fixed deployment rule, online equivalence and measured cost, full evaluator/policy freeze, and untouched evaluation under the project protocol. No new candidate, seed, task, policy, queue action or test access is commissioned by this report. The bounded decision is high-confidence; semantic assurance remains same-family/provisional pending external review.
Sources: runs/text-three-seed-summary-successors-20260929.json; its 18 hashed result/cost sources; all nine refine-logs/CLAIMS_FROM_TEXT_*.{md,json} per-seed reports; TEXT_THREE_SEED_SUMMARY_PLAN_20260928.md, TEXT_LABEL_CONTRAST_PLAN_20260928.md, recovery plans/receipts; and the source evaluator/aggregator. Exact read-source hashes and fresh arithmetic/evidence receipts are in .aris/traces/result-to-claim/2026-09-29_text-full-summary/. The local trace contains findings/routing; global status, findings, paper, queues and test state were unchanged.