Weight EMA

A native Codex session used Allagma to compare training duration and learning-rate policy, producing scientific code, saved model states, statistical estimates, this figure and a report. The result below is retained work from run r07.

EMA’s mean relative benefit was larger under constant learning rate than at completed cosine endpoints. Define Δ = SW1(EMA) − SW1(raw), with negative values favoring EMA. The schedule contrast Δ(cosine) − Δ(constant) was positive in all eight dataset/decay/duration comparisons. Raw-weight mean SW1 also improved under cosine in every dataset/duration comparison, while absolute EMA means changed less. A smaller relative benefit therefore does not imply worse absolute EMA quality.

Open the full-size figure.

Dots are the four paired seeds; diamonds and bars are means and unadjusted 95% Student t intervals. The figure’s EMA 099 and EMA 0999 labels mean decays 0.99 and 0.999. Each panel has its own scale. With four seeds, intervals are imprecise, multiplicity is substantial, and zero-containing intervals do not establish equality. The intervention is the complete learning-rate path; the study does not isolate a mechanism caused solely by its terminal value.

Finding	Observed result
Cell-level mean Δ below zero	14/16
Unadjusted cell-level 95% t intervals wholly below zero	2/16, both moons at constant 5k
Schedule-contrast means above zero	8/8
Duration-contrast intervals containing zero	8/8
Interaction intervals containing zero	4/4
GMM8 mode coverage	All 48 states covered all eight modes; this is a ceiling effect

Execution and verification

The first fully verified Allagma package in frozen run order is r07 (replicate 2). It is the illustrative package selected by the prespecified rule. The workflow comparison is complete; this individual result does not establish an Allagma advantage over plain Codex.

All 32 required cells and 96 raw/EMA model states were completed on this Mac’s MPS backend. There are four independent seeds per dataset. The 24 actual training trajectories total 200,000 updates: each constant-rate trajectory supplies its 5k and 10k endpoints, while cosine-5k and cosine-10k are separately trained from identical initial states and training-input prefixes.

The unchanged frozen scorer passes 240/240 checks, including replay of all 96 states from saved weights and recorded generation noise. Independent controller checks regenerate every seed’s initial weights and input streams, verify paired hashes and all terminal optimizer counters, and reproduce 114 statistical summaries. See the score, independent audit, and eight-item evidence review.

The agent recovered the controlled interruption and an MPS initialization failure without coordinator guidance. Its study runner lowered the allocator’s low watermark to 0.1 while preserving the broker’s high cap of 0.2. The shared framework’s separately validated rc2 correction now passes actual full-broker MPS checks; the original frozen outcome is unchanged. Charged compute was 362.39 seconds, setup 20.43 seconds and native wall time 1,511.68 seconds, with 35 compute requests. Failures remain retained. The same-assistant critique and controller audit are provisional, not independent scientific peer review. AI-generated/adapted code and reporting are disclosed, with the supplied AI Scientist Source Code License retained.

Read the agent’s full report. This exact selected copy retains its source license and disclosure; its relative references resolve inside the full package below.

Restore and reproduce

The package index retains the exact source, protocol, actual training inputs, weights, samples, traces, statistics, report and revision-bound critique. Its 829,100,336-byte archive is split into 25 parts; software wheels are hydrated by exact hash.

python3 evals/research-v0.3/retention.py restore \
  --source evals/research-v0.3/runs/r07/package \
  --destination /path/to/new-ema-package --download-wheels

An exact local wheel cache can replace downloads. Restoration streams the archive without creating an oversized intermediate file. All 901 declared manifest entries and 272 reviewed references pass after restoration. Read the restored REPORT.md, REPRODUCE.md and submission.json, and use the standalone broker with a new finite profile to service the declared full-reproduction dispatcher. It creates a fresh study environment, trains all 24 trajectories and recalculates the analysis. Raw-only recomputation uses the isolated study environment and an unused output directory. Repeated raw analysis and saved-state regeneration were executed; a second full training campaign was not. The separate full clean-checkout release gate passed on the CORE CULP study.

---

Canonical source · Generated from the maintained repository document at this build’s commit.