Authorized prospective half-budget control screens¶

Omar approved the recommended staged design on 13 October local time (12 October UTC): screen unstarted CER and gamma comparisons at half their training budget, retaining the Coins replacements and verification budgets. All submitted scientific jobs continue. Admissions were briefly frozen to record a stable split, with the original ledgers, commands, environment context and deployment files preserved under each root's half-budget-transition-20261013/.

Comparison groups and retained work¶

An entire comparison remains at full budget if any of its profiles had already been submitted. Gamma groups are task/learner pairs and include both discounts and all five seeds. CER groups are task/learner pairs and include all modes and all five seeds. This avoids mixing budgets within a comparison or choosing individual seeds by their outcomes.

The retained gamma groups include Harvest and Classic 2, plus the already submitted IPPO families in Capture 2; their original worker commands and source pins remain unchanged. Both targeted CER cases were underway, so their complete fifty-profile ablation stays at full budget. The broader seven IPPO task comparisons, including matched no-CER, CER and RM-shaping arms, receive the half-budget screen. The complete frozen manifests, rather than this prose, are authoritative for each retained comparison.

The initial workload decreases from roughly 510 million requested training steps to 326 million, before rollout rounding. This saves about 36 percent of training across these controls, rather than 50 percent, because started comparisons are retained. Evaluation costs are unchanged. A later full-budget confirmation is additional work and can reduce or eliminate this saving.

Adapter and provenance¶

scripts/staged_control_study.py runs the frozen worker code through an isolated adapter. The base sources remain gamma c84a075, PR2 c2dbf27 and CER 612145f; the adapter has its own pinned revision and file hash. No frozen worker source, trainer, archived checkpoint or shared remote environment is edited. Each manifest records the base and adapter revisions, all worker commands, source directories, retained groups, original rounded budgets, cutoff budgets and full schedule horizons.

Half runs use the ceiling of half the original number of complete rollout updates. Their learning-rate and exploration schedules retain the original SCHEDULE_TOTAL_UPDATES. They are halfway observations of the planned schedule, rather than shorter runs with a compressed schedule. Batch shape, learner architecture, discount, CER mode/caps/diagnostics, RM potential, thresholds, epsilon and native response-search settings are preserved. Half-budget artifacts live under half-budget-screen/ within their campaign; every worker writes a budget-stage.json alongside the frozen worker's original metadata, result and checkpoint. Development probes are explicitly marked smoke and cannot count as scientific results or unlock a pilot gate.

The controller retains the existing maxima (12/10/10) and global expanded-job ceilings (177/177/185), including held tasks. It shares the existing parent admission lock, preserves each original campaign's controller lock and all submitted job IDs, and uses the existing shared interpreter explicitly. New case/mode seed-zero or seed-100 gates check finite execution only; halfway performance does not remove planned seeds. A short disappearance grace avoids immediately treating delayed result visibility as job failure.

Reporting, promotion and timing¶

Complete matched five-seed screens are reviewed as comparisons. Promising or inconclusive comparisons may be promoted with their matched baseline and all included seeds under the user's authorization. Weaker halfway outcomes can justify deferring full training and investigating diagnostics; they do not establish full-budget failure or an implementation defect. Report the budget actually observed and retain all outcomes. Do not select only the best seeds or combine half/full rows without explicit budget labels.

Current checkpoints retain trained model/optimizer states, but do not retain all environment, random-generator and IQL replay-buffer state needed for an exact uninterrupted continuation. A fresh full-budget confirmation is therefore the conservative current option, with its cost disclosed; a faithful resumption implementation must separately preserve and verify those states. Fresh halfway training followed by fresh full training costs 150 percent of the original training for a promoted profile. No automatic extrapolation from halfway probability estimates is an equilibrium certificate.

Retain the original independent evaluation budgets. Certificates stored by the frozen workers are provisional for their declared family; recompute simultaneous coverage across all examined stages and archived candidates before making paper claims. Gamma screening additionally reserves twice the original per-campaign family size for its provisional payoff-ceiling bounds.

The planning estimate is 24--36 hours for first-stage screening results, roughly 24--48 hours for the core replacements/controls, and 2--4 days for the initial campaign set once workers are placed. Full-budget promotions are outside that first-stage estimate. Commons and strong response searches retain their budgets and may still determine overall finish time; allow about another day for complete auditing and paper integration. These are provisional ranges, not a promised deadline.

Validation¶

The required changed-path selection passed 613 routine tests; the nine focused staging tests also passed after the final controller metadata changes. Ruff and the strict documentation build passed. Three real local development workers exercised frozen c84a075 IQL-CER, c2dbf27 PR2-IPPO-CER and 612145f Territory critic-only CER, each with five rollout batches and two independent evaluation episodes. All saved finite models and the original ten-update development schedule, with only five updates executed. Both gamma probes checked the actual saved checkpoint configuration, as well as the worker's protocol metadata. Probe outputs are under /tmp/pltlf-half-budget-probes-schedules/ and /tmp/pltlf-half-budget-probes/; these are engineering evidence only.

Live deployment and additional headroom¶

The detached adapter checkout is /vol/gpudata/oja24-pltlf/worktrees/control-screen-2c22b5a, pinned to 2c22b5a45f3c9420a9249bccbd1324250a03d6f1. All 16/11/31 previously submitted Gamma/PR2/CER scientific job IDs survived the initial transition. First half PR2 jobs 61960--61962 were running at 00:11 UTC, with live trainer metadata TOTAL_TIMESTEPS=500224, SCHEDULE_TOTAL_UPDATES=3907 and the intended paired gamma values. Job 61963 was queued; these startup observations are not performance evidence.

A subsequent read-only check found background allocations down to 124 CPUs after an unrelated 48-CPU job ended. The user-requested headroom policy now allows Gamma 16, PR2 Gamma 12, CER 14, Commons 14 and archived responses 15; Coins four and Pursuit four remain unchanged. This is 79 physical-core workers, normally 158 Slurm CPUs. Reserving four CPUs for the additional background job observed pending gives a maximum snapshot allocation of 286, below 288. Global submission ceilings remain 177/181/185. No external job, hold, gate or source was altered. These admission-only changes leave the staged scientific split and all submitted workers intact.

The CER observer now honors each staged worker's output_root; its prior source/state/PID are preserved, with replacement launcher 2461653 and Python PID 2461656. The Gamma observer's notices now explicitly refer to staged full/half budgets; its replacement launcher is 2465409. A new read-only comparison watcher, launcher 2464685, records completed matched groups in the responsible agent's shared mailbox and copies artifacts before notices. Existing event deduplication is retained and no chat prompts are injected. Mailboxes still require agents to read them during active work.

Read staged-plan.json to locate results and distinguish budgets; original static full-budget summarizers do not know the new half-output paths. Full confirmation remains a separately labeled, authorized follow-up after matched screening review. The first-stage planning ranges above exclude that additional work; stronger response workers still lack complete end-to-end timings.

Final paper data policy and first full-confirmation wave¶

On 13 October Omar authorized retaining gamma one where it helps and gamma 0.99 where it performs better as final paper data. Full-map Coins is the paper default; Capture 2 retains its original infeasible threshold; mixed CER benefits are reported with corrected data and the implementation limitations.

For the task/learner pairs with matched discount controls, select one discount across all five seeds using the primary batch's best-agent five-seed mean, max_i mean_seed(p_hat_i), matching the paper's individual-satisfaction metric. Ties retain gamma one. Final estimates use the disjoint confirmatory batch and bounds recomputed across all examined seed profiles, discounts, archives and budget stages. A gain for one role is not a joint improvement. Invalid profiles remain failures; they do not receive arbitrary payoffs or silently disappear from a five-seed mean. Halfway rankings are exploratory. No discount comparison uses local-view Coins as a control for full-view Coins.

The first full-confirmation wave is admitted at /vol/gpudata/oja24-pltlf/exports/optimisations/gamma-full-confirmation-20261013. Its immutable manifest has sixty fresh full-budget profiles: Capture 2 IQL with/without CER and Classic 3 IPPO/IQL with/without CER, both discounts and all five seeds. All six input screens were complete, finite, source-pinned and schedule-checked before admission. The review is archived in results/gamma-full-confirmation-wave1-review-2026-10-13.json, including input checkpoint hashes and separate primary scores for each role/discount. Full rankings, rather than these input rankings, determine final selections.

Workers call the original frozen gamma source c84a075 directly at its full notebook budget. The controller reuses adapter 2c22b5a; no scientific worker source, native CER option, architecture, evaluation budget or threshold was changed. Original controllers and submitted workers remain intact. Frozen worker certificates are provisional for their declared family; the larger final paper selection family is recomputed after collection.

The new controller is PID 3858330, max four, cap 177. Original Gamma remains max sixteen. With the other recorded campaign maxima, this permits 83 physical-core paper workers, normally 166 Slurm CPUs. At admission the other user jobs used 104 CPUs; four additional CPUs were reserved conservatively for the pending background task, totaling 274 below the unchanged 288 CPU limit. The global 177/181/185 expanded-job caps and shared admission lock remain unchanged. No other jobs, holds, sources or caps were modified. Jobs 62234--62237 were observed running; startup is execution evidence only.

The comparison observer now includes the confirmation root. Its old source and deduplication state were preserved before restart, with replacement launcher 2863472. Its first four-root snapshot was healthy, errors zero. Completion events are file-mailbox records and artifact copies, never chat prompt injection. Capture 3 remains under complete-screen review for a later wave within the same four-worker confirmation allowance; it is not included in this sixty-profile admitted manifest.

Audited discount-selection collector¶

scripts/summarize_discount_selection.py collects any explicitly listed original and confirmation campaigns from their staged manifests. For example, after copying remote artifacts into the corresponding local export folders:

uv run --no-sync python scripts/summarize_discount_selection.py \
  --campaign exports/optimisations/gamma-one-20261012 \
  --campaign exports/optimisations/pr2-gamma-one-20261012 \
  --campaign exports/optimisations/gamma-full-confirmation-20261013 \
  --output .cache/discount-selection-current.json

The collector checks scientific identities, frozen source revisions, full or half training budgets, preserved half schedules, raw satisfaction counts, disjoint 1000-episode evaluation blocks and checkpoint checksums. Matching full-budget pairs must also share the source-file hashes, trainer options apart from gamma/seed, thresholds and agent sets. Duplicate full-budget seed outcomes require explicit cohort review rather than automatic selection. No checkpoint unpickling is needed for this metadata and checksum audit.

Selection requires both discounts and every seed 0--4 to be complete and finite at the full budget. Integer primary success counts select one discount for the whole five-seed task/learner group; exact ties retain gamma one. The output exposes the selected primary_selection_score separately from its confirmatory role means, so later exports need not choose settings using the confirmatory estimates. Partial, failed and half-budget profiles remain in the inventory and cannot fill a final matched comparison. The output includes the collector checksum and generation time.

Listed-campaign family sizes count all listed profiles, including budget stages and failed attempts. They are not the complete paper-wide selection family: archived candidates and other experiments must also be inventoried before recomputing final bounds. Frozen-worker certificates and approximate response-search values do not become equilibrium certificates through this collector. Twelve focused tests cover seed completeness, selection and exact ties, independent reporting, source/settings consistency, duplicate cohorts, invalid counts, local artifact mapping and checkpoint integrity. The required broader changed checks passed with 625 tests; strict documentation checks also apply to this note.

The read-only comparison observer now invokes a private copy pinned to code revision 3422091 after discount artifact copies and at startup. Collector failures are recorded for review; they do not produce a new selection or prevent scientific completion/failure notices. The prior observer source and event state are preserved, with replacement launcher 2932018. Its first snapshot at 12:02 UTC was healthy and recorded all 34 discount task/learner groups as awaiting complete full-budget pairs. No scientific worker source, job, submission ceiling or concurrency allowance changed.

Selective response retirement and confirmation headroom¶

At the user's instruction to run stronger response searches only when useful, W1 retired ten obsolete local-view Coins workers and preserved their 193 candidate checkpoints, protocols and logs. Its original fifteen-profile plan remains archived with explicit truncated records, not invented failures or certificates. Five Territory response workers remain under a read-only no-admission tracker; their allowance is five, with cap 181 unchanged.

W1 transferred the ten freed physical-worker slots to full gamma confirmation. After confirming the retired jobs absent, checking the own controller identity and acquiring the shared parent lock, we replaced launcher 3858330 with 3949667. Confirmation is now max fourteen, cap 177. The original scientific staged-plan.json bytes remain unchanged; separate admission-plan-max14.json has identical trials and changes the admission maximum from four to fourteen. Prior ledger/deployment/manifest bytes are preserved under allowance-transfer-20261013T155952Z/, and all ten previously submitted scientific job IDs survive. Worker source c84a075 and adapter 2c22b5a remain frozen. The first post-transfer snapshot had six complete profiles and fourteen running, which is execution evidence rather than a matched result.

The coordinated maximum remains 83 physical workers, normally 166 Slurm CPUs: original Gamma sixteen, confirmation fourteen, Coins four, PR2 twelve, CER fourteen, Commons fourteen, archived responses five and Pursuit four. Global 177/181/185 ceilings are unchanged. No diagnostic allowance or scientific settings changed. From measured Capture IQL and Classic 3 IPPO full runtimes, the remaining confirmation planning range falls from roughly 30--40 hours to 10--16 hours, conditional on admission and runtime variation. PR2 repair and the larger Commons controls still affect the overall completion date.