burtenshaw/plan-my-day-v3-source

Plan My Day v3: audited environment and adaptive research loop

Objective and boundary

Produce a reproducible, empirically validated training method for choosing among four offered synthetic calendar plans, including adaptation across normal, deadline, and travel regimes. Preferences are explicitly provided. This is not free-form calendar generation, private preference inference, or demonstrated real-user benefit. v1/v2 evidence and the existing live demo remain unchanged.

The user authorized continued revision and further runs. Initial new-compute cap: $10, with recorded actual running time plus full-timeout reservations for active or ambiguous submissions. No background duplicate runs, secret logging, personal data collection, or social posting. All new source, data and artifacts go to separate burtenshaw/plan-my-day-v3-* repositories and Trackio project plan-my-day-v3.

Findings that change the experiment

The independent CPU audit checks hard constraints without calling the production grader, independently solves feasibility, enumerates reachable repair states, and tests oracle, random and shortcut baselines across nine regime/profile strata. A hard-constraint/completion heuristic achieves full completion on the initial 108-day audit while still making preference mistakes. Therefore completion alone is not a useful learned advantage over a rules baseline. Final repair can also erase bad intermediate decisions, so trajectory safety, original correction flags and regret must remain visible.

The structured v1/v2 observation includes generator seed and seed-derived IDs/date. The existing text policy omits these, but an arbitrary structured agent could reconstruct the label shuffle. The v3 public boundary must redact those values; broker-side original snapshots retain reproducibility. Public action semantics and policy prompt text must stay identical for the same underlying state.

Round 1: controlled optimizer diagnosis

Run exact native four-action reverse KL with the live teacher from the identical v2 pre-travel SFT adapter and pinned training correction ledger. Check 64 and 192 updates using one optimizer. Keep four generated actions per prompt to isolate the loss estimator, although exact one-token loss should be independent of which action is sampled. Native TRL loop/loss remain authoritative; no custom trainer loss or online update loop. Re-evaluate the initial checkpoint and every published checkpoint on v2 development days only, never v2 test days. Compare approximately time-matched native SFT at 110/330 updates, calibrated from v2 measured timings. Report actual timing and unequal prompt exposure, not nominal FLOPs equivalence.

Use the same fixed training-only teacher probe before and after updates. Record active completion/teacher masks, finite gradients, exact checkpoint revisions, optimizer/scheduler continuity, dataset hashes and generated action counts. Native PEFT base disables the initial adapter; native EMA does not initially copy it. Neither may be mislabeled as the frozen warm-start teacher.

Round 1 is deliberately the old raw-policy task for attribution. Its results do not confirm the revised safe preference-ranking system described below. An additional paired guarded-policy evaluation is registered before these runs to measure whether the new preference-ranking objective has useful headroom. It does not alter the training data, loss, or raw-policy comparison.

Subsequent rounds: useful and safe preference ranking

Use an independently checked hard-constraint/non-deferral filter shared by the symbolic heuristic, frozen model and trained model. It uses no soft-preference oracle. The model ranks the remaining feasible candidates. If no acceptable candidate exists, fail explicitly; never silently substitute an oracle action. This deterministic guard supplies safety/completion, not learned weights.

Keep unfiltered model results separately. The guard must not justify claims that raw training improved safety, that raw updates never regress, or that a deployed rollback policy proves monotonic learning.

After round 1, register each revised recipe with a new trial ID and immutable source/data revisions before launching. Candidate changes must follow observed development failures (e.g. action-estimator variance, teacher drift, state coverage, label sensitivity, or retention). Publish the failures alongside any winner. SFT is a legitimate method outcome; there is no requirement to manufacture an SDPO victory.

Prospective validation contract

Before evaluating a new confirmation cohort:

1. Freeze the independently audited environment, filter, data, training recipe, checkpoint schedule, fixed seeds 17/29/43, and all metrics.
2. Use 48 fresh development days per regime (144 total). Use 128 fresh confirmation days per regime (384 total), structurally disjoint from all v1/v2 splits and all v3 training/development/audit examples. No confirmation results may guide its own recipe. A failed cohort becomes exploratory; a later confirmation must be newly generated and registered.
3. Evaluate identical days for the symbolic safe heuristic, frozen model with the same safety filter and phase-limited memory, and the trained system. Publish every raw checkpoint and each seed, not a test-selected best checkpoint.
4. Primary learned benefit: final original correction flags/day improves by at least 0.20 against the frozen model with the same filter, with a paired 95% day-cluster bootstrap interval excluding zero and positive improvement in every optimizer seed. Also require a positive paired preference-quality gain against the deterministic safety-only heuristic. Report cost regret and zero-correction trajectories as supporting outcomes, not substitute winners.
5. The guarded system must have zero independently measured hard violations and no deferred tasks on confirmation. This is a finite synthetic benchmark result, not a guarantee of real-calendar safety. Report all raw unsafe decisions.
6. Retention: between successive guarded checkpoints, original correction flags may increase by at most 0.10/day overall and 0.20/day in each previously seen regime, for each seed. Report paired uncertainty for these differences and distinguish a descriptive margin pass from statistical non-inferiority.
7. An SDPO-specific advantage additionally requires a positive paired improvement over an approximately training-time-matched SFT control with the same guard, memory and evaluation. No SDPO claim if only the full system beats frozen.
8. Bootstrap the shared day IDs jointly across observed seed pairs; do not pretend 1,152 seed/day evaluations are independent days. All seeds share one pretrained model and warm-start adapter, so uncertainty is conditional on those assets.

The new preference-quality target follows the independent environment audit, before new GPU results. It does not retroactively change the negative v2 result or round 1's raw-policy diagnostic. Numerical tolerances are prospective choices, not grounds to relabel an old failed checkpoint as successful.

Round 2 registration (after completed round 1, before fresh development inference)

Round 1 results revision 9861de560aab351b11708b0e3fb63f88eb4e547b shows exact SDPO did not rescue the live-teacher recipe: guarded flags rose from 0.833 to 1.181/day and the fixed training-only feedback-conditioned teacher probe fell from 96.875% to 58.333% accuracy. SFT at 330 updates reduced guarded flags to 0.542/day (paired improvement 0.292, 95% day-bootstrap interval 0.069–0.528). The 110-update SFT checkpoint had regressed. Actual optimization times were 181.1 seconds for exact SDPO and 176.6 seconds for SFT. These are one-seed development results, not confirmation and not evidence that all updates improve.

The next controlled comparison uses SFT at 330 updates per regime, learning rate 3e-5, with historical replay shares 0.50 versus 0.75. Both start from the same pinned initial adapter, use the same eligible correction ledger, and traverse all three regimes. Evaluate each phase's final checkpoint and its pre-training state; also re-evaluate the frozen initial adapter with phase-matched memory. Raw and guarded views are both retained. No checkpoint is silently rolled back. Select a recipe using fresh development preference quality AND prior-regime retention before any confirmation outcomes. End-of-phase checkpoints are the registered acceptance schedule; mid-step monotonicity is not claimed.

Both round-2 arms use the same explicitly versioned viable-target derivative of the fixed student ledger. Of 1,152 states, one accepted local-cost target blocks a remaining task despite a viable alternative; its training target and active memory correction are repaired. The original accepted choice, cost, feedback, proposal, and correction text remain recoverable. Another 106 states already have unsafe/incomplete raw-student prefixes and no viable offered action; their labels remain explicitly marked recovery supervision, not guarded on-policy collection. The other 1,045 original accepted labels are viable. This data repair is shared across both replay arms, never applied retroactively to round 1.

The larger structural audit found a missing offered continuation on development day dev-travel-12320014. The opt-in independent_witness candidate generator adds a public hard-feasible placement only when no offered viable option exists; it does not use preferences, seeds, or an unrevealed interruption. No day is dropped. Covered-mode reward ranks preferences within viable options, because the legacy current-step cost can prefer an action that blocks a later task. Original all-action correction flags remain the conservative registered primary metric; aligned viable-only feedback is separately named. Thus a necessary safe action can still incur a legacy flag, but must not receive a negative aligned reward.

Inference uses batch size 8 only when the actual checkpoint passes a fixed training-only serial/batch check: all raw and guarded selected actions and prompt hashes/lengths agree; maximum absolute A–D probability error is at most 0.03. Failure explicitly falls back to serial evaluation and is recorded. Qualification is per checkpoint; it does not assert bitwise GPU equivalence or prove all states.

Development decision and prospective confirmation

Both round-2 jobs completed on the fresh 144-day development cohort; their final persisted results are at revision 5b3c891a0618dd19658be01ab9ce69c895e44ee3. The 75% replay recipe is selected for confirmation, not declared confirmed:

Guarded system	Normal-phase end	Deadline-phase end	Travel-phase end
SFT, 50% historical replay	0.882	0.521	0.556
SFT, 75% historical replay	0.951	0.646	0.368

Values are original simulated correction flags/day on the same development days. The final phase-matched frozen comparator has 1.028 flags/day; the rules baseline has 1.785. Selected-recipe benefit versus frozen is 0.660/day, paired 95% interval [0.507, 0.819]. Both recipes pass development quality, independent guarded safety, and descriptive retention gates. The 75% recipe also passes every reported retention non-inferiority interval; 50% does not. These exploratory intervals are conditional on one optimizer seed and are not multiple-comparison adjusted.

This is a whole-procedure selection, not causal proof that replay share explains the difference. Even the identical phase-1 datasets produced different adapter tensors across the two GPU runs. Also, the final 75% checkpoint passed the fixed TRAIN-only batch-8 qualification while its earlier checkpoints and all 50% checkpoints fell back to serial inference. The numerical inference mode therefore changes in that contrast; held-out serial/batch equivalence is unmeasured. The same prospective qualification/fallback procedure is retained in confirmation. No checkpoint is selected using confirmation outcomes.

Confirmation restarts the selected recipe from the original pinned warm-start adapter independently for seeds 17, 29, and 43. Each trains 330 native SFT updates in each of three phases, with all pre/post raw and guarded evaluations plus phase-matched frozen baselines on the registered 384-day cohort. No hyperparameter, metric, safety rule, checkpoint schedule, or retention margin is relaxed. An exact published development-selection report and full registration must pass validation before any confirmation inference. Canonical analysis source is frozen with the new source publication before outcomes are read.

The operational Job timeout increases from 60 to 90 minutes because evaluation on 384 days, including explicit serial fallback, is substantially larger than development evaluation. Hardware remains A10G-small and the cumulative new-v3 compute cap remains $10. The four completed pilot/development Jobs used 5,673 GPU-running seconds, approximately $1.576 at the recorded hourly rate; queue time is reported separately. Timeout/cap/price are operational metadata, but the full execution configuration hash still binds them. They do not alter the scientific recipe fingerprint.

Required checks before a final claim

• Independent hard grading and exact feasibility; every reachable included repair state has at least one full feasible action.
• No seed/ID/shuffle-key leak in public structured observations, rewards, or metadata.
• Same-state action-label permutation checks, accepting tied optimal actions.
• Session isolation, deterministic replay, valid reset/step/terminal behavior via the OpenEnv interface; malformed actions cannot earn a successful episode.
• Train-only memory/feedback, disjoint data hashes, no silent prompt truncation.
• Native trainer smoke plus nonzero updates and pre/post diagnostics on real runs.
• Published checkpoints reloaded at exact revisions for evaluation; all artifacts persisted and Trackio logs read back from the hosted project.
• Full held-out report with raw and guarded policies, all seeds/stages, uncertainty, complete cost accounting, and explicit remaining limitations.