burtenshaw/plan-my-day-results

Can a planner need fewer corrections—not just remember more?

A calendar is a useful place to make continual learning tangible. You want time to write, a chance to exercise, and a way to fit both around meetings. Then another meeting appears. A helpful assistant should recover without quietly dropping what mattered.

That is the idea behind Plan My Day: a small, inspectable world built with OpenEnv, a usable calendar demo, and an experiment asking whether training changes planning behavior beyond what a model can do with remembered corrections alone.

Try Plan My Day. Load a sample or enter a day's commitments and tasks, adjust a proposed plan, explain a correction, and export accepted task blocks as a calendar file. Choose the actual trained model for live proposals, or the clearly labeled rules-based reference planner for an instant response. A separate replay view compares recorded model decisions before and after training.

The important boundary: the demo does not fine-tune model weights on your calendar or feedback exports. Lasting corrections can guide model proposals within the current session. Turning consented feedback into a tested model update is the next experiment, not a feature being claimed today.

The experiment behind the calendar

The student is LiquidAI's LFM2.5-1.2B-Instruct. Each synthetic day has three tasks, fixed commitments, explicit preferences, and a later disruption. At each of four decisions, the model chooses one of four prepared actions; OpenEnv applies the choice to the actual calendar state. This makes mistakes visible and replayable without pretending the model can generate arbitrary valid schedules.

Training moves through normal, deadline, and travel scenarios: 48 training days per phase, with a small amount of rehearsal. The same 72 unseen test days are evaluated after each phase, across three training seeds. They are 72 distinct days reused across seeds, not 216 independent days. The primary test calendars are all capable of fitting every task before and after the disruption.

We compared a frozen model with explicit preferences, the same model with correction memory, sequential supervised fine-tuning (SFT) plus that memory, and SFT followed by SDPO plus memory. All arms receive the explicit preferences. Memory-enabled arms have access to the same permitted training correction ledger at each stage; no test feedback is added to it. Frozen deterministic baselines are counted once.

Corrections here are synthetic reference demonstrations, not edits collected from people. The primary metric counts a flag whenever a decision is strictly worse than the best offered action under the public scoring rubric. It measures this constrained task—not human effort saved.

A promising result, with a real trade-off

At the preselected final checkpoints, SFT reduced simulated correction flags from 2.86 to 1.47 per day versus frozen + memory: about 49% fewer flags. The reduction appeared in all three training seeds.

Final policy	Simulated flags/day ↓	All tasks scheduled, no conflicts ↑	Conflict-free final plan, even if tasks deferred ↑
Frozen + explicit preferences	2.85	55.6%	79.2%
Frozen + correction memory	2.86	55.6%	79.2%
Continual SFT + memory	1.47	69.9%	76.4%
Continual SFT + SDPO + memory	1.51	55.1%	77.3%

The practical gain is not just a score: SFT scheduled every task without conflicts on more days, increasing from 55.6% to 69.9%. But the fraction of final plans with no conflicts at all fell from 79.2% to 76.4%. Those measures differ because a planner can avoid conflicts by leaving work unscheduled. Fewer correction flags did not improve every scheduling outcome.

The paired mean reduction was 1.39 flags/day, with an exploratory 95% bootstrap interval of 1.09–1.68 fewer flags, resampling both training-seed and shared-day clusters. Three training seeds are too few for a strong generalization claim; this interval is not evidence of real-user efficacy.

SDPO did not improve the primary final average over SFT, despite adding eight optimizer steps per phase beyond SFT's 48. Its complete, conflict-free schedule rate was also lower. Some secondary metrics moved the other way, so this is a result about this small configuration, not a verdict on SDPO.

Nor did improvement continue smoothly. SFT's average flags/day were approximately 1.44 after normal training, 1.28 after deadline training, and 1.47 after travel training. The reported final model is not a retrospectively selected best checkpoint. The learning curve makes that regression visible.

What the replay shows

The accompanying 13-second animation uses one fixed held-out example, selected before inspecting its outcomes. Both planners receive the same commitments, preferences, and unexpected 10:30 meeting. Both finish without conflicts. The memory-only model leaves the run unscheduled; the SFT model fits writing, exercise, and messages into the day. The simulated flag count is three versus one.

It is an illustration, not a representative success-rate estimate. The animation replays saved model actions and depicts no human corrections.

Small enough to inspect—and repeat

The three parallel pilot jobs ran on Hugging Face A10G GPUs. From first submission to final completion they took about 10.5 minutes, with $0.47 estimated GPU compute across those jobs. That estimate uses their observed running time and the recorded hardware rate; it is not an invoice or the full project cost. Preflights, the separate preference probe, external-teacher requests, and hosting are excluded.

Training and evaluation are tracked in Trackio. The results repository contains the evidence; the pinned experiment source and pinned data identify the exact training inputs. The full report links every evaluated checkpoint revision. Model adapters remain subject to the base model's LFM license; the code's license is not the weights' license.

For clarity about the teacher: SDPO uses the current student with an accepted training correction in a privileged reprompt. A separate Qwen/Qwen3.5-397B-A17B audit is not its teacher and did not supply SFT labels. Of 24 audit requests, 12 returned responses, all matching an optimal offered action; 12 timed out. Provider failures are not wrong model decisions, and this training-only audit is not held-out test evidence.

A separate eight-day, one-seed probe changed the explicit focus-time instruction while retaining the old correction memory. It ran without training. That is a limited instruction-override check—not evidence that the model learned a new preference, and not a matched causal comparison with the original pilot because the derived interruption can also change.

The next useful experiment

The missing piece is a consented feedback loop: let someone use the planner for real, distinguish a one-day exception from a lasting preference, and explicitly choose which corrections can become training data. Build an importer for those approved exports, then compare context-only memory with a small periodic weight update on future days.

Promotion should require more than fewer correction flags. Test complete, conflict-free plans, dropped tasks, respect for the latest preferences, and retention of earlier ones before offering a new checkpoint. Keep a rollback path, and measure whether people actually accept more plans with less editing.

That would move this from a synthetic curriculum toward the applied question worth answering: can a small assistant become easier to live with, without forgetting what already worked?