Build 0.5.1 · report_validation session · started 2026-01-01T00:00:00Z
Your practice memory
Synthetic browser validation; not a human play assessment.
This anonymous profile persists across game launches. Only gameplay is saved locally; nothing is uploaded. Detailed traces expire separately. These are observational estimates, not a diagnosis or proof that advice caused a change.
You · Coach · timing · 0.25x
Eligible returns: 8 / 12 (67%; approximate 95% interval 39–86%)
When a contact window was observed: 8 / 12 (67%; approximate 95% interval 39–86%). No observed window: 0. Excluded exchanges: 0.
Gather more comparable observations before assigning a coaching priority.
Rules: learning-rules-1 · Mint policy: fixed-feeds-1. Cohorts with different speeds, tasks or policies are not combined.
Drills and follow-up
Choose Practice & Learning in the game menu. Timing repeats a central feed; Position alternates left/right; Fixed Benchmark repeats the same twelve-ball sequence. Return contact is the outcome, not whether the outgoing shot wins a point. Initial positioning resets for each feed; these drills do not yet measure recovery between shots.
Mint's learning record
Coach targets a productive 70% human return rate; Competitive targets harder returns. Both start with balanced left/right/lob choices. Speed, reach and contact rules stay fixed. A policy is frozen during a match. Only balanced, eligible human matches train candidates; serves, forced smashes, drills, focus interruptions and game failures are excluded. Selection and evaluation use separate sessions. Mint still reads exact game state; it does not yet have human-like perception.
No promoted policy yet. Mint needs at least 24 eligible outcomes per action in each partition and three sessions per partition, followed by a conservative independent evaluation. Sparse evidence keeps it balanced.
After promotion, every fourth match is labeled CALIBRATION and uses balanced choices for fresh evidence. The latest 24 balanced sessions are evaluated at match boundaries; sufficiently sampled contradictory evidence can restore the baseline automatically. Restore the balanced baseline yourself from the practice menu if a change feels worse, and re-enable adaptation explicitly when ready. Successful evaluation is a screening result—not proof of more match wins or lasting human improvement. The approximate intervals assume independent observations; repeated shots from one session are correlated, so uncertainty may be greater.
Memory: a local SQLite database, schema 1. Session summaries and aggregates persist; the report displays the latest 60 sessions and 40 drill records. Close the game before backing up its Learning folder. To start a fresh profile, preserve a backup and move that folder aside; no reset deletes your records automatically.
Validation session: these synthetic or automated scenarios do not assess a human player's performance.
This report uses recorded game decisions. It separates possible player improvements from game failures. Suggestions are hypotheses to test; they are not proof of skill or fault.
You
67% return completion
8 returned · 4 missed · 0 outgoing errors
0 points won · 0 lost · 0 first-serve faults
0 attempts. Completion excludes incoming balls that ended as the opponent's shot error and interrupted exchanges.
No diagnosed losses yet. Keep playing to collect evidence.
Mint
83% return completion
10 returned · 2 missed · 0 outgoing errors
0 points won · 0 lost · 0 first-serve faults
0 attempts. Completion excludes incoming balls that ended as the opponent's shot error and interrupted exchanges.
No diagnosed losses yet. Keep playing to collect evidence.
Missed-return replays
The most recent eight misses show the incoming flight, sampled every 0.1 game seconds, up to 240 samples plus the outcome. Very long flights are capped and marked in the raw trace. Successful and missed incoming flights are both saved in windows.jsonl. This is a sampled court trace, not a video or a deterministic physics replay. Orange: You. Mint: opponent. Yellow: ball. The height is shown separately.
Evidence behind the advice
The latest 20 completed or interrupted exchanges. Window duration is sampled from the game's eligibility checks, not a measurement of human reaction time. Game seconds slow down in practice; real seconds show the time available to you.
| Ball | Receiver | Attempts | Speed | Window (game s / real s) | Closest (m) | Outcome / cause |
|---|
Recording quality
0 focus losses · 0 of 0 frames exceeded 50 ms. These observations can affect interpretation but do not alone establish why a shot was missed.
Local files: events.jsonl contains timestamped decisions; windows.jsonl contains completed replay windows; summary.json contains both players' totals and recent exchanges. Match restarts remain in the same session. Recording stops at the per-file size limit. Older recorder-owned sessions are retained within a 20-session / approximately 100 MiB budget, checked on launch.