Skip to content

Latest commit

 

History

History
47 lines (37 loc) · 4.71 KB

ROADMAP.md

File metadata and controls

47 lines (37 loc) · 4.71 KB

Roadmap

Core

  • Replay parsing & caching — Parse .replay files into full JSON (including network data) via boxcars, with cached output in parsed_games/
  • Match overview & player stats — Human-readable summaries: game type, score, duration, forfeit detection, per-player scoreboard
  • Interactive replay selector — CLI menu to browse replay categories, select replays, and choose analysis actions

Replay Verification

  • Merkle tree replay signing — Split replay JSON into semantic sections, hash into a Merkle tree, sign the root with hybrid Ed25519 + ML-DSA-65
  • Signature verification — Load .sig sidecar files, verify both signatures, and detect which section was tampered with
  • Replay integrity chain — Extend the merkle/sidecar system to chain replays together for a verifiable tournament match history

Bot Detection

  • Input analysis — Score players based on unique steer/throttle value counts, discrete-only input detection, and platform weighting
  • Kickoff signals — Integrate kickoff pre-hold (human signal) and reaction consistency (bot signal) into the bot score
  • Steer alternation rate — Count direction changes per second to separate keyboard players (human timing, ~4 changes/sec max) from bots (rapid mechanical alternation, 15+/sec)
  • Hold duration variance — Measure how long each steer/throttle value is held; humans have high variance, bots tend toward uniform durations
  • Input entropy scoring — Replace the unique-value-count scoring ladder with Shannon entropy over the steer/throttle histogram; controller players produce a smooth bell curve, bots produce narrow spikes
  • Keyboard-aware scoring path — When a player is identified as discrete-only (keyboard), shift scoring to timing-based signals instead of penalizing low unique counts
  • Multi-input synchrony — Track how often steer + throttle + boost + jump change on the exact same frame; bots show unnaturally high simultaneous input rates
  • Dodge timing consistency — Measure variance of dodge-after-jump delay across the match; frame-perfect double jumps every time = bot signal
  • Post-impact recovery timing — After bumps/demos (velocity spikes), measure time-to-first-input; bots recover in fixed frame counts, humans vary
  • Boost tap duration variance — Measure the length of each boost-on period; low variance across many taps = suspicious
  • Ball prediction accuracy — Compare player heading to the direction of the ball's extrapolated future position; consistently low intercept-angle error = bot signal
  • Rotation period regularity — Measure stddev of time between offensive/defensive transitions; mechanically regular cycling = suspicious
  • Steer-to-ball correlation — Compute ideal steer to face the ball each frame, compare to actual; bots track with low consistent error, humans overshoot/undershoot noisily
  • Aggregate suspicion profile — Combine bot detection + kickoff + boost + rotation into a single per-player composite score with confidence interval

Match Analysis

  • Kickoff analysis — Detect kickoff windows, measure per-player reaction latency, pre-hold detection, steer/throttle sequence variability across kickoffs
  • Boost analysis — Track average boost level, time at zero/full, boost collected/consumed, big and small pad pickup counts
  • Rotation & positioning analysis — Ball chasing %, offensive/defensive split, double commit detection, back-post rotation rate, per-minute breakdowns
  • Goal sequence analysis — Analyze network frames before each goal: ball touches, pass sequences, time-to-goal from possession change
  • Demolition & bump tracking — Parse demo/bump events, track counts per player, and correlate with bot-like targeting patterns
  • Speed & supersonic tracking — Track car speed over time, % at supersonic, acceleration patterns; bots may show unnaturally consistent speed management

Tooling & Infrastructure

  • Replay comparison / diff — Compare the same player across multiple replays to detect consistent bot fingerprints
  • Training data export — Export labeled feature vectors (human vs bot) to CSV/JSON for ML classifier training; one row per player-per-match
  • Batch analysis mode — Analyze all replays in a directory, output a summary CSV; useful for scanning tournament replays
  • Player identity tracking — Track players across replays via OnlineID/UniqueNetId; build a history profile with rolling bot score and play style fingerprint
  • Web API / server mode — Expose analysis endpoints via HTTP (e.g. axum) for integration with Discord bots or web frontends