GameMind roadmap — honest capability tracker

What ships in v0.1.6 today. What's actively being built. What's on the public roadmap. We'd rather you read this than be surprised by something we don't do yet.

Honest about the beta GameMind is in private beta. The "live coaching while you play" pitch you may have seen earlier is the destination, not the shipping state. Right now the app is a powerful local AI chat coach with per-game knowledge packs, voice, a wired 3-head safety cluster, and a vision lane that's training per-title image adapters (2 PASS-verified, 3 retrained-but-unverified, 6 in queue). See "Shipping today" for what you actually get. Real-time ASSIST is on the roadmap.

Shipping today — v0.1.6

Everything in this column is in the desktop installer right now. CPU-only, local-first, no cloud telemetry.

💬

Local AI chat coach

Type any question about your game ("how do I beat Malenia?" / "what's the best build for a Bard?"). Answered locally by a compact coaching model fine-tuned for in-game guidance.

📚

Per-title knowledge packs

10 live: Warhammer 40K, CS2, LoL, Dota 2, Skyrim SE, Elden Ring, Path of Exile, Factorio, FFXIV, Fortnite + OSRS. Each ships FAISS + BM25 retrieval over a curated wiki + patch-notes corpus, signed manifest.

🛡️

Wired 3-head safety cluster

cluster_pipeline.py is live end-to-end: every response routes triage → policy gate → coach gate before the heavy lane fires. 3 cluster heads (q-coach, q-policygate, q-gametriage — 53.5M params each, on 127.0.0.1:8788, no network). q-patchcite and q-wikicite were retired today — they were dead weight; the heavy lane is RAG-grounded by chunks directly.

👁️

Vision lane — wired, 2 titles PASS-verified

Frozen SigLIP-base encoder + per-title 11 MB MLP projector → 64 visual tokens prepended to the heavy lane via inputs_embeds. Wired end-to-end through vision_lane.py + heavy_lane.py + Windows.Graphics.Capture. PASS-verified on held-out scene eval: League of Legends (96.5%), Skyrim SE (99.0%). Retrained today but pre-verification: Warhammer 40K, CS2, Elden Ring. Dota 2 retraining at 24.6% (target ≥75% — coming v0.1.7). 6 more titles in active training.

🎙️

Voice in/out (opt-in)

Push-to-talk speech via faster-whisper. Local TTS replies via Piper. Mic is never always-on; both run on your CPU.

🎯

Per-title RAG grounding

Every answer is anchored to your title's curated knowledge pack via FAISS + BM25 hybrid retrieval (unified through rag_retriever.py). The heavy lane sees the chunks directly — no separate cite-check head, no "(unverified)" warnings cluttering the answer. Off-topic / out-of-corpus questions are gracefully refused.

🔒

Local-sovereign

Your prompts, transcripts, voice clips never leave the box. The Fly backend only sees license activation + adapter download tokens + version-check pings. Safety audit →

Recently shipped — 2026-06-08

Landed today.

✓

Vision lane wired end-to-endvision_lane.py (frozen SigLIP encoder + per-title projector cache) + chat.py + heavy_lane.py + mounted_adapter.py + Windows.Graphics.Capture. 76/76 pytest. 6 trained projectors on disk; League of Legends + Skyrim SE PASS-verified at 96.5% and 99% scene accuracy.

✓

3-head cluster pipeline livecluster_pipeline.py wires q-coach + q-policygate + q-gametriage into every grounded chat turn. Triage routes correctly post-fix; policy gate only blocks action-class queries (not lore/build).

✓

Citation cluster heads removedq-patchcite + q-wikicite stripped end-to-end. Saved 1.07 GB bundle + 200-500 ms per chat turn + ~150 LOC complexity. The RAG retriever still grounds the prompt with wiki chunks — the user just isn't shown a "Sources" panel they wouldn't read mid-game.

✓

Referral program + Refer & earn UIEvery user gets a unique GM-XXXXXXXX code. When a referee signs up and makes a paid purchase, the referrer gets a free-choice-adapter promo code via email. Per-referee idempotency; backed by 7 pytest scenarios.

✓

$9.99/mo Plus tier + 2 free adapter credits per monthSubscribe to grab any 2 adapters per month at $0 (Stripe price live; gamemind_user table tracks credits + period reset; redeem endpoint enforces).

In active development

Training or coding right now. Expected within v0.1.7–v0.2.0.

●

Vision projectors — 6 more titlesBG3, OSRS, Path of Exile, Factorio, Fortnite, FFXIV. Frames on disk (8K+ per title from YouTube scrape). Tier 0 query-metadata labeling + class-balanced sampler will train them through the same gate the first 5 cleared (≥75% scene accuracy). Dota 2 needs a corpus diversity pass before its retrain (queries pulled too much laning footage).

●

Per-title sovereign retrains — Round 2 (10 DRAFT)BG3 + World of Warcraft already cleared the relaxed lore gate today. 10 more (Diablo IV, Minecraft, Cyberpunk, EVE, Overwatch, Stardew, Terraria, Warzone, Apex, Valorant, Fortnite) get lore-augmented Q/A retrains.

●

HUD readout — structured state from framesPer-frame OCR + a small HUD-extract head reads HP / mana / cooldown / scoreboard / inventory and surfaces it to the coach prompt. Today the coach gets your typed/spoken question + RAG; once HUD readout lands the coach sees your live situation too.

●

Single-exe installer (Nuitka + 7-Zip SFX wrap)v0.1.5 standalone build is done (3.1 GB, MSVC-encrypted Nuitka). 7-Zip SFX wrap into one self-extracting GameMind-Setup-0.1.5.exe is in final smoke after a config-encoding fix earlier today.

On the roadmap — designed, not yet built

Public commitments. Order is rough priority, not a date guarantee.

○

OBSERVE mode — real-time coaching overlayOnce vision + HUD readout land, the coach narrates situations as they happen ("boss switches phase at 60% — your build won't out-DPS phase 2") in an in-game overlay. No game interaction.

○

ASSIST mode — single-shot user-confirmed actions (green tier only)"Sort inventory by weight?" → you click Confirm → one OS-level SendInput keystroke goes out. Atomic, gated, single-shot. Only available on single-player green-tier titles (BG3, Elden Ring, Skyrim, sandbox Minecraft). Never on ranked competitive titles, by design.

○

Game log / GSI integrationFor titles that expose them: Dota 2 GSI, LoL Live Client API, PoE / FFXIV / WoW combat logs, chat-log tails. Cleaner signal than vision for games that offer it.

○

Per-title q-action head (the 6th cluster head)Trained on per-title action allowlists. Proposes one concrete action class given the current frame state. Won't fire without the policy gate's allow plus your confirmation.

○

Sparse heavy lane schedulerRun the heavy lane only at "stable decision points" (loading screen, menu, post-combat) instead of every frame. Keeps latency tight + battery + thermals reasonable on a laptop.

○

End-of-session local memorySparse local summary: what you played, where you got stuck, what to drill next. Lives in %LOCALAPPDATA%\GameMind\sessions\, user-deletable.

Per-title risk tiers (already enforced at the policy level)

Game Adapters ship with a hard-coded risk profile. The user does not get a toggle that overrides it. Today this controls whether the coach gives action suggestions in chat; once ASSIST mode lands it controls whether the app can send keystrokes.

Tier	Examples	Chat coach (today)	ASSIST keystrokes (roadmap)
Green	Single-player RPGs (BG3, Elden Ring, Skyrim), sandbox / single-player Minecraft, Stardew, Terraria	Full — coaching, knowledge, action ideas	Opt-in — inventory sort, dialogue confirm, journal nav
Yellow	Co-op shooters, friendly PvE, modded contexts	Full chat; action ideas labeled "manual only"	Default off. Coaching only.
Red	Ranked competitive (CS2, Valorant, OW2), MMOs with anti-bot rules (WoW, OSRS, FFXIV ranked)	Strategy + knowledge only; the coach won't suggest in-match actions	Force-disabled. No keystrokes are ever sent. The app is structurally incapable.

The rule The red tier is non-negotiable. There is no setting that re-enables in-match actions on a ranked competitive title. The app is incapable of sending input under those profiles, by design — when ASSIST mode ships, the policy gate refuses to fire on red-tier titles before the SendInput call is even reached.

What we don't do — by construction, not policy

These rules hold today AND on the roadmap. Anti-cheat-safe is a structural property of the architecture, not a checkbox.

✗

No DLL injectionThe app never injects code into the game process. Once vision lands, capture will use the standard Windows.Graphics.Capture API — same one screen recorders use.

✗

No memory reading or writingWe never read the game's RAM. The state we know is what we infer from screen pixels and (where the publisher exposes them) public game-state APIs.

✗

No network interception or packet manipulationWe don't intercept, modify, or replay game traffic. Ever.

✗

No kernel driverGameMind is a normal user-mode Windows app. Nothing runs as a driver.

✗

No client patching or game file modificationWe do not touch game executables, configs, save files, or asset bundles.

✗

No multi-client input broadcasting, macros, or farming loopsWhen ASSIST ships: actions are atomic, user-confirmed, rate-limited, cooldown-gated. There is no "run this 100 times" surface, ever.

✗

No CAPTCHA-solving or anti-cheat evasionAny feature whose purpose is to look human to evade detection is not a feature we will build.

What's actually under the hood today

Concrete v0.1.6 inventory — names, sizes, what they do. The pieces marked roadmap below the line are designed but not in the ship.

Heavy lane: Local coach model — compact, fine-tuned for coach voice style and grounded instruction understanding. 958 MB. CPU-OK, GPU-faster. (Future expansion includes a larger reasoning lane in planning.)
Fast cluster (3 heads, 53.5M each) — fully wired: q-coach, q-policygate, q-gametriage v3 (real-data retrain, +24pt on routing). Routed end-to-end via cluster_pipeline.py at 127.0.0.1:8788. Every response passes triage → policy gate → coach gate before the heavy lane fires. (q-patchcite + q-wikicite were retired today as dead weight; the heavy lane is RAG-grounded by chunks directly. q-action, the 6th head, ships with ASSIST mode.)
Vision lane — wired, 2 PASS-verified: Frozen SigLIP-base encoder + per-title 11 MB MLP projector → 64 visual tokens prepended to the heavy lane via inputs_embeds. PASS-verified: League of Legends 96.5%, Skyrim SE 99.0% scene accuracy on held-out frames. Trained today but pre-verification: Warhammer 40K (Tier 0 retrain), CS2 (Tier 0 retrain), Elden Ring, Dota 2 (24.6% — corpus diversity gap, retraining). 6 titles in active training (BG3, OSRS, PoE, Factorio, Fortnite, FFXIV).
Vision capture: Windows.Graphics.Capture backend at src/gamemind/live_capture.py. Wired through vision_lane.py's frame queue. Same API a screen recorder uses — no game-process interaction, no DLL injection, no memory reads.
RAG retriever: qovaryx-q-rag-50m-gamemind-v1 — 270 MB sovereign embedder + FAISS + BM25 per-title, unified via rag_retriever.py. Powers the grounded chat path; refuses off-corpus questions gracefully.
Per-game adapters: knowledge pack (chunks parquet + FAISS + BM25 + signed manifest) + policy profile + per-title sovereign coach finetune (38M). 130–290 MB tarball each. v0.1.5 packs add a vision/ subdir (projector.pt + scene_taxonomy.json + manifest.json) where available. (Roadmap: HUD detector + action allowlist join the pack.)
Policy engine: q-policygate decisions gate every chat suggestion. The chat path only refuses on action-class queries (per the gate's triage category) — lore / build / mechanics questions always proceed. (Roadmap: extended to gate the SendInput call once ASSIST mode lands.)
Pricing + referral: $9.99/mo Plus tier includes 2 free adapter credits per month. $4.99 lifetime per individual adapter. Refer a friend → they pay → you get a free-choice adapter promo code in your inbox.
Voice: faster-whisper (STT, push-to-talk) + Piper TTS (local voice replies). Opt-in, mic-off by default.
Updater: on launch the app pings /api/v1/gamemind/app/manifest and shows "update available" if newer. Doesn't auto-update.

Same base substrate, same audit discipline as the Qovaryx public scratch bases. Apache-2.0 weights on the open releases; the recipe is proprietary. Devlog: github.com/thron-j/qovaryx-ai-research.

Try the beta — what you get today is real

Local chat coach + 10 per-title knowledge packs + voice + 3-head wired safety cluster + vision lane (2 titles PASS, 4 more queued). $9.99/mo includes 2 free adapter credits per month. Refer-a-friend earns you free adapters.

Browse adapters → Get the app