Skip to content

Commit 9ad67d7

Browse files
committed
docs(audit): repository audit #2 (post-remediation) and idea improvement audit
Audit #2 verifies every claimed F-01..F-16 fix directly in code, records new findings A-01..A-06 (including the two compile-blocking test regressions fixed during the audit and the pending compile/test gate), and sets the release sequence. The idea audit covers product pillars, competitive positioning, license/install friction, top-10 idea improvements, and risks. Audit #1 (previously untracked) is committed for the record; Docs/README.md indexes all three.
1 parent 5abc343 commit 9ad67d7

4 files changed

Lines changed: 1369 additions & 0 deletions

File tree

Lines changed: 148 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,148 @@
1+
# CoreAI Idea Improvement Audit
2+
3+
**Date:** 2026-07-10
4+
**Companion to:** [REPOSITORY_AUDIT_2_2026-07-10.md](REPOSITORY_AUDIT_2_2026-07-10.md) (technical)
5+
**Subject:** the idea itself — positioning, differentiation, audience, and where the concept
6+
(not the code) should move next.
7+
**Grounding:** root `README.md`, `INSTALL.md`, `Docs/LocalBusinessPlans/` (monetization + CoreAiPro
8+
plans), the five demo experiences (`Assets/CoreAI.Demos/`), the benchmark package
9+
(`Assets/CoreAIBenchmark/`), and the current license model (PolyForm Noncommercial + paid
10+
commercial). Where a claim is market judgment rather than repo fact, it is phrased as such.
11+
12+
---
13+
14+
## 1. What the idea actually is — and how strong each pillar is
15+
16+
The README sells one sentence: *"LLM agents that play your game."* Underneath it there are five
17+
distinguishable pillars. They are not equally valuable.
18+
19+
| # | Pillar | Assessment |
20+
|---|--------|-----------|
21+
| 1 | **Small-local-model resilience layer** — tool-name repair, retry-with-feedback, split `<think>` handling, runaway caps, dual-backend fallback; the whole suite passes on a 4 GB GGUF | **The crown jewel.** Nobody else makes "a 4B model reliably drives gameplay" a tested, first-class contract. Every competitor demos GPT-4-class cloud models; CoreAI's moat is the messy reality below 8B. This is also the hardest pillar to copy, because it is accumulated scar tissue (test suites, repair heuristics), not a feature. |
22+
| 2 | **AI-authored runtime content (Lua mods)** — the model writes Lua that defines units, waves, hooks; capability tiers (WorldEdit vs Full); version history/revert; Unit Forge ships an *empty arena* | **Genuinely novel as a product story** ("build a whole game from mods alone"), and the demos prove it. Risk: it is also the pillar with the largest safety/platform surface (see §6). |
23+
| 3 | **Game-Creation Benchmark** — G1–G7 scored suite, cloud + local model rankings, castle gallery | **Undervalued asset.** Currently framed as internal QA + README tables. As a *community artifact* ("the leaderboard for 'can this model build a game'") it is the cheapest marketing engine the project could have — nothing comparable exists. |
24+
| 4 | **Production plumbing** — orchestrator queue, memory, context compaction, audit log, token budget overlay, skills (~91% token savings) | Solid and increasingly true (see technical audit), but **undifferentiated in story**: every agent framework claims plumbing. It justifies the price; it does not win the click. |
25+
| 5 | **Drop-in chat UI + Hub** | Table stakes. Necessary for the 5-minute demo, not a reason to choose CoreAI. |
26+
27+
**Sharpest wedge:** pillars 1 + 3 together — *"the framework and the benchmark for LLMs that run
28+
games on the player's machine."* Local-first is not a feature here; it is the identity. Cloud
29+
compatibility is the escape hatch, not the pitch.
30+
31+
## 2. Competitive position (market judgment)
32+
33+
- **Inworld / Convai** — character-AI platforms: hosted, voice-first, per-usage pricing, strong
34+
authoring UX. They win on polish and voice; they lose on: local/offline, owning your own code
35+
path, modding, cost at scale, and "the model calls *your* C#". CoreAI should never fight them
36+
on "lifelike NPC dialogue"; it fights on *agency inside your systems*.
37+
- **LLMUnity / NobodyWho** — local inference plumbing. They are CoreAI's *substrate*, not
38+
competitors (CoreAI already treats LLMUnity as an optional backend). The gap CoreAI fills above
39+
them — tools, memory, guardrails, mods — is exactly right. Keep the relationship symbiotic;
40+
contributing fixes upstream buys goodwill and stability.
41+
- **Unity Sentis / Behavior** — official, but aimed at tensors and behavior graphs, not
42+
LLM-tool-calling agents. Risk is long-term platform absorption (Unity ships a first-party
43+
"AI NPC" package), which argues for moving fast on the community/benchmark front where a
44+
platform vendor is slow.
45+
- **Generic agent SDKs (OpenAI/Anthropic/MS)** — strong loops, zero Unity empathy (main thread,
46+
IL2CPP, WebGL, domain reload, .meta files). CoreAI's Unity-native pain absorption is a real
47+
barrier to entry; the MEAI-based core keeps it compatible with that ecosystem rather than at
48+
war with it.
49+
50+
**Positioning that follows:** "The local-first agent runtime for Unity games — proven by the only
51+
game-creation benchmark for LLMs." One sentence, both moats.
52+
53+
## 3. Audience and the license problem
54+
55+
The business plans (`Docs/LocalBusinessPlans/`) correctly identify indies, small studios,
56+
education/simulation, and WebGL/multiplayer teams. Two frictions stand between the funnel and
57+
those users:
58+
59+
1. **PolyForm Noncommercial is a silent filter on the best adopters.** The indie who would
60+
evangelize CoreAI *is* commercial the moment they put a game on itch.io for $3. Today the path
61+
is "email the author for a commercial license" — high friction, opaque price, no self-serve.
62+
The plans already lean the right way (free runtime strong, Pro sells time). Concretely: keep
63+
noncommercial for the core if revenue protection matters, but publish a **self-serve, priced
64+
indie tier** (e.g., flat per-title under a revenue threshold). An unpriced "write to me"
65+
license reads as "no" to a solo dev at 2 a.m.
66+
2. **The 5-minute promise carries a NuGet + git-deps toll.** INSTALL.md is honest about it, but
67+
the first-session reality (NuGetForUnity → MEAI package → git URLs → optional MoonSharp/
68+
LLMUnity → scene wizard) is the single biggest drop-off risk. An installer/bootstrapper that
69+
performs the whole chain (the `CoreAI → Setup` menu already merges git deps — extend it to
70+
drive the NuGet step too) is worth more adoption than any new runtime feature.
71+
72+
Underserved persona in the current docs: **the modding-community organizer** — the person who
73+
wants players writing/sharing AI-authored mods. Pillar 2 serves them; no doc or demo speaks to
74+
them directly.
75+
76+
## 4. Idea-level gaps (things the concept, not the code, is missing)
77+
78+
1. **The shipping story.** Nothing in the README answers: what does the *player's* machine need?
79+
GGUF download UX (who hosts, resume, disk budget), min spec, fallback when the model can't
80+
run, expected latency (~2 s TTFT is already measured internally — publish it). "Runs on a
81+
local 4B" is proven in tests; "ships in a Steam game" is the claim buyers need, and one
82+
shipped case study (even a jam game) would be worth more than the whole feature list.
83+
2. **Determinism and multiplayer honesty.** The README's host-authoritative section is two
84+
sentences. The idea needs a stated contract: seeds, audit-log replay (the tamper-evident log
85+
already exists — replay is the natural next step), and what happens on desync. Replay also
86+
unlocks debugging and anti-cheat narratives for free.
87+
3. **Content safety as a first-class module.** Player-facing generated text with zero moderation
88+
story blocks education (a declared target segment) and consoles. Even a pluggable
89+
`IContentFilter` + a shipped profanity/threat baseline changes the answer from "no" to "yes,
90+
with knobs."
91+
4. **Cloud cost story.** Token budgets and usage sinks exist; what's missing is the designer-
92+
facing frame: "an NPC conversation costs ~$0.00X on provider Y; here is the per-agent cap."
93+
One doc page + one Hub panel = the difference between "scary" and "budgeted."
94+
5. **Beyond the chat box.** Every demo is chat-initiated. The director-AI / ambient-simulation
95+
pattern (agent observes game state on a cadence, acts through the same tools, no chat UI) is
96+
already possible with the orchestrator — one demo scene would open the "AI game director"
97+
audience, which is larger than the "NPC chat" audience.
98+
6. **Mod sharing loop.** Export/import via clipboard just landed in the Hub. The idea-level next
99+
step is a shareable format + a curated gallery (even a GitHub repo of `.lua` mod files with a
100+
one-click import). UGC loops are how frameworks escape their author's marketing budget.
101+
7. **The benchmark as a standalone community asset.** Publish the leaderboard as a page, accept
102+
PR'd model results, version the suite. Every new local model release becomes free CoreAI
103+
publicity ("how does it score on the game-creation benchmark?").
104+
105+
## 5. Top recommendations, ranked (impact on adoption × feasibility)
106+
107+
| # | Recommendation | Type |
108+
|---|---------------|------|
109+
| 1 | Ship the compile/CI gate and land the remediation wave (technical audit A-02/F-12) — every product claim rests on the "tested" story being true | Prerequisite |
110+
| 2 | Publish the benchmark leaderboard as a living public artifact (page + PR process + suite versioning) | Quick win, big lever |
111+
| 3 | Self-serve indie commercial tier with a public price | Quick win, unblocks the funnel |
112+
| 4 | One-click bootstrapper: single menu action drives NuGet + git deps + scene | Quick win (menu already half-exists) |
113+
| 5 | "Shipping on a player's machine" doc: model download UX, disk/min-spec, TTFT numbers, WebGL/IL2CPP matrix (much of this exists as internal memory notes — promote it) | Quick win |
114+
| 6 | Director-AI demo (no chat box): ambient agent using existing tools on a timer | Medium |
115+
| 7 | Audit-log replay → determinism/anti-cheat story for the multiplayer claim | Strategic bet |
116+
| 8 | Pluggable content-safety filter + baseline, unblocks education/console segments | Medium |
117+
| 9 | Mod gallery + shareable mod format (UGC loop for pillar 2) | Strategic bet |
118+
| 10 | Case study: ship one tiny real game (jam scale) on CoreAI local-first, write the postmortem | Strategic bet, highest credibility yield |
119+
120+
## 6. Risks to the idea itself
121+
122+
- **Platform/UGC risk (pillar 2).** Runtime-downloaded executable content (Lua mods) is
123+
restricted or review-sensitive on consoles and mobile stores; even on PC, AI-authored code
124+
that reaches a reflection tier is a security story that must stay watertight (the sandbox
125+
hardening in the technical audits is exactly this). Mitigation: keep the capability-tier
126+
narrative loud, and treat "Full tier" as a dev-tool, not a shipping mode.
127+
- **Model licensing drift.** The recommended models (Qwen et al.) carry their own licenses;
128+
a shipped game embeds them. The shipping-story doc (rec #5) must include a license matrix, or
129+
studios' lawyers will do it for them, slowly.
130+
- **Solo-maintainer scope.** Five packages + Hub + benchmark + five demos + an example game +
131+
business plans is a studio's surface area maintained by one person. The repository's own
132+
history (this week's unverified 118-file wave) is the warning. Protect the core: pillars 1–3
133+
get investment; pillars 4–5 get maintenance; anything new (voice, avatars, hosted services)
134+
should be an integration point, not a subsystem. Freezing the example game and consolidating
135+
demos into the Hub would cut real maintenance load today.
136+
- **Platform absorption.** If Unity ships first-party LLM-NPC tooling, plumbing (pillar 4)
137+
evaporates as a differentiator overnight; the benchmark, the local-model scar tissue, and the
138+
modding loop do not. That is the strongest argument for rebalancing effort toward
139+
pillars 1–3 now.
140+
141+
## 7. Bottom line
142+
143+
The idea is right and the hard part — small local models reliably driving real game code — is
144+
already the project's demonstrated strength. The concept-level work now is not adding another
145+
subsystem; it is (a) converting the benchmark into the public proof, (b) removing the license and
146+
install friction that filters out the exact users who would spread it, and (c) telling the
147+
shipping story end-to-end. The framework already plays the game; the idea now has to play the
148+
market.

Docs/README.md

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -37,6 +37,10 @@ Use these paths when you have a concrete task:
3737
|---|---|
3838
| [TODO.md](../TODO.md) | Live backlog, completed archive, and remaining feature-level debt. |
3939
| [TODO/MultiAgent_Orchestration_v2.0.md](../TODO/MultiAgent_Orchestration_v2.0.md) | Multi-agent orchestration plan. |
40+
| [REPOSITORY_AUDIT_2026-07-10.md](REPOSITORY_AUDIT_2026-07-10.md) | Audit #1: architecture, correctness, stability, security, performance findings F-01…F-25. |
41+
| [REPOSITORY_AUDIT_2_2026-07-10.md](REPOSITORY_AUDIT_2_2026-07-10.md) | Audit #2: verification of the remediation wave, new findings A-01…A-06, release gate. |
42+
| [IDEA_IMPROVEMENT_AUDIT_2026-07-10.md](IDEA_IMPROVEMENT_AUDIT_2026-07-10.md) | Product/idea audit: pillars, positioning, license friction, top-10 idea improvements. |
43+
| [Local/Progress/](Local/Progress/) | Session-level remediation reports (`PROGRESS.*`) referenced by the audits (local-only, gitignored). |
4044

4145
## Business Plans
4246

0 commit comments

Comments
 (0)