Personal R&D · Knowledge Engineering / Python Tooling
Game-Forge
A private knowledge base of game-design requirements, decisions and system specifications, frozen as a versioned, hash-pinned snapshot for AI-HQ to build games from.
- Status
- Completed · frozen
- Source
- Private source
- Period
- October 2026
- Role
- Owner: game decisions, review and approvals
- Python
- Git
- GitHub Actions
- AI Agents & Multi-Agent Systems
- Software Architecture & System Design
- Markdown
- CSV / JSONL
- SHA-256 provenance
01Interactive exhibit
Stages
S01–S36 · 206 requirements
Specifications
Specifications across 36 modules, written from my own source documents, with 168 dependency edges and 49 invariants. Stages 0.1 to 0.3 were approved; the Stage 0.4 drafts were closed as superseded when the mission changed.
02
Context & challenge
Building games with AI agents needs my gameplay requirements, design preferences and technical decisions written down precisely, with what is approved kept apart from what is only proposed.
That knowledge also has to reach another system intact: an agent must never read a proposal as a decision, or quote a document that has changed since.
03
Approach
Game-Forge began on 18 October 2026 as the specification phase of a game foundation. After a mission change on 24 October it became a temporary knowledge source for AI-HQ, which will build the games. Nothing was deleted: every requirement and decision was reclassified as binding game knowledge, reusable guidance or a retired framework obligation.
Approval states come only from structured sources (CSV columns, decision tables and dated ledger records), never from prose. Python tooling validates the documents and exports them from a pinned commit into a versioned snapshot in which every record keeps its source text, path, commit and SHA-256.
04
My role
Personal R&D project. I made the gameplay decisions and approved every stage, the mission change and the final freeze. AI agents drafted most of the documents and tooling under my direction, and other models reviewed them before I approved.
05
What was delivered
Specification corpus
Specifications S01–S36: 206 requirements across 36 modules. They also hold 49 invariants and 949 documented acceptance scenarios, which have not been run because there is no game code yet.
Decisions and ledger
62 decisions and a dated ledger at revision 69. The decisions sit in 14 architecture decision records, and the ledger holds 129 owner-decision records. Open conflicts and pending decisions are kept, not hidden.
Validator and exporter
Two Python tools, each tested by injected faults. gfdocs runs 17 documentation checks and gfexport 22 snapshot checks; their tests inject 144 and 139 faults that the checks must catch.
Knowledge snapshot v1
3,292 records, each with its provenance. Handoff contract v1.0 defines what AI-HQ will receive, and a re-export under another time zone, locale and hash seed matched byte for byte.
06
Engineering decisions
No state flattening
Approved, proposed, deferred, superseded and withdrawn stay distinct; 420 records are flagged never to be applied as approved.
Approval outside the artifact
A snapshot cannot contain its own approval without changing its hash, so it is approved in the ledger by its ID and SHA-256.
Owner-only gates
The validator accepts a stage as approved only with evidence from the named pull request's merge, so agents cannot approve their own work.
Deterministic, portable export
No wall-clock time in the output, line endings pinned to LF and a Windows CI job that compares real bytes, so the same commit always yields the same snapshot.
07
Validated
- Final verification at the freeze: 17 of 17 documentation checks, 144 of 144 injected faults caught and 22 of 22 snapshot checks, including a byte-identical re-export (run locally on Windows 11).
- CI on Ubuntu and Windows; every run on main passed. The exporter's tests ran locally on Windows only.
- Each approved stage and the freeze carry a dated owner approval in the ledger; reviews by other AI models ran before each approval.
08
Roadmap: not yet completed
- Importing the snapshot into AI-HQ's knowledge fabric once AI-HQ's importer is approved. AI-HQ's own handoff format differs from this snapshot, so a bridge is still needed.
- A pilot game built by AI-HQ from this knowledge, after which Game-Forge is retired.
09
Status & outcome
Frozen on 31 October 2026 with my approval of handoff contract v1.0 and snapshot v1. AI-HQ does not read it yet; the import is AI-HQ's next step.
10
Attribution & publication notes
- The stage view on this page is an original illustration. The knowledge itself is private and is not published here.
- Game-Forge is not a game, an engine or an SDK, and it contains no game code. Repository is private.
Next project
Lore-Forge