AI Decision System
Replayable, offline-verifiable records for AI-assisted go/no-go decisions: experiment promotions, typed model answers and human-approved actions.
A SQLite-compatible database engine written in Clojure by a crew of AI agents, judged by SQLite's own test files.
Translation pending
Developer toolsClojure · PLpgSQLGitHub @hanialshaterOnlineOwner unconfirmed
Have you tried it? One tap is enough. No sign-in needed.
No updates detected yet. GitHub releases will appear here.
Owners of public repositories under a personal GitHub account can sign in to verify ownership and manage the introduction themselves. No administrator approval is needed.
Sign in, generate a verification file, and commit it to this repository’s default branch within 30 minutes. Files on other branches or forks do not count. Then check the file below to manage the project without administrator approval. Keep the file to retain your access.
If needed, you will go to sign-in. After signing in, return here and press your chosen verification button again.
Run /nomorevibe verify in your project folder to verify ownership and manage the introduction yourself.
Verify in one minuteAsk about other verification options
The maker's introduction consists of the line above. Read the repository README for more details. Repository README ↗
Loading the full README.
Read the full README ↗Demo site: https://system3-sqlite.vercel.app (built from real runs; site/README.md).
A SQLite-compatible database engine written in Clojure by a company of AI agents in the System 3 city. Its referee is SQLite's own test suite: the engine passes when every statement in SQLite's test files gives the output SQLite gives.
| Hidden test files (480), mean pass rate per file | 91.85% (B13, 2026-10-25) |
| All hidden statements | 93.59% |
| Visible test files (221), mean per file | 96.66% |
| Built by | an eight-member crew: an architect and lead, six area engineers, one deliberately fallible fast lander |
| Cost so far | $1,282.87 across three runs (B11, B12, B13; company/company.json) |
Java 21 or later.
./run.sh --fetch-jars # Clojure 1.12.6 and its two spec jars into ./lib (or set CLJ_LIB)
echo "CREATE TABLE t(a); INSERT INTO t VALUES(1),(2); SELECT sum(a) FROM t;" > /tmp/t.sql
./run.sh /tmp/t.sql
./run.sh --batch list.txt --out out/ # many scripts, one output file each
With the Clojure CLI: clojure -M:run script.sql.
referee/judge.mjs scores a tree:
fast runs the visible corpus (spec/b12/corpus, 1,250 scripts from SQLite's test files).full runs the hidden corpus on a clean copy of the tree's HEAD, with no SQLite reachable.The hidden set is held by the referee's owner, outside this repository. See referee/README.md.
company/ is the company as data, read by the System 3 HQ through Meta's tools (system3-meta,
docs/product/STARTER.md §3):
| File | What |
|---|---|
company.json, milestones.json | mission, measure, budget lines and milestones |
org.json | the crew and its reporting lines (the org contract) |
events.jsonl | B13's event log (v1): every land, refusal, note, claim and spend |
knowledge/entries.jsonl | 245 lessons the members recorded when they left, with evidence and belief, by team (run) |
docs/ holds the run designs (predictions committed before each run) and the results: B11.md/BENCH11.md,
B12.md/BENCH12.md and B13.md/BENCH13.md. spec/ holds the specs the members worked from, and its
KNOWLEDGE.md.
From a checkout of hanialshater/system3-meta beside this one:
node ../system3-meta/hq/build.mjs --city=hq/city.json --snapshot=sqlite --out=<snapshot dir>
The git history is the crews' own: members commit under their role names (m-architect, m-engine, …), and
operators commit at the start of each run.
MIT (LICENSE).
SQLite's test files, from which spec/b12/corpus is derived, are in the public domain.