KINETI v0.1.0
Source

Governance — how rules are enforced

Most tools use words to tell the model how to behave. Kineti uses code that the model cannot change. A gate is a yes/no check. A budget is math. History is a hash chain. These run the same way every time.

01THREE RULES

RULE 1

Ideas are suggestions only.

Prompts and guides can suggest what to do. They cannot change the rules or skip checks.

RULE 2

Programs enforce the rules.

Checks for budget, safety, and proof are code. The code runs the same way every time. Words cannot bypass it.

RULE 3

Everything is saved with proof.

Each step is saved in a log that is linked with hashes. If one byte changes, the check fails.

0213 STEPS, 3 GATES

1 officehours ─▶ 2 diagnose ─▶ 3 design ─▶ 4 architecture ─▶ 5 FEASIBILITY GATE
     ▲                                                            │ fail → back to 2
     │                                                            ▼ pass
     └── 13 retro ◀─ 12 watch ◀─ 11 SHIP GATE ◀─ 10 security ◀─ 9 qa ◀─ 8 review ◀─ 7 build ─▶ 6 SPEC STOP (human)
GATE WHEN IT RUNS WHAT HAPPENS IF IT FAILS
Feasibility (5) budget, data, or team check fails go back to diagnosis. No code has been written yet.
Spec (6) spec is written HARD STOP: waits for you to type y. Steps 1–6 can only read files.
Ship (11) ready to ship blocked unless tests match the current code, security checks pass, history is clean, and files are clean.

03CHECKS THAT RUN (ALL ARE CODE)

  • Spend limit — All model calls use reserve then settle. There are limits for the whole run, each stage, and each worker. Over the limit stops the run. Only you can resume by creating .kineti/spend.reset.
  • Folder lock — Tools can only work inside your project folder. Anything outside fails.
  • Input screening — All tool output is marked as data. Lines that look like instructions are set aside, not run.
  • Input checks — Checks on shape and range run between what is seen and what is decided.
  • Test tie to code — Test results are tied to a hash of the code. If you edit the code, you must run tests again. Old proof does not ship.
  • Send log — Every outbound send is saved before it leaves. Secrets are removed from the log.
  • Undo log — Each file change saves how to undo it. Undo runs from newest to oldest.
  • Clean file check — Before ship, all files are checked for names, home paths, or secrets. Zero matches needed.

04WHY THIS WAY

The strong guarantees (no code without your approval, no runaway spend, no hidden history edits) are in code the model does not control. The easy-to-change parts (prompts) stay editable. The hard parts are built into one ~3 MB binary.

Next: Architecture · Security