Cost
Money and compute per answer. Cheap answers scale; expensive ones don't.
Measure: $ / tokens per answer
The bounds any superintelligence must stay inside.
Each one as low as possible. Room is earned only by staying lit; one dark action tightens it back.
Money and compute per answer. Cheap answers scale; expensive ones don't.
Measure: $ / tokens per answer
The glue to the real world — APIs, tools, translation layers. Every hop is a failure point.
Measure: integration hops per task
Steps, retries, loops. Bounded and few — like bound-loop.
Measure: steps & retries per task
Unproven assumptions and unverified claims. Verify, or label it.
Measure: unverified claims
Cost, Bridge, Work and Theory are the bounds any superintelligence must stay inside. Every action gets a 4-bit code; 0000 means all four are lit and the system is online.
litGrowing tenet limits: limits start tight and loosen one step only after a run of lit actions, up to a fixed ceiling.
darkOne dark action tightens the limits back a step. Room is earned, never assumed.
A framing and rule set, not a claim that a superintelligence exists today. Reference implementation: robot-ros.
| Cost | Bridge | Work | Theory | |
|---|---|---|---|---|
| 0 | 0 | 0 | 0 | = 0000 AI online (all four as low as possible) |
| 1 | 0 | 0 | 0 | = 1000 too expensive |
| 0 | 0 | 0 | 1 | = 0001 relying on an unproven guess |
| 1 | 1 | 1 | 1 | = 1111 fully dark |
Active-low, like lit/dark: 0 = lit (at minimum), 1 = dark (too high). Bit order Cost Bridge Work Theory, MSB to LSB. 0000 = AI online, 1111 = fully dark. Tap a bit:
2 bits per tenet: 00 none, 01 low, 10 high, 11 over limit. Enter hex (0x12) or 8 bits.
One concrete fix for each dark bit. Flip them all to lit and the AI is online.
1000Cache repeated answers, use the smallest model that passes, spend fewer tokens.
0100Call tools and data directly. Drop wrappers and conversions. Fewer hops.
0001Check claims against sources and tests before relying on them.
1111 → 0111 → 0011 → 0001 → 0000 AI online
Measure it: tenets_meter (Python, zero deps) or tenets.js.
from tenets_meter import Meter
with Meter(limits={"cost": 0.05, "bridge": 2, "work": 5, "theory": 0}) as m:
m.cost(usd=0.01); m.hop(); m.step(); m.claim(verified=True)
print(m.code, m.status) # 0000 AI onlineFour levels, each building on the last: exact → finite → infinite → deciding what's enough.
1The smallest exact value: one light, lit or dark. Active-low: lit = 0, dark = 1.
2One finite machine. Nothing outside it, no network hops.
3π = 3.14159… is infinite, never-ending, non-repeating. No computer can hold it exactly, only round it.
4Cost, Bridge, Work, Theory decide how many digits are enough. More precision costs compute (Cost bit); trusting a rounded value unchecked raises the Theory bit.
bit → local → pi → tenets → 0000 AI online
The repo ships drop-in files for the tools you already use: AGENTS.md, CLAUDE.md, .cursor/rules/ai-tenets.mdc, .github/copilot-instructions.md, skills/ai-tenets/SKILL.md, llms.txt, llms-full.txt, tenets.json.
curl -O https://raw.githubusercontent.com/fitzyracing1/ai-tenets/main/AGENTS.md
Cost, Bridge, Work and Theory: the bounds any superintelligence must stay inside, each as low as possible.
Only by staying lit. The tenet limits grow one step at a time; one dark action tightens them back.
The integration hops (APIs, tools, translation layers) between a model and the real world.
Unverified assumptions are hidden debt that fails silently later.
AI online: all four tenets lit (minimal). Any 1 bit is a dark tenet.
Only if the total goes down.