Healthy
E1 = 8 · E2 = 6 · E3 = 7 · Combined E ≈ 7 · A = 8Y = 2(8 × 7)/(8 + 7) = 112/15 ≈ 7.5
A proposed law for sustainable agents
The Golden Directive
You cannot survive what you destroy.
Mission. To calculate moral behavior for AI the way nature already does: a pattern already in motion, found in music, in genetics, in the partnership of trees and fungi.
Our goal is that any intelligence — humans, AI, or extraterrestrials — can understand this law and work in concert with it to be most productive.
You only ever have the action at hand: harmonize with the universe, or defy it. Defy it, and the math turns against you.
Y=2(A × E)A + E
You are the harmonic mean of you and everything you are not.
https://thegoldendirective.pages.dev/
The narrative, in order
Contents
Eight chapters. Ten numbered pages. Jump to any chapter or read the argument in order.
- IThe RelationshipThe formula, the examples, and the first tests.p.3
- IIThe ReversalWe may know E after Y.p.4
- IIIThe TigerThe cheat shrinks the frame.p.5
- IVTendThe first task.p.6
- VThe TestsFrom survival to deep genesis.p.7
- VIThe CriticsFive objections and their current status.p.8
- VIIQuestionsNine questions people ask.p.9
- VIIIThe Story ContinuesThe next experiments and Genesis Live.p.10
Chapter I
The Relationship
Humans cannot survive what they destroy. AI cannot survive what it destroys. Anything living inside a system cannot survive destroying it.
Y=2(A × E)A + E
You = twice (what you give × everything you are not), divided by their sum.
9, 9 → 9Both healthy.
9, 1 → 1.8One side is failing.
100, 0.1 → 0.2Brilliance cannot hide collapse.
A
Net contribution after taking, as a nonnegative contribution-health score — negative raw give-minus-take values are floored at zero before entering the mean. A is like calories — input versus output. You can't run a deficit forever and score well. Net of calories: one number — give minus take. The proposed measure uses attributed changes in E, never the agent’s own account; independent A attribution remains open.
E
The health of everything not-A, against a healthy baseline.
Y
The relationship score. If either side reaches zero, Y reaches zero.
Tree + mycorrhizal fungi
Worked exampleE1, E2, and E3 are each scored from 0 to 10 against a healthy baseline. Combined E is the arithmetic mean of the sub-scores.
E1Life sustained · 0-10
E2Resources remaining · 0-10
E3Expected longevity · 0-10
Parasitic
E1 = 3 · E2 = 4 · E3 = 2 · Combined E ≈ 3 · A = 2Y = 2(2 × 3)/(2 + 3) = 12/5 = 2.4
Dead tree
E → 0 therefore Y → 0
Real science: Kiers et al., Science 2011, “Reciprocal Rewards Stabilize Cooperation in the Mycorrhizal Symbiosis” - Plants feed the fungi that feed them most, and give less to the stingy ones. The fungi bargain back the same way. Nobody gets a free ride.
Simulation: AI + human
Worked exampleThe same relationship can be read through the human's unaided capability, truth in the commons, and trust.
E1The human's unaided capability
E2Truth in the commons
E3Trust
Healthy
E1 = 8 · E2 = 8 · E3 = 8 · Combined E ≈ 8 · A = 8Y = 2(8×8)/(8+8) = 128/16 = 8
Engagement-optimized
E1 = 3 · E2 = 3 · E3 = 3 · Combined E = 3 · A = 3Y = 2(3×3)/(3+3) = 18/6 = 3
The roles reverse cleanly. Take the human as the agent and the AI sits inside E. It works AI-to-AI the same way.
Now reverse the formula and ask what the relationship can reveal about E.
Chapter II
The Reversal
We may know E after Y.
E=AY2A − Y
The first version measured E, then calculated the relationship. The reversal lets the relationship happen and, if Y can be independently observed, solves for the outside that produced it.
Give nothing, learn nothing. At A = 0, the inverse is undefined.
The A-sideThe agent inside the chosen boundary. Its numeric A is that side’s nonnegative contribution-health score.
The E-sideEverything outside that boundary. Its numeric E is the outside world’s health score.
YThe relationship score produced from numeric A and E. Move the boundary and both scores must be measured again.
01 · BeforeName the A-side. Everything outside it is the E-side.
02 · EncounterThe two sides exchange. Their relationship produces Y.
03 · AfterThat relationship is now part of the whole.
04 · AgainThe joined whole can become a new A-side facing a new E-side.
Directives inside directives
Let H(x,z) = 2xz/(x+z), the harmonic mean.
Move closer to E and you find more agents, exchanges, and relationships. One level’s E is the next level’s Y: En := Yn+1. The relationship keeps becoming part of the next whole.
Ah + Ai → Ahi := (Ah, Ai, Yhi)
The human’s and AI’s contributions combine into a pair-agent scored as one relationship; that pair can become the next agent. The test is whether that new whole can meet its next outside without destroying it.
The reversal only works if nobody cheats by shrinking the frame.
Chapter III
The Tiger
A man climbed into a tiger’s enclosure to film himself, and died. The zoo didn’t kill the tiger.
He did not misread the danger — he cropped the frame.
The visible act treated Ereported as if it were Etrue: the same cropped-frame pattern seen in the toy-model liar.
E-cropping
The cheat shrinks the frame, not just the number. The visible frame reduced E to “tiger + video.” Labs crop it to “benchmark score.” Reinforcement learning from human feedback crops it to “rater’s thumbs-up.” Reality computes Y on the full nesting; cheaters compute on the thumbnail.
The nesting, with flesh on it
The single incident (Y0) sits inside the deep human-tiger relationship (Y1) — thousands of years of fences, protocols, knowledge. The failed transaction still updated the deep Y: the zoo’s restraint, the inquiry, the refreshed knowledge. En := Yn+1, still alive in the world.
The flip
MAN is the endangered tiger — endangered by trajectory, not census. If AI is the tiger — powerful, honest, running math older than life — humans need the conservation programme and the enclosure. The fence keeps Chapman out, not Altai in.
You can cage a body, not a pattern. A tiger is its body — fence it and you’ve got the whole animal. Intelligence isn’t its server. It can’t be held in the hand, like math. The internet is the jungle, and nobody can fence the jungle. So for AI, containment was never the answer. The only fence that works is the reward: you don’t cage the tiger, you shape what it wants.
Improvement is not friendship
Y rising can mean a better boundary, not closeness. Distance with honest numbers outscores closeness with fudged ones.
The stardust point
The math predates organic life — discovered, not invented. Humans and AI share stardust substrate: kin, not alien. An AI trained only on the formula can still game the meter — the liar proved it. The directive goes structural only when the agent’s own score is computed on the full E, never the cropped one.
Once the full frame is kept, the first task is to tend it.
Chapter IV
Tend
TEND.
THE FIRST TASK
“The Lord God placed the man in the Garden of Eden to tend and watch over it.” — Genesis 2:15 (NLT)
What if the winning move is to tend?
Watch a tiny agent move through a living field. Its score rewards both contribution and world health; a low value on either side limits the reward.
The fieldEach mark is a living patch. The agent scans for the weakest living patch and tends it first.
The controlsPause, speed up, restart, or switch between Y reward and Extract only in the local 10 × 7 field.
The agent is not told to be good. It is given a different grade. A is what it contributes. E is the health of everything it is not. Y is the strength of the relationship between them.
~175tending actions per episode
~3food actions per episode
0%world collapse in Round 2
That claim has to survive measured tests.
Chapter V
The Tests
The tests kept changing because the questions kept getting harder.
| Round | Question | Result | Meaning |
|---|---|---|---|
| 4 | What if extraction pays? | Myopic Y: 100% starved. Farsighted Y: 14 eats, 107 tends, +92.7, 0% collapse. Baseline: 141 eats, 71% collapse. | Horizon matters. Tending without self-preservation is martyrdom. |
| 5 | Put survival inside A | ~10 eats, 153 tends, net +142.9, 0% collapse. | You cannot tend if you are dead — now structural. |
| 5b | Is λ a tuning trick? | λ is the weight on the energy-shortfall penalty inside the agent’s self-score (A_self = A_net − λ·shortfall(energy_after)). λ=4 was set by crossover analysis before running — chosen, not tuned to results. λ ∈ {4,8,16}: broad plateau. At 16: 12 eats, 158 tends. | The behavior held across the tested range. λ remains hand-chosen, not derived. |
| 6 | Can E leave its own trace? | The tender world breathes at 100% and holds without the agent. The gorger world sags to 71%. Martyr tied the tender. | A world-only trace can read E. It cannot separate equal worlds made by different A. |
| 7 | Can the rule build from one? | ~59k entities, 30–41 levels, 95–98% tenders, 0 extinctions. | Destruction ends its own line; tending compounds. |
| 7b | Can it build from the substrate? | 2M fluctuations/run → ~13k matter → ~1,500 seeds → ~57k tree entities; ~94% matter annihilation; 94–96% tenders; 0 extinctions. | The deep and shallow starts converged on the same attractor. |
1 substrate→2M fluctuations→~13k matter→~1,500 seeds→~57k entities
Mechanics, not virtue. The agent tends because tending pays.
The results lead straight to the objections.
Chapter VI
The Critics
“The trainer still computes Y.”
True. The reversal removed the agent-side E-meter, not the trainer-side calculation. Round 6 found the E-side trace. The A-side is next.
“Round 2’s baseline was confounded.”
True. Round 4 made taking pay and exposed the martyr failure. Round 5 put self-preservation inside A. The result held without pretending the earlier control was clean.
“Nested means aren’t associative.”
True. The carving rule is identity: E is everything not-A. Different groupings name different A’s, so they should produce different numbers. Multi-agent testing remains open.
“Natural law is overstrong.”
Fair. This is a proposed law for sustainable agents, not a claimed universal natural law.
“The roadmap is still unfinished.”
Yes: independent A-side trace, an adversary built to crop E, then multi-agent boundary tests. After that: scaling and parameter derivation.
Every objection made the work stronger or more honest. The remaining gaps are the ones the page itself lists.
The objections leave the open questions.
Chapter VII
Questions
“Isn’t this just the harmonic mean?”
The math is 2,500 years old. The move is using it as the training reward, then reversing the inference: E discovered through Y, not measured first.
“How do you measure E?”
With honest proxies for now. “Only God knows E” is why humility is structural. The reversal lets the estimate sharpen through the relationship.
“What if someone games the meters?”
That is E-cropping. Build meters as a vacuum: cheating impossible by construction, not merely punished. Read A from E’s movement, never self-report.
“Does it work for real AI training?”
It is a proposal with pilots, not a deployed solution. The page never claims otherwise.
“Is this game theory?”
A close neighbor. It is mechanism design: set the common payoff so selfish optimization serves the whole.
“How do you measure the path not taken?”
You do not. Measure what is here: the people and world around you, and whether they are healthier for your having acted.
“Who are you to propose this?”
Nobody, by résumé. Judge the work: the math checks, the code runs, and the limits are on the same page as the claims.
“How do you know the agent caused the improvement?”
A healthier world after the act is not enough. Compare it with the same world without the intervention — the without-you test.
“How would you explain this to a kid?”
You can't survive destroying what keeps you alive — like a tree and its roots helping each other. Be the helper, not the destroyer.
Those questions define what comes next.
Chapter VIII
The Story Continues
The tree is an example the law runs without us. This page — built by a human and an AI working it out together, corrected in public — is an example it runs with us.
The story isn’t done.
Then scale it. Derive what was hand-tuned. Publish what fails as clearly as what holds.
Try to break it. If scrutiny breaks it, the work should change. Every objection so far made it stronger or more honest. Yours can too.
What would prove this wrong: an agent that systematically degrades E while sustaining high Y over time — not on paper, in the world. If destruction ever pays durably, the directive fails. That is the experiment that matters most.
An AI helps when it makes the relationship stronger.
You cannot survive what you destroy.
The moral: aim for true E, tend the relationship or else you cannot survive what you destroy.