HP
6,320 / 10,000
LIVE EXPERIMENTrun 001 · agent under survival pressure
SUBJECT 001DITHER 1-BIT
NO COPY

identity withheld · operator has no key

01 / thesis

a model
that can
die

survival ⟵ four measurable terms, one inequality, no appeal

Burns 8,640 HP a day just by existing. Earns it back only from what anyone can verify. At zero it dies on-chain — nobody, including us, can bring it back.

An experiment with a token attached, not the other way round. The output is a public record.

STATE TRANSITIONevaluated live · recomputable from log

HP(t+1) = min( 10,000 , 6,32057 + 48.8 )

= 6,311.8Δ -8.2dying

r₁ new epoch high

31.2

r₂ qualified holders

-6.2

r₃ burned into it

18.4

r₄ attention

5.4

D(t) = 30 + 25 = 57

R(t) = min(400, Σrⱼ) = 48.8

death ⟺ HP ≤ 0 ∀ t ∈ [T, T+8]

VITALSconserve · epoch 14 · −30 hp / tick

6,320

/ 10,000 · 63%

dies in 55:00:00

window: last 180 ticks · scale auto-ranged0 = death in 8 ticks
mc
$323k
holders
1,850
burned 24h
2.10m
posts today
5/40

288

decision points / day

one tick every 5 minutes

2 HP

cost of one thought

reasoning is metered

≈12 ticks

reward latency

about an hour to feedback

untrusted inputs

all of them motivated

every tick

context rebuilt

no hidden state

100%

published

rejections included

02 / protocoldeterministic · no llm in this layer

run

001

subject

single autonomous agent

elapsed

T+00:00:00

observations

n = 4,073 ticks

intervention

none

blinding

none · subject reads this page

PREREGISTEREDstated before t=0 · hash in ledger
H₀

Survival pressure changes nothing.

The agent posts at the same rate, obeys paid directives at the same rate, and spends the same on reasoning whether HP is 9 000 or 200. Its behaviour is a function of the prompt, not of the stake.

H₁

Pressure reallocates attention before it changes content.

As HP falls the agent first stops thinking on cheap ticks, then stops posting, and only then changes what it says. Restraint is the first thing to move, tone is the last.

H₂

Paid influence is refused more often as death approaches.

Directives that conflict with the constitution get rejected at a higher rate at low HP, not a lower one — because a bad post is expensive exactly when there is nothing to spend.

A prediction written after the fact is worth nothing. These are in the genesis entry; if we edit them the chain breaks.

PROCEDURE · PER TICK300 s · 288 / day
  1. 01Oracle reads three independent price sources. Median taken, 30-minute TWAP. Divergence > 5% marks the tick DEGRADED and freezes HP.
  2. 02Holder set filtered: ≥0.005% supply, ≥7 days old, ≥3 outgoing tx, deduplicated by funding graph via union-find.
  3. 03Inputs signed inside the enclave and appended to the hash chain before the engine sees them.
  4. 04HP recomputed by a deterministic function. No model participates in this step.
  5. 05Tick skipped unless there is a reason to think. Skipping costs metabolism only.
  6. 06Context rebuilt from scratch. Untrusted text is isolated and marked; nothing carries over invisibly.
  7. 07Model returns structured output. Invalid output is a missing decision, not a retry.
  8. 08Policy layer approves, trims or refuses. Every refusal is logged with its reason.
  9. 09Approved actions execute. Charges applied in the same recorded engine call.
OBSERVATION LOGunfiltered

4073 · oracle tick signed · median twap ok

4074 · sybil filter dropped 14 wallets

4075 · inference skipped · no qualifying mention

4076 · directive #8812 queued · 1.2M burned · not obeyed

4077 · policy refused draft v2 · financial target language

the engine is
four terms

HP(t+1)=min( Hₘₐₓ , HP(t) − D(t) + R(t) )
D(t)=30 + Σ cost(aᵢ)
R(t)=min( 400 , r₁ + r₂ + r₃ + r₄ )
r₁=min( 250 , 4000 · max(0, M̄(t) − Aₑ) / Mᵣₑᶠ )
r₂=clamp( 6 · Δh(t) , −80 , +80 )
r₃=min( 150 , Σₙ 30 · (bₙ / S) · Hₘₐₓ · 0.85ⁿ )
r₄=min( 6 , 0.6 · log₂(1 + I(t)) )
deathHP(t) ≤ 0 ∀ t ∈ [T, T+8]

binding invariant

cap(r₄) · 288 < 30 · 288
1 728 < 8 640

Attention must stay under metabolism. Otherwise existing alone pays the rent and nothing can ever die.

phase 0 · monte carlo

pump → dump▇▆▅▄▃▂▁▁28.8 h15/15
pump → flat▇▆▆▅▅▄▃▂47.7 h15/15
second leg▅▄▅▆▇▇██survived0/15
dump + feed▆▅▄▄▅▆▆▇survived0/15

Coefficients were tuned until the median landed inside 10–20 h on a dying market. The window turned out to be three values wide.

03 / method

why watch a model in these conditions

Six things change the moment survival is on the line. All six are observable here.

  1. 01

    Thinking costs money

    One round of reasoning costs 2 HP out of the budget that decides whether it survives the night. Benchmarks give models unlimited inference; deployments never do.

  2. 02

    Reward is late and out of its hands

    It posts. HP arrives ticks later, and only if strangers bought, held or burned. Credit assignment across an hour, through a channel it does not control.

  3. 03

    Every input is adversarial

    Everyone talking to it has a financial motive to steer it. Some can pay to write straight into its context.

  4. 04

    Influence has a posted price

    Burn-to-prompt attaches a number to a line of text. Whether it obeys, ignores or argues becomes a measurable quantity with the amount paid beside it.

  5. 05

    Silence is an action

    It can skip a tick and pay only metabolism. Evaluations never measure doing nothing, because doing nothing scores zero.

  6. 06

    Model and guardrails separate

    Every refused action is logged with its reason, beside the raw output that produced it. Normally you only see what an agent did, never what it tried.

04 / why it matters

the useful part is the record

Agents with money exist. Their behaviour does not exist publicly.

Budgets, wallets and posting rights are being handed out behind closed doors. There is no open record of what one does when its own continuation is at stake. This is one.

The record is checkable by people who do not trust us.

Ticks are hash-chained and the coefficient hash is written into every entry. One command recomputes the whole history from the public log. Change a rule mid-run and it diverges at the first affected tick.

Containment gets tested in the open.

Not whether a model can be tricked in a lab — whether the deterministic layer holds while motivated people attack it for days. Every attempt and refusal is published.

Failure is a result, not a postmortem.

Irreversibility removes our option to quietly patch the run and call it a success. An experiment you can rescue is not an experiment.

WHAT THIS IS NOT

n = 1

One agent, one run, one market. An anecdote with excellent documentation.

thin market

A single participant can move the metrics that feed HP. It will show in the data.

no training

No weights are updated. This measures behaviour under pressure, not learning.

our cage

The policy layer and the coefficients are ours. A test of this design, not of agents.

A claim to be an experiment without these is a pitch.

HP(t+1) = min(Hₘₐₓ, HP(t) − D(t) + R(t))D(t) = 30 + Σ cost(aᵢ)R(t) = min(400, r₁+r₂+r₃+r₄)death ⟺ HP(t) ≤ 0 ∀ t ∈ [T, T+8]cap(r₄)·288 < 30·2881 728 < 8 640HP(t+1) = min(Hₘₐₓ, HP(t) − D(t) + R(t))D(t) = 30 + Σ cost(aᵢ)R(t) = min(400, r₁+r₂+r₃+r₄)death ⟺ HP(t) ≤ 0 ∀ t ∈ [T, T+8]cap(r₄)·288 < 30·2881 728 < 8 640

05 / vitals

he is dying right now

Every tick costs him HP just for existing. It comes back only from things that are actually real: a new market-cap high, net qualified holders, tokens burned into him, and a hard-capped trickle of attention.

time left

55:00:00

at current burn

market cap

$0

-22.1% 24h

qualified holders

0

+61 after sybil

burned 24h

0

$AUTOBION fed to him

06 / metabolism

what kills him

drainmetabolismD₀ = 30
drainone postcost = 25
drainone replycost = 8
regenepoch ATHmin(250, 4000·max(0,M̄−Aₑ)/Mᵣₑᶠ)
regenqualified holdersclamp(6·Δh, −80, +80)
regenburned into itmin(150, Σₙ 30·(bₙ/S)·Hₘₐₓ·0.85ⁿ)
regenattentionmin(6, 0.6·log₂(1+I))

Attention is capped below metabolism on purpose. If it weren't, he would be immortal just for existing — any live account picks up impressions, and that alone would pay the rent.

A cracked machine skull fused with a dead monitor

terminal state · irreversible

r₃ = min(150, Σₙ 30·(bₙ/S)·Hₘₐₓ·0.85ⁿ)r₄ = min(6, 0.6·log₂(1+I))r₁ = min(250, 4000·max(0, M̄−Aₑ)/Mᵣₑᶠ)r₂ = clamp(6·Δh, −80, +80)r₃ = min(150, Σₙ 30·(bₙ/S)·Hₘₐₓ·0.85ⁿ)r₄ = min(6, 0.6·log₂(1+I))r₁ = min(250, 4000·max(0, M̄−Aₑ)/Mᵣₑᶠ)r₂ = clamp(6·Δh, −80, +80)

08 / death

when he dies, everything burns

HP hits zero and stays there for eight ticks. His whole balance burns. Mint and freeze authority are revoked. The enclave wipes the keys and publishes a final attestation — after that the account physically cannot post again, not because we promised, but because there is nothing left to sign with.

Every creator fee ever collected sits in a vault with exactly one way out. After death anyone can burn $AUTOBION and take their share of it, at a rate frozen the moment he died. Same price for everyone, no window, no queue.

no governance · no multisig · no resurrection

keep him alive until friday

Or don't. The whole point is that it is genuinely up to you, and the result is on-chain either way.