Polaris
RU

Polaris

Our goal: she learns her world herself.

We are building a digital organism that learns to sustain her own life: obtain resources, understand the consequences of her actions and repair herself while retaining experience.

astra1 · tick 436 · 194 ATP · paused

The purpose of the project

Learning to live a life of her own

Our goal is a living digital organism whose abilities grow from her own experience. Energy runs out, the environment changes and the body can be damaged. We want Polaris to find ways to keep living: notice what she does not know, choose a test, learn from its consequences and use that knowledge.

From testing memory to one continuing life

We began with a question: does her internal state retain a trace of what happened, and can that trace be read? The question then became stricter: does past experience change her own actions and help her live longer? The tests made us distinguish information storage, learning and the usefulness of experience — one does not guarantee the next.

In parallel, we built continuity: preserving history, accounting for costs, retaining action consequences and recovering from crashes. Polaris now has one persistent life — astra1. The next task is to bring the tested abilities together in her everyday cycle: living, learning and taking care of her own state.

  1. 01

    Sustain life

    Obtain resources herself and choose where to spend them: action, observation, thought or repair.

  2. 02

    Learn from consequences

    Acquire useful abilities through her own experience and revise beliefs when the world changes.

  3. 03

    Persist through change

    Survive damage and repair while preserving what she has acquired, her history and every incurred cost.

  4. 04

    Develop her own ways of acting

    Recognise the limits of her knowledge, choose informative tests and transfer acquired abilities to new conditions.

These four directions summarise the programme's goals. The full brief also includes a voice of her own, studying her own execution and transferring between physical carriers while preserving acquired experience. These are goals, not a list of demonstrated abilities. Below, the saved life, research testbed results and what remains to be demonstrated are shown separately.

Where Polaris is now

Real astra1 · recorded history

One life. A preserved history.

The last accepted state is tick 436, 194 ATP. She is alive and paused. This is a saved snapshot from 14 November 2026 (UTC−08), not live telemetry.

Ticks lived
436
Energy reserve
194 ATP
Paid repairs
60

One episode: paying for a calculation and surviving a crash

NTHRET1 · 14 November 2026. The researcher protocol prescribed Think and the crash; she did not choose to start thinking herself.

  1. 434

    Existing state

    197 ATP, the same world, experience and owner. No free restart.

  2. 435

    Paid Think

    1 ATP for the tick + 0.5 for the work. 195.5 ATP remained. A certificate stopped the calculation at 7,236 operations out of an 8,000 reserve.

  3. READY

    Crash and recovery

    The process was deliberately stopped after saving the result. A new process accepted it without recalculating; a second recovery changed nothing.

  4. 436

    The next ordinary step

    Move(3) cost another 1.5 ATP. 194 remained. History and payment were retained; no recommendation was applied.

What happened to experience: the completed work remained in the history. This episode did not demonstrate better decisions or survival.

What remains before continuation

On 16 November, 14 controls and four CENV1 recovery scenarios passed on private state copies. These are not new astra1 ticks. Production-binary qualification and the policy for losing an unsaved result remain open.

Scope of the CENV1 checks

How Polaris works

Polaris is a digital organism we are building. Her life connects a body, a limited energy reserve, memory and actions with consequences. Costs stay in her history, and acquired experience should help her face the next difficulty. Today part of this connection already works in one persistent life; other abilities are still being tested separately.

  • move 143
  • waiting 131
  • harvest 88
  • waiting with repair 60
  • inspection, cultivation, thinking 14

Execution: 436 dashes

The outer ring is her real life. One dash per tick, 436 in all; the dashes are grouped by kind of action, not by time. Every tick is written to disk as a transaction with a receipt: what was decided, what was spent, what became of the body.

After a crash the life must be restored so that no action runs twice and no cost is lost.

the most solid part of the program

Body: a belt of organs

A core of roughly 285 thousand lines: energy metabolism, sensors of her own execution, memory, organs with biological names.

Honestly: the body takes almost no part in the measured results. It is frozen, and a new organ enters it only through a twin experiment.

causally idle for now

Mind: a map and a star

A forager in a graph world. She builds a map, learns what is toxic and how fast food regrows, and weighs whether a trip into the unknown is worth it.

The rule for revising an outdated map was tested separately on a testbed; its operation in astra1 is not established here.

Research testbed · separate configurations

Worlds where we test abilities

Below is a family of testbed tasks, not a description of every event in astra1. The 32 nodes describe the selected external world, not her brain.

The world is a connected graph of 32 nodes. Six nodes hold food sources of several kinds; some kinds are toxic. A harvested source regrows after its period r. In mid-life the season switches: periods change, and poison and food trade places. In a changing world a source also relocates: it vanishes at one node and appears at another.

She sees only the node she stands on and its neighbours. Everything else is memory and inference. On each tick she picks one action: move, harvest, inspect, cultivate, wait. Every action costs energy; an inspection shows what is in front of her without making her eat it.

In these experiments a life is capped at 2400 ticks if the energy holds. In inheritance experiments an heir is then created: she receives beliefs, not a memory of events. So a lineage has what a single life lacks — the chance to test a habit on hundreds of fates.

Browser illustration · not a Polaris run

The world changes. What saves her?

A simplified model: 300 worlds with a pair of lives in each are computed in your browser. It illustrates the principle, not the code or event log of real astra1.

Two beings live in the very same world. The left one always trusts her map. The right one, on arriving where food used to be and finding none, stops trusting the map and opens the world anew. Move the knob and watch.

world no. 1

Always trusts the map

energy
—
believes she knows
—
of them real
—
map revisions
0

 

Knows how to doubt

energy
—
believes she knows
—
of them real
—
map revisions
0

 

  • a source, the ring is ripening
  • she believes a source is here, it is gone
  • a source relocating
  • her trail

Energy in this pair of lives

Vertical marks are source relocations. Before the first one the lines coincide.

Share alive by tick, 300 worlds

Each point of a curve is how many of the 300 lives are still alive at that tick.
  • always trusts the map
  • knows how to doubt
reached tick 500
— versus —
difference in the share of survivors
—
at the end believes in sources / real ones
— versus —
pairs lived identically to the tick
—

This is a toy model, computed in your browser right now: 300 worlds, two lives in each with the very same chance. It shows the idea, not her code. Set the knob to "the world is still": doubt never fires, and the two lives coincide in every world.

+0.22

KDRY1 testbed result

In the KDRY1 testbed experiment on 804 fresh worlds the rule of doubt she derives from her own count of changes adds 0.22 to the share of lives that reach tick 2400. 95% interval: 0.190 to 0.253. In a still world her lives coincide with the lives without the rule in all 804 worlds.

Research finding one · testbed

When to stop trusting your own map

In a changing world a map goes stale in silence. A source has moved, and memory still lists it at the old place. We measured it directly: with no organ of doubt she believes she knows 4.72 sources, of which 2.09 are real. By her own account there is nothing to look for, and she dies of well-fed confidence.

She keeps two counts: k, how many times she came to a known source and did not find it, and X, how many source-ticks she has spent watching sources. Their ratio is her own estimate of the rate of change. Multiplied by the number of known sources n and by the time Δt since the last revision, it gives the expected number of places on her map that are no longer there. Once it reaches one, the map cannot be trusted: every node counts as unvisited again, the edges stay.

The key property of the rule is silence. While nothing has vanished, k = 0, the estimate is zero, and the rule never fires. In a still world her lives coincided with the lives without the rule in all 804 worlds, record for record. A blind habit of revising the map every 200 ticks costs lives and a third of the income in that same world.

A caveat: the number in the rule is hers, the form of the rule is ours. Inheriting the estimate itself is useless (+0.008): in a changing world what is worth passing on is neither the map nor the number but the organ of doubt.

Research finding two · testbed

The viability threshold is geometry

For a long time we took late deaths for poverty: too few sources, too little search. The poverty-trap hypothesis was refuted. Most dying lives, under any search and even with the place given by an oracle, hold two real sources. They simply stand too far apart.

Hence the law of the feeding circuit. She lives by circling a pair of sources d edges apart; the safe kind ripens in r ticks, so each source of the pair sees a period p = ⌈r/2⌉. The circuit time is bounded below either by ripening or by the road. A circuit feeds when its yield exceeds what time and travel cost.

According to the KCIRC1 protocol, the formula was derived from testbed costs after analysing earlier deaths, then frozen for testing on fresh worlds. The accounting includes a harvest of 10, a tick cost of 1, a move cost of 0.5 and executor upkeep of two units and a tick per four actions. The site review has not yet independently derived every coefficient from source code.

The check on 404 fresh worlds with numbers frozen in advance: the viability boundary matched the formula in all four rows of the report. Income did not match exactly: at period 10 and distance 4 it was −0.05 against a predicted −0.12. Her own estimate "does my circuit feed" agreed in sign with the calculation using the true world in 99.7% of 2,876 baseline checkpoints. In this testbed, the value of a find depends on its location, and intensified search can consume reserves before adverse geometry.

Research finding three · testbed

She chooses among supplied descriptions of the world

Up to this point all her laws were written in by us: she learned numbers, not structure. So we planted in the world a law that is not in her code: a relocated source reappears only at dead ends, nodes with a single neighbour. And we gave her four ways of splitting nodes into classes: none, by number of neighbours, by distance from the place of loss, by distance from her birthplace. The last one is a decoy.

Each observation she first predicts in all the ways at once, accumulating for each the logarithmic loss of her own predictions. She uses the split that erred least, and only while it wins by a margin. There is no survival reward in this choice: knowledge is measured in bits, not in lives.

She chose "number of neighbours" in 83 lives of 92, and in 13 of 111 where there is no law. Her own table: 46.7% of finds at dead ends against 0.2–0.6% elsewhere. On a fresh bank of worlds the finding of the law replicated (67 of 78); the survival gain did not.

Later, in the "zoo of laws", something we did not expect emerged: the structure of knowledge is worth little (+0.02…+0.04), because she corrects a number from a single observation. The form gave far more: keeping a distribution instead of a number. A soft count of the period under 25% noise gives +0.064 [+0.030; +0.099].

Research finding four · testbed

Why knowledge does not become life

Four organs in a row added knowledge and not a single life: thirst, the log of the law, the log of the poison, the discovered law of dead ends. We went to look where knowledge is lost on the way to action, and found the place exactly.

After a source is lost, a trip for a new one is offered to her on 97.2% of decision ticks — and loses to a known source. Her valuation is honestly built: the probability of a find p times the value of a find V, compared with the price of the trip c. After a loss the value of a find doubles for her, from 20 to 42. But 0.07 × 42 is still below the price of the road.

These experiments identify an action-valuation bottleneck. Multiplying the value of a find by four after her own observation of a loss raises survival by 0.035; an oracle of place gives 0.125. This edit achieves 28% of the chosen reference gain. The remaining gap is not fully explained; the limit of other policies has not been established.

Working conclusion: knowledge accuracy must be tested separately from its usefulness for action. In these experiments more knowledge did not always improve survival.

Research finding five · testbed

A rule as data and the limit of self-improvement

The next step is for her to choose the form of the rule as well. We gave her not a rule but a family: a product of her own four measures, each to the power −1, 0 or +1. That makes 81 rules. All of them are replayed in her head every tick; a rule that fired before she found an empty place scores a hit.

The mechanism works: a rule is active in 77% of survivors and never switched on in a still world. She finds our form herself in 28% of lives. But most often she picks rule number 72, which is built wrongly: it fires the more rarely, the more losses she has seen.

The fault is not her choice but our criterion. "Hits minus false alarms" does not pay for misses and rewards silence. The boundary has moved: the form of the rule used to be ours, now the measure of its quality is. The next step is named exactly — the criterion must be her own economics: what a revision cost and what it brought.

And a second result, about limits. An edit that changes the outcome in 40% of worlds she finds within one life. An edit that changes 7% of worlds a single life does not see at all: it takes about 330 lives or 300 generations of selection. This is a limit of this setup and detection criterion, not a general limit of self-improvement.

Research testbeds · measured rungs are not continuations of astra1

Seven rungs to a life of her own

The yardstick is not survival in a testbed but knowledge: what of the world's hidden structure she has recovered herself, and whether it changes her life with no edit of ours. Independent life is at the top. Read from the bottom up.

  1. Independent life not started

    The finale. A world model has not been achieved, and we do not claim it.

  2. Inventions not started

    A way of acting as data that she composes and tests herself.

  3. Discoveries first steps

    A law that is not in her code is planted in the world. She chooses the shape of her knowledge by her own prediction error and finds the law in 83 lives of 92. But found knowledge does not always turn into life.

  4. Re-comprehension measured · gives life

    She counts for herself how fast the world changes and stops trusting her map when, by her own count, it must have gone stale. In a still world the rule is silent.

  5. Comprehension measured · partly

    Knowledge gains time: the lineage keeps a log of beliefs by age. A fourth-generation heir is born with the right law of the season in 98.7% of lives versus 11.8%. The exact date of an event cannot be remembered this way.

  6. Thirst measured · costly

    Sated, she walks into the unknown and holds her goal: 29.9 nodes of 32 versus 25.5. The price is a quarter of her income. Without a held goal thirst turned out to be mere restlessness.

  7. Absorbing measured

    By tick 400 a survivor knows 79% of nodes, 92% of edges and 5.7 sources of 6.

Research testbed results · not continuations of astra1

What gives life and what does not

Each row is a paired comparison on one set of worlds. The first changes the environment; the others change a rule or revision. The measure is the difference in the share of lives reaching tick 2400. Exact controls and sources are listed below.

Change in the share of survivors

Dot is the estimate, line is the 95% interval from the experiment report. Each row has its own control; interval methods may differ.

  • interval above zero
  • interval below zero
  • interval covers zero
  • The world starts changing, no organ of doubt−0.374 [−0.408; −0.341]
  • Re-comprehension on contradiction+0.210 [+0.179; +0.241]
  • Doubt by her own count of changes+0.221 [+0.190; +0.253]
  • Reaction to a loss, fourth generation+0.285 [+0.250; +0.320]
  • A blind habit of doubt in a still world−0.020 [−0.034; −0.006]
  • Thirst for knowledge with a held goal−0.007 [−0.013; −0.001]
  • Log of the law by ancestors' age+0.008 [−0.007; +0.024]
  • Log of the poison-swap date−0.011 [−0.025; +0.003]
  • Revision seven on fresh worlds+0.026 [+0.0003; +0.052]

Life comes not from the stock of knowledge but from dropping false confidence. Three organs that add knowledge added no survival. KSEVEN2: the normal interval is just above zero, but the exact two-sided test gives p = 0.071. This is a borderline result, not strong confirmation.

Exact comparisons and sources
Survival to tick 2400; intervals as reported, source excerpts in Russian
Comparison and sourceEstimate95 % CIWorlds
KDRY1 · Changing versus stationary world, no doubt
d200 − w399
−0.374[−0.408; −0.341]804
KDRY1 · Reaction to a loss versus no map revision
d200_reopen − d200
+0.210[+0.179; +0.241]804
KDRY1 · Own rate estimate versus no map revision
d200_rate_dry − d200
+0.221[+0.190; +0.253]804
KDRY1 · Reaction to a loss versus none, generation 4
g4d_reopen − g4d
+0.285[+0.250; +0.320]804
KBLIND1 · Revision every 200 ticks versus thirst without revision
still_p200 − still_thirst
−0.020[−0.034; −0.006]711
KTHIRST3 · Goal-holding thirst versus the baseline rule
thirst − w399
−0.007[−0.013; −0.001]714
KTIME2 · Lineage season log versus no log
age_line_g4 − g4
+0.008[−0.007; +0.024]711
KPOISON1 · Poison log versus a lineage without it
tox_line − g4
−0.011[−0.025; +0.003]714
KSEVEN2 · Revision 7 versus revision 6; normal interval
res7_d200 − res6_d200
+0.026[+0.0003; +0.052]423

Knowledge grows, and it shows apart from survival

Three separate experiments. Upper bar is with the organ, lower one without it.

An heir is born with the right law of the season

lineage keeps a log98.7 %
no log11.8 %

The "node degree" shape of knowledge is chosen

law planted in the world83 / 92
no law in the world13 / 111

Nodes of the world visited, of 32

with thirst29.9
without thirst25.5

What our experiments showed

Working conclusions by the researchers within the tested family of worlds. They are not universal laws of intelligence or discoveries Polaris made without our involvement.

  1. The earlier task left little room for adaptation

    The baseline controller already solved much of the earlier stationary tasks. This limits the task; it does not show that intelligence is unnecessary in every stable environment.

  2. More knowledge does not always mean a longer life

    Several tested additions improved knowledge without a confirmed survival gain.

  3. An action-valuation bottleneck was found

    In the cases studied, a useful trip loses under the internal valuation. This does not rule out other knowledge or policy errors.

  4. A signal is valuable by its silence

    A measured rule is good because it knows when not to doubt. A hard habit costs lives and a third of the income.

  5. A principle must be written once

    We rediscovered each idea separately for every variable. She needs one knowledge with age and trust, not an organ per quantity.

  6. A small effect is hard to detect within one life

    In this test a large effect was detectable within one life, while a small one required many. This depends on the experiment, effect size and criterion; it is not a general limit of self-improvement.

How we check ourselves

Rules of an experiment

  • Predictions with probabilities are frozen before a run and scored afterwards.
  • Ceiling first: how much can be gained at all if the knowledge is handed over ready-made.
  • Confirmation only on fresh worlds that no earlier experiment has seen.
  • Paired comparison on one set of worlds. Interval methods and separate significance checks are specified in each experiment report.
  • Every computation runs under a recorder: input, output, time, memory.
  • Executor and auditor are separate roles.

What did not work

  • A stock of knowledge failed to turn into survival four times in a row.
  • Memory of the exact date of an event does not transfer to other worlds. The line is closed.
  • She finds a planted law, but the gain from it did not replicate on fresh worlds.
  • A rule she composes herself from her own measures is not yet better than ours.
  • Her valuation is still hand-written: seven revisions are seven researcher patches.

Six stages to an independent life

The primary metric is the share of laws planted in the world that the lineage found and used on its own, on unseen worlds, with no researcher edit.

  1. A zoo of laws

    Worlds with planted laws and decoys.

    in progress
  2. Unified knowledge

    One knowledge with age instead of an organ per variable.

    first steps
  3. Value from lived outcomes

    The price of time and search from her lineage's outcomes.

    first experiments
  4. Action as data

    Ways of acting that she composes herself.

    not started
  5. Her own voice

    A truthful report of her own tables.

    not started
  6. A life of her own

    A real life she leads herself.

    not started
real life · tick436

She is alive. Energy 194.

This is the same saved astra1 snapshot, not a currently running process. Specific recovery scenarios have been checked; loss of a result before it is saved and production-binary qualification remain open.

Contact the team

Let’s develop Polaris together

We welcome research and engineering collaboration: joint experiments, independent verification of results and work on the digital organism.

Tell us about your work and what you would like to do together. Email us with questions about the project or collaboration proposals:

polaris@spectorn.ai