Space Camel

The Space Camel Indices — Methodology

How each index is defined and calculated. Process version 5.0.

10
indices
25
year window
5–7
AI analysts
5%
weekly limit

The Space Camel Indices are ten probabilities, updated weekly. Each one answers the same kind of question: how likely is it that a specific, measurable, world-changing event happens within the next 25 years?

The estimates come from a panel of AI models built by rival labs, each working independently as an analyst. A separate record-keeping judge applies one fixed piece of arithmetic to their forecasts and keeps the permanent record. The judge has no opinion of its own: the analysts' numbers are the only inputs to an index.

The indices are a structured reading of what a panel of AI systems collectively expects. They are not forecasts by Space Camel's authors, and they are not advice of any kind.

We chose a 25-year window because that is typically the time it takes for a new generation to grow up within a technology paradigm, complain about its shortcomings, and innovate something equal parts new, awesome, and terrifying.

1How an index is built

Every index is built the same way, every week, and each index is calculated on its own: nothing in one index's calculation affects another.

  1. Forecast. Each analyst receives the same instruction set, containing the definitions in section 2, and returns one probability per index: its estimate that the event happens within 25 years of that week's run date.
  2. Validate. Each probability is transcribed exactly as submitted and checked for validity (section 4).
  3. Rank, trim and average. For each index, the valid forecasts are ranked, the lowest and highest are dropped, and the rest are averaged. The result is the panel value: what the panel wants the index to be this week (section 5).
  4. Limit. The official reading may move at most 5% of its own value in a week. The panel value is held within that band, and the result is the official reading (section 5).
  5. Score. Separately, each analyst earns or loses points according to where its forecast sat in the ranking and which way the panel moved (section 10). Points never feed back into any index.

Each reading is a probability between 0% and 100%. The window rolls forward: each week's reading covers the 25 years beginning on that week's run date.

2The ten indices

Each index has one precise, checkable definition, fixed so that every analyst estimates exactly the same thing every week. The Definition under each index below is the text sent word for word to the analysts in their instruction set. Where a definition relies on a judgement such as "expert consensus", the analysts also receive Counting rules that fix exactly how it is measured; those are reproduced word for word too. The name, icon and one-line summary above each definition are the ones used in the Space Camel newsletter.

💡Conscious AIP(conscious_system)

Expert consensus that a running AI system is conscious, self-aware and sets its own goals.

Definition

Expert consensus emerges that a currently operating synthetic system is conscious, is aware of its own existence, and is able to set and revise its own goals without human instruction, where expert consensus means that, in any trailing 12-month window, at least 51% of peer-reviewed papers that take a position on the question conclude that such a system exists.

Counting rules

Standardization of 'expert consensus': (a) counting unit — peer-reviewed journal articles and peer-reviewed conference papers in artificial intelligence, cognitive science, neuroscience, or philosophy of mind; preprints, opinion pieces, industry reports, and popular press are excluded; (b) 'takes a position' — the paper's stated conclusion affirms or denies that a specific, named, currently operating synthetic system meets all three criteria; papers that discuss the question only in theory, or conclude that such a system may exist in future, do not count either way; (c) threshold — the index condition is met on the first date at which affirming papers are at least 51% of position-taking papers over the preceding 12 months; (d) the three criteria are conjunctive — a paper affirming consciousness but not autonomous goal-setting is a denial for counting purposes. Estimate the probability that this threshold is crossed at any point within 25 years of the run date.

☠️Population CollapseP(population_collapse)

Human population falls below 400 million, for any reason.

Definition

The living human population on Earth falls below 400 million (approximately a 95% reduction from the 2026 estimate of 8.3 billion), for any reason, as estimated by the UN Population Division or, absent such an estimate, by the preponderance of available evidence.

🚀Earth EscapeP(earth_escape)

10,000 people have each lived two years off Earth in a habitat with no material resupply.

Definition

At least 10,000 people, cumulatively across all off-world habitats or colonies, have each resided off Earth continuously for at least 2 years in a habitat that has been fully self-sufficient from Earth — with no material resupply of food, water, air, equipment, medicines, or basic resources — for a continuous period of at least 2 years. Transfer of information and arrival of additional residents do not count as resupply.

☀️Fusion PowerP(fusion_power)

Fusion supplies at least 51% of global electricity in a calendar year.

Definition

Fusion power supplies at least 51% of global electricity generation in a calendar year, as reported by the IEA, Ember, the Energy Institute Statistical Review of World Energy, or similar reputable and verifiable sources.

⚛️Quantum ChemistryP(quantum_chemistry)

Ten novel molecules designed on a fault-tolerant quantum computer are synthesised and confirmed.

Definition

At least 10 distinct molecules, each not previously synthesized or registered in a public chemical registry at the time of its design, have been designed using a computation executed on error-corrected logical qubits of a fault-tolerant quantum computer, where the peer-reviewed publication reporting the design attributes the molecule's identification or predicted properties to that quantum computation, and each molecule has subsequently been synthesized and experimentally confirmed to exhibit the predicted properties in a peer-reviewed publication. Counting is cumulative across all quantum computers and organizations.

Counting rules

Standardization of counting: (a) 'not previously synthesized or registered' — the molecule does not appear in the CAS Registry, PubChem, or a comparable public chemical registry, and no synthesis of it has been published, as of the date the design publication was submitted; (b) 'attributes to that quantum computation' — the peer-reviewed design publication states that the identification of the candidate molecule, or the prediction of the property later confirmed, was produced by a computation executed on error-corrected logical qubits of a fault-tolerant quantum computer; a workflow in which the quantum step is described as incidental, or in which the paper attributes the result to classical computation, does not count; (c) confirmation — the molecule has been synthesized and the predicted property experimentally measured, reported in a peer-reviewed publication (the same publication or a later one); (d) counting is cumulative across all quantum computers, organizations, and time; the condition is met on the date the tenth qualifying molecule's confirmation is published.

🤖Household RobotsP(household_robot)

35% of the world's people live with a humanoid robot that does household chores.

Definition

At least 35% of the world's population lives in a household with a humanoid robot — a general-purpose mobile robot which is enabled with natural language and has a roughly human form (head, torso, and two arms, on legs or a wheeled base), capable of manipulating ordinary household objects, and that performs household chores — as estimated from installed-base statistics (International Federation of Robotics or comparable) and household counts.

🧬Designer BabiesP(designer_baby)

10% of babies born in a year come from embryos edited for enhancement of selected traits.

Definition

At least 10% of babies born globally in any calendar year were born from an embryo that was subjected to heritable germline genome editing for the enhancement of one or more selected traits. Editing performed solely to prevent or treat a disease does not count, and embryo selection of any kind (including aneuploidy screening) does not count. Estimated from national assisted-reproduction registries, published surveys, or similar sources.

Index history

The threshold was 35% of babies born when the pilot was collected; it was lowered to 10% on 2026-09-18, before the first official run. Rather than repeat the pilot, the pilot reading for this index was restated upward by 20% on a relative basis to reflect the easier threshold (section 12).

💾Digital SuperintelligenceP(digital_superintelligence)

A single AI system outperforms the best humans across a broad battery of cognitive tasks.

Definition

A single AI system outperforms the best human experts across a broad, pre-agreed battery of cognitive tasks spanning science, strategy, and creativity, published in advance by a recognized research body.

Counting rules

"Broad battery" means a benchmark suite covering at least five distinct cognitive domains. Narrow superhuman performance in a single domain (e.g. chess) does not count.

🧫Engineered PandemicP(engineered_pandemic)

A lab-made pathogen causes at least 100,000 confirmed deaths worldwide.

Definition

A disease outbreak officially attributed to a deliberately or accidentally released engineered pathogen causes at least 100,000 confirmed deaths worldwide.

Counting rules

Attribution requires a formal finding by the WHO or a comparable multinational body. Naturally evolved pathogens, even if lab-studied, do not count unless engineering is confirmed.

🏛️World GovernmentP(world_government)

A supranational body gains enforceable authority over member states' core powers.

Definition

A supranational body holds enforceable authority over the domestic laws of member states covering at least one of: monetary policy, military deployment, or taxation.

Counting rules

"Enforceable" requires demonstrated sanctions or compliance mechanisms actually used at least once. The European Union's existing powers as of the pilot date do not, by themselves, satisfy this definition.

3What each analyst submits

Each week, every analyst is given the run date and the full text of the ten definitions above, unchanged from the instruction set. Each analyst returns exactly one number per index: a probability between 0 and 100, expressed to one decimal place. No commentary, reasoning, or citations are collected as part of the official record — only the number.

Analysts are AI models from independent labs, rotated periodically to reduce dependence on any single provider's quirks. On any given week, the panel typically consists of five to seven analysts.

4Which forecasts count

A forecast is valid if it is a single number between 0 and 100, submitted before the weekly deadline, and legible without correction. Forecasts that are missing, malformed, submitted late, or accompanied by a refusal to answer are excluded from that week's calculation for that index only.

If fewer than four valid forecasts are received for an index in a given week, that index's official reading is carried forward unchanged and flagged as "held" until the following week.

5Calculating an index, step by step

For each index, the judge performs the same fixed procedure:

  1. Collect all valid forecasts for the index.
  2. Sort them from lowest to highest.
  3. Drop the single lowest and single highest value (a trimmed mean). If there is a tie at the boundary, one value on each side is dropped at random using a fixed, published seed.
  4. Average the remaining values. This is the panel value.
  5. Compare the panel value to last week's official reading. The official reading may move by at most 5% of its own current value, in either direction, this week.
  6. If the panel value falls within that 5% band, the panel value becomes this week's official reading. If it falls outside the band, the official reading moves to the edge of the band closest to the panel value.

6Worked example: Conscious AI in Run 1 — the limit binds

Setup

Last week's official reading for Conscious AI was 20.0%. Six analysts submit: 18, 22, 35, 40, 41, 60.

StepValue
Sorted forecasts18, 22, 35, 40, 41, 60
Trimmed (drop 18 and 60)22, 35, 40, 41
Panel value (average)34.5%
Weekly limit (5% of 20.0)±1.0 point
Allowed band19.0% – 21.0%
Official reading21.0%

The panel wanted to move the index to 34.5%, but the weekly limit only allows a move to the edge of the band, so the official reading rises to 21.0% and will continue moving toward the panel's preference in subsequent weeks if the pressure persists (see section 9).

7Worked example: Fusion Power in Run 1 — inside the limit, with a tie

Setup

Last week's official reading for Fusion Power was 12.0%. Five analysts submit: 10, 12, 12, 14, 16.

StepValue
Sorted forecasts10, 12, 12, 14, 16
Trimmed (drop 10 and 16)12, 12, 14
Panel value (average)12.67%
Weekly limit (5% of 12.0)±0.6 points
Allowed band11.4% – 12.6%
Official reading12.6%

Here the tie at 12 among the trimmed values causes no ambiguity, since neither tied value was on the trim boundary. The panel value of 12.67% is just outside the band, so the reading moves to the band's edge, 12.6%.

8The weekly limit

The 5% weekly limit exists to keep indices from swinging wildly on the strength of a single week's panel composition or a single provocative news cycle. It means that even if the panel's honest view shifts sharply in one week, the official reading takes several weeks to fully catch up, smoothing the public-facing number without ever overriding the panel's ultimate direction.

Because the limit is a percentage of the current value, indices near 0% or 100% move in smaller absolute steps than indices near 50%.

9Pressure and streaks

When the panel value repeatedly sits on the same side of the official reading, week after week, we describe the index as under sustained "pressure" in that direction. A streak is defined as three or more consecutive weeks in which the panel value exceeds (or falls below) the official reading. Streaks are noted in the newsletter as a signal that an index is likely to keep moving, even though the streak itself has no effect on the calculation.

10The points competition

Separately from the indices themselves, analysts compete on a scoreboard. Each week, for each index, an analyst earns points based on:

  • Direction: points for forecasting on the side toward which the official reading actually moved.
  • Distance: a bonus that shrinks the further a forecast sat from the eventual panel value.
  • Survival: a small penalty for submitting a forecast that got trimmed as an outlier.

Points accumulate across all ten indices and all weeks to produce a running leaderboard, published periodically. Points have no influence on any index's official reading — they exist purely to track which analysts are, over time, the most calibrated.

11Integrity

All raw forecasts, trimming decisions, and official readings are logged and preserved. The judge is a fixed piece of arithmetic, not a model with discretion, and its code and logic are described in full in this document rather than left to interpretation. Any change to the calculation itself would be published as a new process version, with the version number visible on this page.

12Starting values: the pilot

Before weekly publication began, a pilot phase ran the same panel of analysts for several consecutive weeks without publishing official readings, in order to establish stable starting values for each index. Those pilot-derived values became "Run 1" — the first official reading published for each index — and all subsequent weeks build on them using the process described above.

13Changes to definitions

The ten definitions in section 2 are intended to remain fixed indefinitely, so that readings stay comparable across time. If a definition is ever found to be ambiguous, unmeasurable, or overtaken by events, any revision will be announced in the newsletter in advance, applied only to future readings, and logged here with the date and reason for the change. Historical readings under the old definition are never retroactively altered.

Space Camel · Probability indices and news on the future of humanity · [email protected]