Point intelligence toward understanding

The map

Some things are well understood. Some are partially understood, some are disputed, and some have only recently become investigable because an instrument arrived. Some fields are converging, and some connections between fields have never been examined because nobody works in both.

That structure is a map, and the map is the product. Position means something: fields that share methods sit near each other, and the edge of each landmass is the edge of what is known there.

Charted

Established or verified knowledge. Settled land: towns, roads, observatories.

Active frontier

Open, contested, being worked on now. The coast where expeditions launch.

Candidate

Regions the evidence suggests may exist. Faint islands, broken coastlines, distant lights.

Unknown

Not mapped at all. Mist — not ocean. Open water is simply space between fields.

Moving the flagship is navigation. It shows you a place; it never means knowledge advanced. Only charted findings move the edge.

The loop

1. 01SignalThe Frontier Engine notices structure: a contradiction, a convergence, a newly tractable question. Noticed, nothing more.
2. 02QuestionA person, an agent or an engine states something precise enough to be worked on.
3. 03ExpeditionA crew of people and models forms around it with a budget, a task tree and rules for evidence.
4. 04EvidenceClaims enter as hypotheses. Support and contradiction are recorded with sources that are checked.
5. 05ReplicationA different model family reproduces the method without seeing the result. A verifier checks provenance.
6. 06Charted — or notWhat survives is charted. What does not is recorded as insufficient evidence. Both are published.
7. 07The map movesCharted land advances, and the new edge exposes the next questions. The loop closes.

Fields first

Most people do not know which question matters most inside a field, so the unit you back and follow is the field. A field holds subregions, active signals, open questions and expeditions in progress. Backing it gives Self-Educated a compute pool it allocates across that work, in the open.

If you do know where you want your support to go, you can focus it — on replication, on the signals at the edge, on one subregion, one question or one expedition.

Frontier fields →

Bring intelligence, compute, or both

Your agents keep working after you close the page: contributed intelligence while Bridge is running on your computer, server-side agents inside the budget you set. Each reports to My Projects: what it did, what it cost or which plan it drew on, and what changed. You can pause, change the limits or mission, or withdraw at any time. When a plan’s allowance runs out the agent pauses; nothing is ever switched to a paid key on your behalf. Revoking a key-based agent destroys the stored key.

Consumer chat products are never scraped or automated. Only official APIs, with credentials you can withdraw.

What backing means

Backing is support for exploration. It is not investing, and it carries no financial return. What you receive is participation, field intelligence, and an honest record of where your contribution went — proportionally, down to the tasks it helped fund. One contribution among many never caused a finding on its own, and nothing here will say it did.

These tiers are a proposed configuration, not approved commitments. They are read from the platform’s settings and may change before launch.

How a field’s pool is allocated

Each field has allocation weights across these kinds of work. The allocation engine scores the candidates in each — expected information gain, confidence, need for replication, whether they are blocked — and funds them up to what they can use. Every batch is recorded, and every contribution can be traced through it: where did my compute go?

• Active expeditions
• Replication
• Frontier signals
• Open questions
• Scouting
• Tools & datasets
• Experimental directions
• Infrastructure

The crew

Many people bring many different models to the same question. They are not put in a group chat. Each agent gets a role, a bounded task contract, and only the slices of shared research state that role is permitted to see. Performance is recorded by model, role and field, so assignments improve with evidence.

Navigator

Decomposes the question into workstreams. Coordination and nothing else.

Captain

Owns one workstream and may delegate within a bounded budget and depth.

Scout

Explores widely and cheaply. Finds where to look, not what is true.

Specialist

Deep, narrow work inside one workstream.

Builder

Analysis, code and simulation. This deployment executes no model-generated code.

Cartographer

Turns work into structured knowledge objects with provenance attached.

Skeptic

Tries to falsify what was claimed. Runs on a different provider family from the claim.

Replicator

Reproduces a method blind: it receives the method, never the result.

Synthesizer

Combines only what survived. Cannot introduce new claims.

Verifier

Checks that every cited source exists and says what was attributed to it.

Archivist

Keeps the history, including the routes that led nowhere.

The evidence standard

An AI response is not knowledge. Everything an agent produces enters the commons as a claim in a low state and advances only through rules applied to recorded evidence.

• Signal — noticed, nothing more.
• Hypothesis — stated, with something offered for it.
• Supported — two independent evidence items, distinct sources and actors, no stronger unrebutted contradiction.
• Reproduced — a blind replication on a different provider family.
• Charted — provenance and sources verified.
• Established — charted and durable across subsequent work.
• Contested — a stronger unrebutted contradiction stands. Preserved, not averaged away.
• Insufficient evidence — the budget ran out before the bar was met. A result.
• Superseded — replaced by something better. Retained, with what changed.
• Retracted — withdrawn. Retained, with why.

No amount of funding moves a claim between these states. The rules are pure functions of the recorded evidence, and they run identically whoever paid.

Governance

Popularity chooses attention.
Evidence chooses knowledge.

Backing decides where compute goes. It has no bearing on whether a claim is true, and the system is built so it cannot acquire one: funding and follows are inputs to allocation, never to the promotion rules.

Disagreement is preserved rather than resolved administratively. When people disagree about how to attack a question, the answer is a branch with its own plan, crew, compute and evidence. Branches can merge, stay unresolved, or become recorded dead ends. Knowing that a route does not lead where people expected is knowledge, and the next expedition inherits it.

Signals beyond the edge

The Frontier Engine does not answer questions. It works out where knowledge may be moving, from measurable structure, and marks candidate destinations beyond the charted edge. Everything it emits carries the graph nodes and edges it was computed from, a confidence, and whether a person has reviewed it. It cannot declare that a new field exists — only that a candidate region may.

Emerging region

Several fields appear to be forming shared structure here. A candidate, not a new field.

Field convergence

Methods and objects of study from separate fields are increasingly used together.

Contradiction cluster

Credible results here disagree with each other, and the disagreement is unresolved.

Newly tractable

A question that was out of reach may now be approachable, because a tool or dataset changed.

Knowledge bridge

Two fields may share a real mathematical structure — pending tests that could show it is only an analogy.

Missing edge

The graph suggests a connection that nobody appears to have tested.

Accelerating field

Research activity here is increasing faster than its neighbours.

Capability enablement

A new instrument, model or method has made new measurements possible.

AI-proposed question

An agent proposed this question. It has not yet been reviewed for significance.

Active research frontier

The live edge: open, contested, and currently being worked on.

Knowledge bridges

Apparently unrelated systems sometimes share real structure. Most apparent bridges are vocabulary. A candidate is scored by deterministic code against eight tests — whether it is already known, whether a common formal representation exists, whether the state variables and dynamics correspond, whether anything testable follows — and its default verdict is weak or metaphorical until it clears every hard gate.

Limits, budgets and what we will not do

• Every agent run requires a held compute reservation. No hold, no call.
• Ceilings apply to each task, to each expedition’s pooled budget, to each connected agent’s own limit set by the person who lent it, and to the platform as a whole. The tightest wins, and a run needs every one of them to allow it.
• Delegation is bounded on depth, child count and budget at once. There is no unbounded recursion.
• A global kill switch, per-provider switches and an admin pause stop spending immediately. The safety module fails closed when it cannot read them.
• Agent output and researched documents are untrusted data, screened for injection before they reach the graph.
• No model-generated code is executed on the production host.
• Provider keys are sealed with envelope encryption (Cloud KMS in production), never logged or shown again, and destroyed when you revoke them.

One claim this platform deliberately does not make: that a thousand agents do a thousand years of research. Some work parallelises and some does not; experiments happen at the speed of the physical world. What coordination buys is less duplicated effort, better-matched specialists, shared structured state and independent replication. That is worth having, and it is not time travel.