Formation Protocol Reaches the Build Threshold

Remove the Privileges / Round Two

The filing cabinets were still standing

We have spent much of the Consequential Formation work asking an uncomfortable question:

What have we declared irrelevant before giving consequence the opportunity to decide?

The first Remove the Privileges research round attacked inherited nouns: gravity, dimensionality, spacetime, boundaries, conventional scientific categories, and our own terminology.

The second round exposed a deeper problem. We had removed many of the nouns while quietly leaving the filing cabinets standing. Disciplinary domains were still deciding which variables were allowed to meet. Reference frames were still deciding what counted as background. Scale was still being treated as a ladder of pre-existing levels. And our proposed formation mechanism contained an embarrassing hole: the distinction being tested simply appeared from somewhere.

That second round changed enough of the picture that we believe it is now time to build. Not everything. The smallest thing that can fail.

Research state

What we think the kernel is

The present working sequence is:

This is not being presented as a new law of nature. It is the current candidate grammar we intend to implement and attack.

The important step for the software work is much narrower:

A system experiences something. A distinction within that experience is shown, not merely asserted, to alter consequence. That distinction is retained in a form that changes what the system recruits, admits, excludes, trusts, resolves, or recomputes when sufficiently similar conditions recur.

Research state

Learning before intelligence

We kept reaching for the word intelligence too early. That encouraged arguments about definitions when the lower mechanism had not yet been built. So we are separating the staircase.

The research team does not unanimously agree about where the word learning becomes fully justified, or whether strong operational intelligence begins at G8 alone or requires G7 plus G8. We are preserving that disagreement.

Software can give us better questions than another round of naming. The first build therefore stops well below the intelligence claim.

Research state

The missing office: proposal

An earlier formation equation assumed a candidate distinction d was simply available for testing. That is not good enough. Where did it come from?

We call this the proposal operator, Pi. Its purpose is not to establish truth. Its purpose is to prevent the current representation from monopolizing the hypothesis space.

A system should be permitted to propose an unfamiliar distinction. That does not entitle the distinction to survive. Proposal and adjudication are different offices.

For the first build, Pi will remain deliberately restricted. Candidate distinctions will come primarily from residuals, contradictions, collapse-twin divergence, provenance gaps, and deliberately available-but-unmodeled channels. Open-ended scientific hypothesis generation can wait.

Research state

Remove the domain

One of our larger methodological corrections is simple to state: domains do not interact. Physical processes do.

Physics, biology, astronomy, geology, chemistry, cognition, and the rest are extraordinarily useful human organizations of knowledge. They are not automatically physical boundaries. Absence from a discipline's conventional variable list is not evidence of physical inconsequence. Neither is the absence of an accepted mechanism.

That does not mean every measurable variable should be retained. A candidate is allowed into testing. Consequence decides whether it stays.

Research state

Remove scale as a ladder

A higher-scale participant does not need to be an ontologically different kind of thing. It may be a set of lower distinctions whose internal differences can safely remain unresolved for the next operation.

If a later operation exposes consequential divergence inside that closure, the distinction must reopen. Scale becomes less like a stack of boxes and more like a changing resolution of consequential sufficiency.

This is deeply related to established coarse-graining, quotient construction, state minimization, bisimulation, and renormalization ideas. We intend to import that mathematics openly.

Attribution

Wade Marr's contribution to this step

Wade Marr's recent closure and composition work helped us solidify ground we had been circling. The useful connection is specific.

Wade's work gives us a sharper formulation of this question: when may distinctions be safely collapsed because the reduced representation remains sufficient for a declared operation?

Consequential Formation asks the complementary developmental question: when does a previously collapsed distinction have to be reopened because consequence has made it necessary again?

We are not claiming that Wade's framework and Consequential Formation have been proved formally equivalent. They have not. Nor are we claiming CF invented quotienting, coarse-graining, bisimulation, or partition refinement.

The value of Wade's work here was that it helped make the boundary between safe compression and consequential reopening much clearer. That materially influenced the direction of the Formation Kernel.

Research state

The bounded mathematics underneath it

For a finite state set and finite operation family, splitting states when a discriminator reveals consequential divergence and merging states when operation sufficiency is established is not mysterious new mathematics. It belongs to established families including state minimization, partition refinement, and bisimulation.

The CF-specific engineering question begins when the partition itself becomes developmental state: candidate distinctions can expand; operations can change; every split and merge has provenance; a previous compression can reopen locally; retention has computational and energetic cost; and formation must justify that cost through later utility.

The bounded mathematics anchors the machine. It does not diminish the experiment.

Research state

Formation must pay rent

Reasoning is expensive. If a system pays to discover something consequential today and then repeatedly pays nearly the same price to reconstruct it tomorrow, it received a poor return on that expenditure. Formation should turn expensive past reasoning into cheap future selectivity.

The formed system may initially be more expensive because it has to investigate, verify, and write. The question is whether recurrence changes the economics while preserving or improving verified consequence.

• Touch less state
• Perform less search
• Invoke less expensive reasoning
• Respond faster
• Consume less compute or energy

If not, formation may simply be an elaborate memory system. That is a failure we want to be able to see.

Research state

The first machine

We are calling the first implementation the Formation Kernel for now. It will be deliberately bounded. Two otherwise identical agents will share the same frozen host capability, initial representation, environment, and task family. One will repeatedly reconstruct. The other may form.

• Raw receipts
• Executable formed state
• A causal witness engine
• Typed formation writes
• Operation-sufficiency and recompression
• Provenance-preserving reopening
• An append-only formation ledger
• Resource accounting
• A restricted proposal operator

The environment will include hidden consequential distinctions, interesting-looking distinctions that do not matter, and a later distribution change that makes one previously safe compression insufficient. The formed system must learn the useful difference without worshipping the useless ones. Then we will break its compression.

Research state

The causal test that matters

If a formation is genuinely causal, we should be able to move it. After the treatment agent develops a verified narrow advantage, transplant the formed state into an otherwise identical fresh host, predict the specific behavioral difference that should move with it, and test. Then ablate the formation and test whether the advantage disappears.

If the behavior does not move with the formation, the developmental claim has failed. That causal test is part of the first build, not a future embellishment.

Research state

What would kill this direction

We should demote or abandon the proposed mechanism if the bounded implementation shows that:

• Formed distinctions do not causally change later recruitment or admissibility.
• Transplant and ablation fail to move the predicted behavior.
• Maintaining formation costs as much as or more than repeated reconstruction without compensating performance advantage.
• Compression cannot reopen locally when the operation changes.
• The proposal mechanism produces unusable candidate explosion.
• Ordinary transcript, RAG, static rules, or conventional state-minimization baselines reproduce the claimed benefits at equal or lower cost.

A failed result remains publishable. The failure would tell us which arrow in the chain does not survive.

Research state

Why publish now?

Waiting until a result looks finished produces a false history of research. The interesting part is often the path that died. This release marks the state before the Formation Kernel is built.

The old failure stays failed. The new result stands beside it. And today's confidence is allowed to become tomorrow's mistake.

Current research checksum

Trace the relations until consequence either survives or dies

For the Formation Kernel, add one more line:

That is the machine we are building next.