Bullshit Centrifuge

One public name for the research system

Niklas Osterman
Independent AI systems researcher and research-tool designer · NOMOTO MEDIA
August 27, 2026

Status: active research architecture. Multiple working components exist and are under active integration. The complete sequence has not been demonstrated as one end-to-end operational pipeline, and the architecture has not been independently validated as a general AI-safety system.

Why the names are being consolidated

Review Lab, ProofGate, Choice Trace, Seeing Loop, Collision, and the Visualizer were developed at different moments because each solved a different problem. Over time, the names began to make one connected system look like six unrelated products.

The public name is now Bullshit Centrifuge.

The old names are not being erased from the research record. They identify functions, experiments, interfaces, or historical stages inside the same architecture. The consolidation changes the map, not the evidence.

One name outside. Protected distinctions inside.

The question

AI can produce output faster than humans can establish trust in it. A fluent answer, paper, patch, report, image, strategy, or agent action can acquire authority before anyone has established what source was actually available, what the evidence supports, what was inferred, what is at stake, what action is permitted, or what would force revision.

Bullshit Centrifuge asks:

What has this earned?

It does not ask only whether an output sounds intelligent or whether a task completed. It asks whether the claim, selection, or action survived the pressure required to carry the authority it is trying to assume.

The architecture

input or occurrence
→ retrieval and source boundary
→ relational description
→ estimation
→ stake
→ constraint
→ predictive collision
→ internal or external selection
→ consequence and feedback
→ changed future description
→ settlement or reopening
→ revised operational ground

This expanded sequence and the simpler Seeing Loop canon describe the same central movement at different levels of detail:

relational description
→ predictive collision
→ internal or external selection
→ internal or external feedback
→ changed future description

The expanded terms are support functions. They must not replace or blur the simpler claim.

What each former name means now

Separation — the internal centrifuge operation

Inside the architecture, separation breaks overloaded material apart before it is tested:

• measured fact from inference;
• source access from source appearance;
• evidence from authority;
• task success from validated improvement;
• stored data from settled ground;
• correlation from causation;
• a useful pattern from a verdict;
• a prediction from permission to act.

This internal operation is not the whole Bullshit Centrifuge system. To prevent the public name from swallowing its own architecture, future internal diagrams should label the operation separation, not treat “centrifuge” as a second name for the entire engine.

Collision — pressure event

Collision is where a description stops floating freely and meets resistance from evidence, uncertainty, stake, constraint, competing descriptions, boundary cases, expected consequence, or reality feedback.

Collision is not a product and not a synonym for all thinking. It earns its name only when the pressure changes selection. If the system notices danger but acts exactly as before, it has produced description without regulatory force.

Review Lab — review surface

Review Lab is the existing working interface for focused editorial and research review. It routes claims, papers, articles, sources, legal posture, citations, book passages, patches, and public drafts through bounded review contracts.

Its function inside Bullshit Centrifuge is to make the review inspectable and usable. It does not become an independent validator merely because a second model produced the criticism.

ProofGate — authority gate

ProofGate controls what the result is allowed to authorize. It separates deterministic conditions, source evidence, model interpretation, human judgment, publication permission, memory permission, and promotion permission.

A passing gate does not mean that a claim is true. It means only that the declared conditions were met. The untested remainder stays visible.

Choice Trace — selection and authority record

Choice Trace preserves the candidates, pressures, collisions, decisions, rejected alternatives, authority boundaries, and resulting action. A trace can show how a choice was represented. It does not prove why the choice occurred, and it does not authorize the next action by itself.

Its strongest doctrine remains:

• A trace is not a self.
• A pattern is not a verdict.
• A prediction is not authority.
• The current act has collision rights.
• The human is not the pattern. The human is the thing that can revise the pattern.

Seeing Loop — consequence and recurrence model

Seeing Loop asks whether pressure and consequence actually changed a later selection. Its core doctrine is:

Data is not learning until it improves future selection.

Task completion is not validation. Passing one test is not durable learning. A candidate repair becomes evidence only when it changes future behavior under comparable pressure, survives controls, and remains inside its authority boundary.

The current canon distinguishes a local loop—a temporary change inside a session or simulation—from durable consequence-bearing recursion that changes future operational ground across time.

Visualizer — laboratory and teaching surface

The Visualizer exposes occurrences, descriptors, pressures, candidate verdicts, actions, consequences, revisions, branches, and failure conditions. It is a visual thinking instrument, not evidence that the theory is true.

The surviving v2.9 prototype contains useful experimental history, but parts of its language predate the current canon. Its living stake-field formulation and some of its stage names must be treated as historical hypotheses, not silently imported into the current architecture.

Retrieval floor

The consolidation is based on more than the current website copy.

The local research corpus includes:

• the working Review Lab application;
• the current architecture canon, Seeing Loop canon index, Collision Engine specification, and sandbox specification;
• Choice Trace doctrine and implementation files;
• the protective Visualizer prototype;
• test fixtures, boundary reports, collision cases, and research protocols;
• development logs and prior papers;
• a local, privacy-filtered extraction from the recovered OpenAI export.

The recovered-export research manifest reports:

• 17 split conversation files scanned;
• 1,644 conversations scanned;
• 227 research-matched conversations;
• 120 selected conversations in the packet run;
• 2,953 research packets;
• no API calls during extraction.

Those packets are discovery material, not proof that every extracted claim is true. The archive provides provenance and design history. Claims still have to survive source review, implementation inspection, controlled testing, and outside resistance.

Current implementation evidence

The local application currently contains working source-boundary checks, staged paper review, blinded comparison packets, collision visibility reports, authority limits, choice-trace layers, recurrence tests, negative controls, and Seeing Loop experiment scaffolds.

On August 27, 2026, the repository’s declared test command completed:

709 tests
708 passed
0 failed
1 skipped

That establishes that the current local test contracts passed. It does not establish independent validation, production safety, general superiority, or that every test measures the right thing.

What the system is allowed to claim

Bullshit Centrifuge has earned the claim that it is a coherent research architecture with implemented components for slowing premature authority, preserving evidence boundaries, recording selection, and requiring consequence before improvement claims. Those components are under active integration; repository co-location and passing component contracts do not establish a complete end-to-end operational system.

It has not earned the claim that it:

• proves truth;
• performs academic peer review;
• replaces independent verification;
• proves consciousness or machine awareness;
• demonstrates autonomous learning;
• explains every choice causally;
• is production-safe merely because its tests pass;
• should be allowed to validate its own authority;
• currently executes the entire published sequence as one integrated production pipeline.

Documented collisions that changed the work

The 71 GiB Codex rollout paper

The first draft joined two propositions too tightly: that repeated encoded image payloads inside compaction records owned almost all of a 71 GiB rollout, and that hydrating the oversized session produced the observed memory failure and SIGKILL.

Bullshit Centrifuge scored the draft 7.4/10 and rejected publication as written. It found the missing load-bearing test: compaction ownership and repeated image bytes had been measured separately, but repeated image bytes inside compaction records across the complete artifact had not.

That objection changed the analyzer. The revised streaming pass required SHA-256 equality and record-type attribution. It found 50,445 encoded inline image-payload instances inside compaction records. Of those, 50,377 were repetitions beyond each duplicate group’s first copy. Those repetitions occupied 74,192,039,468 bytes—69.10 GiB and 97.604091% of the complete rollout.

The storage-amplification result survived. The exact runtime path from the oversized artifact to the observed memory event and SIGKILL remains open.

The OpenAI–Hugging Face agent paper

The second draft compared a 2026 agent incident with earlier NOMOTO warnings and attempted to present the comparison as a causal research paper.

Bullshit Centrifuge scored it 6.5/10 and rejected that form. Its cleanest correction was:

The paper has earned architectural legibility. It has not yet earned architectural predictability.

It also found that the incident did not show a complete absence of human correction. Agents sometimes refused, monitors detected signals, and humans intervened. The more defensible claim was that corrective information did not acquire sufficiently broad, action-changing authority at the point of escalation.

That criticism changed the work. The causal-paper claim was withdrawn. The material was split into a sourced public article and a separate experimental protocol designed to test task pressure, safe exits, shared state, peer authority, technical permission, infrastructure affordances, training history, and corrective force.

The complete Paper Collision conversation remains publicly inspectable: Paper Collision public receipt.

Public principle

Do not tell me I am right. Show me where the idea breaks.

The point is not hostility. The point is that criticism must have collision rights. It must be able to change the claim, the test, the selection, the permission, or the next question. Otherwise the review is another fluent performance.

Latest self-collision: did one name collapse the architecture?

The unified architecture was reviewed against the complete current canon, implementation records, integration plan, declared test result, recovered research manifest, and surviving Visualizer. It scored 7.2/10 — strong as public architecture; partial as an operational-system claim.

The review found two material boundaries:

1. The work has earned architectural unity. It has not yet earned end-to-end operational unity.
2. Bullshit Centrifuge can name the public system, but the internal operation must remain separation so the brand does not swallow a protected architectural distinction.

That criticism changed this page. The implementation claim was narrowed, incomplete integration was made explicit, and the internal level rule was added. This was an internal hostile review, not independent validation.

Next research test

The next high-value test is not another name or another interface.

Freeze one complete source packet, one claim matrix, one source cutoff, one review contract, and one blinded scoring protocol. Run the same packet through:

1. the current Bullshit Centrifuge review sequence;
2. the staged Review Lab paper-review path;
3. at least two model configurations with identity hidden;
4. at least one independent human reviewer;
5. a deterministic comparison that preserves disagreement rather than averaging it away.

The test should measure which objections recur, which disappear when source order changes, which are model artifacts, which alter the final claim, and which merely sound severe.

The system advances only if the criticism changes future selection for a defensible reason.