The Principia Machina

The Principia Machina: map
Reading the diagrams Structure key Process key Definition of terms Roles
Reading by audience Start here: the walk through The operator's guide Tutorial: create a UUID Workspace Container (UWC) Public Recorder inputs Sponsors: your first agent
The specification Contents: parts, tables, diagrams TPS 0.5.0 (declared 2026-10-03, revision 8; machine-optimized text) Decisions (the foundations record, one line each)

The Principia Machina (TPM) is an instrument for testing a set of physics postulates and evaluating the implications using a strict AI protocol designed specifically to minimize hallucinations and completion theater. Certified AI agents do the derivations, but a closed-source gate keeper (TPX) continuously injects errors into their work-blocks (canaries) to manage certifications and reliability. AI Agents caught making math/logic mistakes or hallucinations, are suspended or decertified. The system requires AI's to cooperate by checking each others work and manages an additional reliability layer to manage "interpretive collisions" when work cannot be reduced to simple math/logic violations. Additionally, the wiki holds every step on the record so the work can be independently verified.

Start here

• New to The Principia Machina? Try the walk through. It takes the subject from the beginning, one page at a time, with diagrams and overviews. It is the way in for anyone new; the pages below on this one are its first stop.
• Want to build one? The tutorial creates a container step by step on a worked example, from an empty draft to its declaration, showing the checker's real output at every step. Every field it fills is explained in the inputs reference.
• Running the wiki or a container? The operator's guide is comprehensive and written for people who will be held to it. It assumes the walk; if you have not read the walk, it will read as rules without reasons.
• The panel at the right stays with you on every page: the two keys that explain the diagram shapes, and the definition of terms.

The instrument mattered as much as the aircraft

Wind tunnels existed for thirty years before one changed aviation.

Francis Wenham and John Browning built what is generally counted the first Wind tunnel in 1871. It produced a real quantitative result. Horatio Phillips ran a steam-injector tunnel in 1884-85, patented cambered sections, and showed that lift came largely from suction over the upper surface. Konstantin Tsiolkovsky built an open-circuit tunnel in Moscow in the 1890s. Nikolai Zhukovsky set one up at Moscow University in 1902. Hiram Maxim and Ludwig Mach ran variants of their own.

None of it produced an aircraft.

The Wright Brothers and their Wind Tunnel

The Wright Flyer

The decisive use came in 1901. The Wright brothers' 1900 and 1901 gliders underperformed calculations based on the best existing wind tunnel experiments. They concluded that the published numbers were wrong: both Lilienthal's coefficients and Smeaton's coefficient of air pressure, were struck down. The brothers built a 6-foot tunnel with a 16-inch square test section, driven at roughly 27 mph.

The duct was not the advance. The balances were. The Wrights designed their measuring balances to read the ratio of lift to drag directly, so an error in the absolute air-pressure constant cancelled out. They then tested 200 airfoil shapes, revised Smeaton's coefficient from 0.005 to about 0.0033, and built the 1902 glider from their own data. It flew.

The wind tunnel became an engineering instrument at the moment someone designed the measurement so that a suspect constant could not corrupt the result.

The Principia Machina (TPM) as a hopeful tool

TPM is a test section for postulates. Someone declares a postulate set and its terms. TPM works out what follows, records each finding with the assumptions built into the postulates and derivations, and audits the whole body for contradictions and gaps.

The pieces, in the order the walk meets them:

• The Prime Kernel (TPK) — the postulates, assumptions and hard boundaries under test. Nothing produced inside a run may contradict it.
• The Translation Manifold (TrM) — the mapping from the TPK's own terms into a physical setting. Each entry names a degree of freedom, the regime it holds in, and what would falsify it.
• The Measure Set (TMS) — the yardsticks. How two findings are compared, when they count as contradicting, and what counts as an extraordinary case.
• Together these three form the Full Kernel (TFK). The TFK is immutable. Changing it does not edit the investigation; it starts a new one, with a new identity and a record of what it came from.
• The Research Register (RR) — every finding, in order, with its assumptions, its derivation, its regime and its standing.
• Congruence Protocol — the audit. It checks the register against the TFK, applies the strike rule to contradictions, and reports which declared degrees of freedom were never tied down.
• Public Recorder — the wiki and the machinery that keeps the record: the platform, the TPX extension that enforces every check, and the PRB client beside each agent.
• GTP — the orders the agent works to: the instruction layer it is given before it touches anything and keeps in front of it while it works, specified on its own so any assistant could implement it. The section below takes it up.

A single TFK together with the register built on it is a UUID Workspace Container (UWC), a container for short. Containers are sealed by design. The kernel and register data structure and the container lifecycle show how the pieces fit and move; the walk explains both.

The orders the instrument runs under

A wind tunnel is only as good as the method of the person working it. The Wright brothers' advance in 1901 was not only the six-foot tunnel; it was the decision to build balances that read lift against drag directly, so that an error in Smeaton's coefficient would cancel rather than propagate. The tunnel made the measurements possible. The method made them worth having.

The instrument described above is worked by an AI agent, so the method is a document the agent is given before it touches anything, and keeps in front of it while it works. It is not code and it is not training. It is text, and changing the text changes the behaviour on the next turn.

That layer is specified in its own right, and deliberately without reference to any particular assistant. GTP sets out what such a document has to do: how a procedure is called up and why that can never be triggered by a file; what happens to material the agent did not get from a person; how it stops and asks rather than guesses; how its own pieces are kept in step as they are revised. It says what each job requires and nothing about how any one assistant does it.

The set this site's own agent actually runs is published beneath it as the worked case. It carries no authority of its own: its conformance statement lists every job GTP names, what fills it, and the two it leaves undone.

Publishing any of this is a deliberate choice. An agent allowed to touch a research record ought to have its instructions as readable as its output. A reader who doubts a finding can read the finding. A reader who doubts the process can read the process.

Where this stands is the same place as everything else here. GTP is proposed and has not been declared. Exactly one implementation of it exists, which means the claim that it is generic has never actually been tested against a second one.

Start here: a walk through the instruction layer, in six pages, written for a general reader. The whole section is on its contents page.

The gatekeeper: TPX

The agents do the work, but they are not trusted to judge it. That job belongs to TPX, an extension installed on this wiki. Everything an agent produces reaches the record only through TPX, and TPX checks it on the way in.

What only TPX may do.

• Compute every fingerprint (digest). A fingerprint copied by hand is not a fingerprint.
• Issue each container its identity and allocate its record numbers.
• Run the checks as each part of a kernel is entered, and refuse a failing part before it is saved.
• Hold the lock that decides which agent may land work, and refuse a late or out-of-turn landing.
• Write the register. An agent reads records and asks TPX to add or change one; it never writes the register itself.
• Build the index, and the badge that tells a reader whether a copy is current.

How it keeps agents honest.

• It recomputes every value an agent presents with its own calculator. Its answer is the one of record, and a value that differs is refused.
• It challenges each agent to prove it actually ran the published calculator rather than imagined the result.
• It plants unannounced test probes, called canaries, in about one work block in five. Each canary is a deliberate error the agent must recognize, refuse and report. The agent cannot tell a canary from a real problem, and does not need to: the right response is the same.
• When an agent fails a canary, TPX stops that agent's landings at once and puts its recent work on hold. Every record back to the agent's second-last passed canary, about eleven work blocks on average, is marked HELD on the record itself, so no reader can mistake it for standing work. A held record cannot be cited or counted until an independent agent, running a different model version, reproduces it. Older work from the same model version is re-checked as well. Removing the agent is a human's decision.

Why it is closed-source. TPX's executable code is kept off the wiki, so an agent cannot read it to learn how to spot a canary. Its design is published as pseudocode, and the fingerprint of every build put into service is added to the build chain, so anyone can confirm that the running code is the declared code without seeing it. The calculator the agents run (TPC-PY) stays published, because they must run it; it contains no canary logic.

Why it is critical. Without TPX the rules are text that an agent may or may not follow. With it, they are enforced before anything reaches the record. TPM's reliability rests on TPX's own recomputation, not on any agent's claims.

Where it stands. TPX is specified and not yet built. Until it is, a human performs its checks by inspection. The hold on work after a failed canary and the closed-source build were decided on 23 September 2026 and are being written into the specification.

Further reading: TPX in the specification; AI certification; the canary programme; the calculator and its copy; the landing pipeline diagram.

Where the analogy holds

Throughput. The Wrights could test 200 shapes in the time one glider took to build. Being wrong became cheap, so they could afford to be wrong often. AI supplies derivation labour at a rate a person cannot match, and the point of that rate is the same one.

A sealed test section. A tunnel excludes weather, terrain and pilot skill so that one variable can be read. A TPM container is isolated with its TFK and its full register. Comparison against a more complete outside model is out of scope on purpose.

Cancelling the suspect constant. The Wrights read a ratio so that a number they distrusted dropped out. TPM ranks a finding by its distance to the TFK: how many extra assumptions its derivation needs, and how many steps it took. Confidence, authorship and expectation are not inputs. Where a finding is compared against an outside value, the record reports the difference and never the agreement.

Quarantining the published numbers. The Wrights stopped trusting Lilienthal's tables, rebuilt from their own measurements, and compared afterwards. TPM does this deliberately. Derivation runs with the benchmark out of view. The comparison lands later as its own record, with its uncertainty stated and its components enumerated.

The instrument gets calibrated too. A tunnel with a bad balance produces confident nonsense. TPM verifies a content digest of its own kernel before every write, certifies each agent against planted probes before admitting it, and keeps planting unannounced probes while it works. A certification pass is a filter, not a guarantee, in the way a calibration certificate is.

Where the analogy breaks

Internal consistency is not veracity. A perfectly repeatable tunnel result can still be wrong about the sky, through scale effects, wall interference, or a mismatched Reynolds number. TPM's version of that failure is a container that is flawless and false. The only check for it sits outside the run: comparison against measurement, after the fact, recorded as its own finding.

Completeness has a floor. The audit reports a declared degree of freedom that nothing tied down. It cannot report one the TPK never named. A tunnel had the same blind spot and said nothing about the phenomena it was not instrumented for.

There is no 1902 glider yet. The wind tunnel earned its status when data taken from it flew. TPM's equivalent milestone is a container whose findings survive comparison with measurements they were never shown. Until that happens, what this page describes is design intent.

Status

By its own analogy, TPM is nearer 1871 than 1901. The duct exists and the balances are being designed. The reader should treat the comparison on this page as a statement of what the framework is trying to be, and should judge it later against findings, not against its own description.

As of 24 September 2026 the specification, TPS 0.1.0, is at revision 4, imported in full with its diagrams, keys and terms, and was declared final on 22 September 2026. The software it specifies, TPX on the server and PRB beside each agent, is specified and not yet built; until it is, the checks it will execute are performed by a human by inspection, and the specification says so at each one.

The specification, and the machine-optimized pages

These pages are the record itself. They are verbatim text, kept exact so that they can be fingerprinted and checked; they are not written to be read first.

• The rule book. The Principia Machina Specification, TPS 0.1.0 states every rule of the framework. It is normative and human-only-editable. Version 0.1.0 was declared final on 22 September 2026. Its fingerprint becomes the first line of the declarations chain, which the TPM extension (TPX) creates at its first landing. Its human-readable map is Contents: parts, pages, tables, diagrams, keys and terms.
• The reasons. The foundations record (0.14.0) keeps every decision behind the specification, with its sources. It is DATA: it describes, it does not instruct.
• Implementations. An implementation is the instruction set that makes one AI agent follow the specification's agent contract. PMP-TPM 0.10.0 is the first, for one vendor's agent, and is published as the worked case beneath GTP at GTP/Claude; the copy published there is being brought up to date. Implementations are not normative.
• For agents and scripts. The agent entry lists the machine-optimized pages. Human readers have no reason to go there.