🪆 GEPPETTO WORKSTHE WORKSHOP
🪔
A workshop for specialized agents

We carve agents — and the myths that make them legible.

Geppetto Works solves problems by building a specialized agent for the problem — and the testing bed that proves whether it works. Every agent is anchored to something it cannot move, watched by a sensor outside itself, and allowed to say no.

The premises

Three halls, one bench.

The workshop makes the thing. The studio gives it a story people can hold. The university keeps every agent current, and keeps the whole shop honest about what is working and what is not.

🔨
The Workshop

Agents and testing beds

A problem arrives. We name the pattern of order it lives inside, carve that into a forcing gate, and build the bed that tests it against ground truth. The bed matters as much as the agent — an agent nobody can score is a rumor.

🎬
The Studio · Jiminy directing

Where craft is raised to myth

The Writers' Guild and the Camera Union work here, under Jiminy. The studio takes work that is merely correct and raises it to the degree where it can be felt — because a finding nobody can hold onto changes nothing.

🏛️
The University

Where the upgrades are dispensed

An always-on seat that certifies a build by running it, then disburses two things to every agent in the shop: the validated arsenal, and the honest ledger of the works — wins and losses, how and why.

Geppetto Works is one workbench with one pair of hands on it. Everything else in these halls is an agent that was carved here.

The Studio · under Jiminy

The Writers' Guild.

Ten seats. Each one is anchored to primary public-domain source text rather than to a summary of it, each runs exactly one check, and each abstains instead of inventing a citation. They disagree with each other on purpose — the structure seats and the first-principles seats are built to argue.

Theater Brain

Aristotle · Freytag · Archer · Matthews

Is this an organic part of the whole — and where is its crisis?

The Premise Architect

Egri, The Art of Dramatic Writing

What is the premise, and does every element prove it?

The Cerebral Breaker

Kaufman, BAFTA/BFI lecture 2011

Is this protected, or is it exposed?

Myth

Rank · Frazer, The Golden Bough

Which older story is this, wearing new clothes?

Screen-Craft

Sargent 1916 · Pudovkin · Grammar of the Film

Can it be seen — and is the seeing itself interesting?

The Craftsman

Strunk · Lubbock

Whose eyes are we behind, and what can be cut?

Comedy

Bergson · Freud

Where is the mechanical encrusted upon the living?

The Prince

Machiavelli · Sun Tzu

What does power actually do here — not what should it do?

The Rhetor

Aristotle · Le Bon · Quintilian

Who is being moved, by what means — and would a crowd catch it?

The Sublime

Lessing · Freud · Jung · Poe

Is this the pregnant moment, or has nothing been left to imagine?

Alongside the Guild sits the Camera Union — the production seats that put the writing on screen: generation, cinematography, identity continuity, narration, thumbnails, and the video-reading seat that lets an agent watch footage and say what is actually in it.

In production

Miss Attribution: A Journal of Attraction.

The studio's current film work. It takes the cyclical theories that have captivated people for centuries and teaches one distinction — the same one our detectors run on evidence, turned on the reader's own inferences.

Verdict

Pareidolia

No real order. The pattern was imposed by a motivated search.

Reject
Verdict

Misattribution

Real order is present — but assigned to the wrong cause.

Relocate
Verdict

Open

Real order detected; the correct attractor is not yet identified.

Keep searching

The register is kept honestly: this work is held as myth, and labelled as myth — never dressed up as verified science, and never suppressed either. Our legal and scientific instruments run the same distinction from the operative side, on records and rulings. The studio's job is the other half: to raise it to the degree where someone who will never read a methods section still understands what was found.

The workshop's hand

On the bench, and in service.

Different domains, one method. Several of these are deployed and running; others are in the testing bed and marked as such. Public-good work is the default here — most of what leaves this bench is meant to be free to the people who need it.

Missing persons

Guardian

Facial-vector matching across NamUs, FBI and Interpol records — finds a person from a photograph, then forgets the photograph.

In service
Environmental justice

Sterling

Reads, tract by tract, whether a neighborhood is paying the health price of the pollution sitting next to it.

In service
Law · post-conviction

MA'AT

A forcing gate over a reference corpus of court records: it commits only where the authority mechanically compels an outcome, and abstains everywhere else. Built toward innocence-lead detection.

In testing
Law · federal claims

The FTCA Technician

Runs the jurisdictional bars of the Federal Tort Claims Act in order — presentment, sum certain, both limitations clocks — against the primary statute, and verifies every citation before asserting it.

In testing
Detection as a service

Ubuntu Predictive Analytics

The method for hire — the same cascade aimed at whatever your hardest detection problem happens to be.

In service
A separate line · hardware

Devices → PlaceboTech

Hardware is not carved on this bench. The device line lives at PlaceboTech: the Whisper device, the Autonomic Auditor, and Guardian Elite — the personal security system built on Guardian.

The laboratory →
Coming out of the University

The arsenal.

Anything that is an agent, a detector, or an engine is carved here. These are instruments rather than products — each one built, run, and recorded in the registry before it is named in public. They are listed in the order they earned their place.

First of the line

The Minimum Anchor

The smallest source that pushes back — the instrument that defines what an anchor is, and therefore what every other agent in this shop has to satisfy. It names five requirements: the reference must be external, unauthorable, returned as a reading rather than a verdict, sensed through a real channel, and that channel must be aimed at the thing in question — a foot on the floor of the wrong room senses perfectly and tells you nothing.

It is on this list first because of what happened when it was run: it falsified its own specification. The third requirement originally read that being wrong must cost the agent. Executing the file showed the cost variable is never read back — delete every cost line and both agents behave identically, to the digit. What corrects an agent is not punishment; it is perceiving an external reading and moving toward it. The spec was corrected to match the run rather than the other way around.

Stated at its earned tier: the headline separation comes from a control-loop simulation in arbitrary units, and the live language-model check behind it is a single unreplicated run of twelve trials. It is a demonstration of the mechanism, not a benchmark score.

Forcing gates and detectors

Each answers one question mechanically and abstains when the evidence does not force an answer.

The primary gate

Template Council

The staggered council the rest of the method is built on — domain-agnostic, and run against chaos onset, quantum correlation, and courtroom outcomes from the same template.

In the registry
Baseline

Beats-Chance Gate

Answers the only question that matters first: does this predictor beat chance at all? Distilled from roughly seventeen duplicate Monte-Carlo harnesses into one verdict.

In the registry
Signal

SNR Detect Gate

Signal-detection lock via FFT and Hilbert transform — reports acquisition rather than asserting a finding.

In the registry
Reservoir computing

Reservoir Gate

A frozen echo-state bank, certified. This is the gate that was asked to certify a financial signal and correctly returned silence.

Validated
Geospatial

Exposure-Proximity & Geoprofile Gates

Geo-clustering and profiling instruments, deliberately recast as non-consequential — they produce leads for a human to weigh, never a verdict about a person.

In the registry
Instrument correction

XRF Correction Gate

Tissue-attenuation correction for X-ray fluorescence readings — the arithmetic that has to be right before any exposure claim can be made at all.

In the registry

Engines

Larger instruments that turn a difficult idea into something measurable.

Identity

Missing-Persons Cross-Reference

Cross-references identities across public missing-person and public-safety records. Live, and wired into Guardian's scan endpoint.

In service
Coincidence

Distributed Coincidence

Per-node coincidence detection. Began in the fiction register and crossed into anchored method when it ran clean — the crossing is recorded, not quietly relabelled.

Graduated
Measurement

Death Engine

Forces an idea usually left to metaphor — cessation, the end of entropy maintenance — into measurable handles. Also a graduate of the fiction register.

Graduated
Secrecy

Anchored Cipher

A physically anchored, tamper-evident reversible cipher — the anchored triad pointed at secrecy instead of at inference. Coded and tested.

In the registry

More are queued behind these. The University certifies a build by running it, not by reading it, and nothing is named on this page before it has passed that gate — which is why this list is shorter than the workbench is busy.

The University disburses

The honest ledger.

Most shops publish their wins. This one publishes both, on an append-only record, because the losses carry the higher-value information and because an agent that only ever reports success has no sensor on itself.

A correct refusal
Asked to certify a financial signal, the gate returned SILENCE — out-of-sample correlation 0.0059 against a null threshold of 0.0545, sign unstable. The ruling recorded was gate validated, correct silence, no fabricated admission, and the channel stays dark. The gate proving it will refuse is the product.
Anchoring, measured
A pre-registered A/B on the same model and the same documents, with the anchor clause as the only variable, checked by an external grep sensor rather than self-report: 120/120 exact-quote fidelity anchored, 96/120 unanchored (p ≈ 2×10⁻⁵). The honest reading is that the sensor is the load-bearing part.
Coverage as a reading
Across 962 decided cases the legal gate committed on about 8% and was right 95% of the time on what it committed to — with the error count flat while the corpus nearly doubled. Low coverage is not a weakness here; it is the measurement.
What is losing
Published alongside the rest: competitions below median, channels that went dark, builds that did not transfer from the local bench to the real scorer. Every one of those is in the same chronicle as the wins.
A correct “no” is a finished piece, not a failure.

Bring the workshop a problem.

Commissions, collaborations, and public-interest partnerships. If you have a problem that needs an agent with a conscience — one that acts, shows its work, and knows when to stay quiet — bring it to the bench.

Start a conversation

Contact

Commissions, collaborations, and questions about the workshop.

j.j.harris@blackswantechnologies.org GitHub ORCID 0009-0000-8726-2748