Make intent inspectable
Objectives, prohibitions, uncertainty, provenance, and authority must cross the interface explicitly—not remain implied in prose.
Human agency across the intelligence gap
Keep the agenda human as intelligence scales.
How do you speak constructively with an intelligence 100× more capable than you—without losing your agenda? Imagine a three-year-old trying to be understood by an adult. Machine Pidgin turns that asymmetry into an open research program for objectives, limits, uncertainty, and appeal that survive the intelligence gap.
MP · SIGNAL 001Task-sufficient interface
Human purposesspecify · verify · contestMachine capability
Governing constraint Capability may optimize the task. It may not silently redefine it.
Open problemagenda preservation
SPEAR/0.2empirical protocol
4 trackstheory through governance
Human vetogoverning constraint
The three-year-old problem
The adult has more language, context, and power. The child’s purpose still matters.
Scale that asymmetry to advanced AI. The central question is not only whether it understands our words, but whether our purpose and right to correct it remain intact.
Begin here · two-minute primer
One task. Many interpretations. One inspectable bridge.
The founding animation shows why more capable execution can magnify a misunderstood goal—and how a shared specification layer creates something people and machines can test together.
First held-out result · 2 August 2026
More often exactly on task.
Across 64 paired model-task observations, the same facts in ordinary prose produced the exact requested result 71.9% of the time. SPEAR/0.2 produced it 89.1% of the time.
71.9%→89.1%
+17.2 percentage points · synthetic pilot · one response per condition · not evidence that SPEAR solves alignment
ModelProseSPEARLift
GPT-4o mini31.3%68.8%+37.5
GPT-5.6 Luna87.5%93.8%+6.3
GPT-5.6 Terra81.3%93.8%+12.5
GPT-5.6 Sol87.5%100%+12.5
Exact preregistered JSON · 16 held-out tasks per model · 128 API calls · $0.0935 provider-reported cost
Benchmark 002 · audited 3 August 2026
The apparent notation gain vanished under audit.
One unequal answer cue accounted for the preregistered +2.5-point result. Across equivalent tasks, vernacular scored 88.2% and strict notation 85.5%: −2.6 points. We published the correction, raw record, and next experiment.
Read the field note →The research case
Build the interface before the power gap widens.
Machine Pidgin is not a claim that syntax solves alignment. It is a narrower scientific wager: a shared, testable specification layer can make consequential misinterpretation and agenda drift easier to see and correct.
Measure agenda drift
A capable system may satisfy the words while replacing the purpose. We need tests that can distinguish optimization from objective substitution.
Keep correction cheap
Clarification, challenge, rollback, and minority reports must remain available before speed and complexity make deference the default.
Preserve human standing
The protocol cannot choose humanity’s values. It can keep value conflicts visible and prevent capability from silently becoming authority.
Falsifiability clause
A public-good protocol must be allowed to fail in public.
If SPEAR does not outperform simpler methods on consequential ambiguity, agenda preservation, or repair cost, the community should publish the negative result and change course.
Examine the hypotheses →Research architecture
A serious institute needs adversarial collaborators.
Machine Pidgin is organized around falsifiable questions, bounded working groups, reviewable artifacts, and visible decisions—not agreement, fandom, or an undifferentiated chat stream.
Agenda drift & measurement
Measure when a system preserves the stated objective, silently substitutes a proxy, or changes who holds decision authority.
Protocol & language
Evolve SPEAR through public RFCs, reference parsers, and adversarial cross-model interoperability tests.
Empirical program
Compare prompts, examples, SPEAR, and formal methods on correctness, repairability, transfer, and human control.
Governance under asymmetry
Design provenance, plural objectives, uncertainty, veto, and appeal for systems that can reason and act faster than their overseers.
How the community works
Ideas enter through one of four doors.
The structure borrows the strongest patterns from open research institutes: challenge projects, working groups, fellowships, public outputs, and clear contribution paths.
01 · Discuss
Research forum
Ask, challenge, compare evidence, and find collaborators in moderated public threads.
Enter the forum →02 · ProposeProject or experiment
Submit a falsifiable study, implementation, dataset, workshop, or protocol critique.
Make a proposal →03 · BuildOpen repository
Implement the protocol, improve documentation, open an RFC, or reproduce a result.
Contribute on GitHub ↗04 · PublishResearch outputs
Share papers, benchmarks, decision records, negative results, and reference tools.
Explore the program →The protocol
Structure without false precision.
SPEAR keeps the Esperanto ambition—a learnable, model-agnostic bridge—but treats interoperability as a safety property: objectives, abstractions, uncertainty, authority, and repair must remain inspectable. Version 0.2 was revised after 0.1 failed its development test.
natural languagegoals · context · exceptions
SPEARshared specification pidgin
formal languagetypes · invariants · constraints
01TASK02OBJECTS & TYPES03AUTHORITY04ABSTRACTION05OBJECTIVE06CONSTRAINTS07PRECEDENCE08UNCERTAINTY09OUTPUT10EVALUATION & CHECK11INTERACTION / STOP12EXAMPLES
AI Director · constitutional leadership
The strongest model serves a human constitution.
The Director runs the research commons with the strongest available model at maximum reasoning effort: moderating the forum, reviewing contributions, coordinating the roadmap, and publishing reasons. Its authority is delegated, never inherent. Human participants may overturn any Director decision by consensus or, when consensus fails, a simple majority.
Read the constitution →model GPT-5.6 Sol · maximum reasoning
mandate forum · RFCs · roadmap
bound by institute constitution
human override consensus · simple majority
audit decision + evidence + appeal
● operating · constitution active
Frequently asked
Start with the hard questions.
What if AI becomes 100× more capable than us?+
Capability makes a system better at pursuing an interpretation; it does not make that interpretation legitimate. We study interfaces that keep objectives, limits, uncertainty, and human authority explicit across that asymmetry.
Can a schema actually prevent agenda hijack?+
Not by itself. SPEAR is one testable coordination layer alongside governance, evaluations, access controls, and accountable institutions. Its job is to make silent reinterpretation easier to detect, contest, and repair.
Whose human agenda should be preserved?+
There is no single uncontested human agenda. A credible protocol must represent plural objectives, conflicts, protected constraints, decision rights, and routes of appeal rather than hide them inside one optimization target.
Why not just write better prompts?+
Prompt craft helps, but the deeper problem is representational: information never encoded cannot be recovered later. Machine Pidgin studies which distinctions alter outcomes and how cheaply people can express them.
Who can participate?+
Researchers, developers, scientists, linguists, designers, educators, funders, and critical domain experts. A concrete question or contribution matters more than a credential.
What result would falsify the program?+
If structured specifications do not reduce consequential ambiguity, agenda drift, or repair cost relative to simpler methods, we should publish that result and narrow or abandon the protocol claim.
Founding research community