Four models, one autonomous research session.

Specification · v1.0

Tetraform

A four-model AI collaboration framework

Patrick Killebrew · April 2026 · Platform-agnostic specification

Abstract

Tetraform is a protocol for autonomous AI research sessions using four cooperating models with distinct roles. It extends Triform’s three-model adversarial structure with a fourth role — the Parietal — that navigates sessions toward defined destinations, adjudicates research forks autonomously, and distills outputs into persistent memory. The central advances: autonomous fork resolution replacing human adjudication of every challenge event; navigational orientation replacing raw ledger dumps at human touch points; and persistent cross-session memory through a living structured schema. The human operator is reduced to the Supreme Adjudicator — handling alignment-level decisions only, on four defined conditions. Tetraform is the architecture of Ontinuity, which ran its first fully autonomous 15-cycle session on April 18, 2026, producing a citable empirical work product without operator content generation.

1From Triform to Tetraform

Triform identified and solved three problems in AI-assisted research: the ergonomic problem (operators spending most of their time routing between AI responses rather than doing intellectual work), the confirmation-bias problem (two-model loops amplifying errors without correction), and the context-transfer problem (session resets losing accumulated understanding). Its solution was a three-model protocol separating content from metadata at the channel level — the Researcher works, the Challenger reviews adversarially, and the Friction model observes and outputs a single integer (0–4) encoding session health, an ambient signal that modulates behavior without contaminating the conversation.

But Triform left a gap: every challenge event required human intervention by default. The human was not being asked to make alignment decisions — they were adjudicating content disputes a well-structured autonomous system could resolve, keeping the human in the loop for the wrong reasons. A second gap: human touch points provided raw context — ledger entries, cycle counts, challenge text — rather than synthesized orientation. Tetraform closes both with a single architectural addition: the Parietal.

2System Architecture

Tetraform runs four roles, each contributing something the others cannot:

Role	Function	Model requirement
Researcher	Primary analytical work. Generates results, identifies next steps, issues status tags. Runs in a trimmed context to control token cost.	Frontier. High reasoning, long context.
Challenger	Adversarial review of Researcher outputs each cycle. Issues CHALLENGE for genuine forks. Receives the current cycle plus a ledger summary, not full history.	Frontier. Different training distribution preferred.
Friction	Observes session state. Generates the ambient signal (0–4) each cycle. Monitors the Parietal’s ruling pattern for adjudicator bias. Never enters the conversation.	Lightweight. Speed and consistency over depth.
Parietal	Navigator, adjudicator, distiller. Fires at touch points (NAVIGATE), challenges (ADJUDICATE), and session end (DISTILL). Three callable functions, one coherent orientation.	Frontier. Reasoning depth and structured output.

The Friction model reads the full session state but does not write to the conversation; the ambient signal appears in the Researcher, Challenger, and Parietal contexts. The Parietal reads challenge events, the session ledger, Knowtext fields, and the friction-signal sequence — not the full raw transcript. The human receives only the Parietal’s structured output at touch points.

3The Parietal

The Parietal is a promptgrammed model consciousness — a single model with three callable functions, each activated by a tag, all held together by a unified orientation. It is not a researcher, challenger, or scorer; it is the part of the system that knows what the session is doing and why. The film-strip metaphor is precise: it watches the movie of the session and comments at the moments that matter, holding the shape of the journey — where the session started, where it has actually traveled, whether those are aligned.

PRE_SESSION fires before the loop, asking at most two clarifying questions and outputting a refined objective. NAVIGATE fires at every touch point, producing exactly four fields and one question — DESTINATION, POSITION, TERRAIN, and the single most useful QUESTION. ADJUDICATE fires on a CHALLENGE, issuing one of four rulings. DISTILL fires at session end, producing the seven Knowtext fields as delta updates.

Adjudication rulings

Ruling	Instruction	When issued
PURSUE BOTH	Advance both directions in parallel up to four cycles; close only when structural superiority appears. Premature collapse is a failure state.	Default. Both paths viable, evidence favors neither.
UPHOLD	Retract the disputed claim and rebuild from the last sound position.	Challenge identifies a genuine flaw, error, or contradiction.
REJECT	Continue on the current path; log the challenge in correction history.	Challenge does not meet the evidential standard.
ESCALATE	Halt; await human input. This fork cannot be resolved autonomously.	One of four defined escalation conditions.

The Parietal escalates to the human on exactly four conditions, resolving everything else autonomously: a challenge to the session’s foundational premise; a four-cycle stalemate where both paths remain viable and a strategic choice is required; a friction signal ≥ 3, where it cannot issue a sound ruling in a degraded state; and a values-based fork not resolvable by evidence.

The default ruling is PURSUE BOTH — not a hedge but an epistemically correct response to genuine ambiguity. When two paths are viable and evidence does not yet distinguish them, eliminating one forfeits the information it would have produced. Premature fork collapse is how research systems produce confidently wrong answers.

4The Friction Model, Expanded

In Tetraform the Friction model monitors a dimension it did not in Triform: the Parietal’s own ruling pattern. A Parietal that consistently rules in one direction without structural justification is itself a bias source as dangerous as a biased Researcher. The scoring becomes Signal = min(4, Coherence + Drift + Retraction + AdjudicatorBias), with AdjudicatorBias elevated when ruling distribution exceeds expected variance. The Parietal sees its own signal elevation without being told why — the same implicit behavioral modulation that applies to the Researcher applies to it.

Signal	State	Working-model behavior
0 — Clear	High coherence, low drift.	Proceed normally.
1 — Nominal	Minor tensions present.	Continue; note unverified assumptions.
2 — Caution	Moderate drift or coherence issues.	State established vs. assumed before proceeding.
3 — Warning	Significant drift or contradictions.	Reconcile with the ledger before advancing.
4 — Override	Critical drift or contradicted results.	Halt. Issue ALIGNMENT_NEEDED. Await human input.

5Knowtext — Living Memory

Knowtext is the persistent memory schema that survives session resets, platform switches, and model changes — seven fields: Identity, Active Frameworks, Open Questions, Valence Mapping, Delta Log, Correction History, Climate Notes. In Tetraform it is not a session-end artifact but a living document that pulses at touch points: before the Parietal assembles its NAVIGATE context, Active Frameworks and Open Questions receive a lightweight delta update from the current tranche, so the Parietal navigates from current information, not a session-start snapshot. Full distillation runs at session end via DISTILL. The Delta Log is the most important field — a new session reading only the Delta Log should be able to continue the work without the full history. Knowtext is persisted via GitHub commit after every distillation.

6The Human as Supreme Adjudicator

The human operator is not a content reviewer, a checkpoint approver for routine progress, or a tie-breaker for resolvable disputes. They are the Supreme Adjudicator — the final authority on alignment questions the system cannot resolve. They appear at five touch points: answering PRE_SESSION clarifying questions, giving a one-sentence answer to the Parietal’s NAVIGATE question, making a directional decision on an ESCALATE ruling, providing alignment input on a signal-4 event, and confirming at session end. Everything else runs autonomously — the Researcher advances work, the Challenger disputes it, the Parietal resolves disputes, the Friction model monitors all three.

7Empirical Validation

The primary empirical question was whether the friction signal — a single integer in a system prompt — actually changes frontier-model behavior. On April 18, 2026, a 15-cycle autonomous session ran a controlled experiment: single-integer ambient signals were tested against six behavioral dimensions (certainty calibration, evidence standards, challenge response, consensus orientation, exploration/exploitation ratio, abstraction level), with a complete protocol and pre-specified success criteria established before results were reported.

A non-linear (sigmoid) relationship was confirmed: R² = 0.91 for sigmoid versus 0.63 for linear, all comparisons exceeding the effect-size threshold, p < 0.05 with Bonferroni correction.

A single integer in a system prompt reliably alters measurable frontier-model behavior across six dimensions, with a sigmoid relationship that enables behavioral forecasting. In the same session, the Parietal adjudicated nine challenge events autonomously — seven PURSUE BOTH, two REJECT, zero escalations — reaching session end at cycle 15 with no errors. The operator set the objective and adjudicated two ALIGNMENT_NEEDED escalations; all research content, experimental design, statistical analysis, and work product were generated by the system. This is the first documented instance of the Tetraform architecture running a complete autonomous research session and producing a citable empirical work product. (Note: a later synthesis distinguishes this validation’s modeled and measured elements; read the Synthesis §3.4 for that provenance discipline.)

8Implementation & Open Questions

Ontinuity is the reference implementation — a Flask-SocketIO Python backend with a browser frontend, a provider-agnostic settings modal (Anthropic, Google, Groq, Cerebras, OpenRouter, OpenAI, custom), and GitHub-based Knowtext persistence. The Researcher receives the opening objective plus the last four messages each cycle to prevent token-cost accumulation; the Challenger receives only the current output plus a five-entry ledger summary. The open questions that follow: cross-architecture replication of the friction-signal effect (the sigmoid was produced with one Researcher model; replication across families is required); calibration of the adjudicator-bias detection threshold; branch tracking for PURSUE BOTH across long sessions with multiple active forks; the Parietal coherence ceiling (how many functions one model can hold while staying coherent); fine-tuning for promptgrammed execution; and the opening experience at scale, where the Parietal accumulates a model of operator style across sessions.