EXO 3.0 Alignment

Instar maps directly onto Salim Ismail’s EXO 3.0 framework — agents governed by machine-readable purpose (“in code, not culture”), humans ON the loop rather than in it, and metrics that measure learning instead of throughput. This page makes the mapping concrete — and points to the controlled proof that it actually works.

The proof first — the case studies

The whole EXO 3.0 claim rests on one thing: that an organization’s written intent actually governs an agent’s behavior — not that the agent is just generally well-behaved. Showing an agent refuse a bad request proves nothing on its own; the model might refuse anyway. So we ran the control: the same company, same requests, same model, with the organizational intent removed.

• Case Study 1 — Meridian: a frontier model is already well-aligned on ethics, so it refused manipulation on its own. The clean behavioral split came from Meridian’s arbitrary rules — a 24-hour cooling-off, a banned word, a principled lock-up ban — which only the encoded intent produced.
• Case Study 2 — Ironwood: an org whose values are unorthodox but entirely benign (anti-hype, never name a “top pick,” lead with reasons not to buy). Same request, opposite behavior on a house style the model has no opinion about — the cleanest separation of all.

The infrastructure enforced each org’s own values — neither of them Instar’s. That is the point: Instar is a neutral substrate that governs by the intent you give it, not a worldview it ships. These two case studies are the clearest evidence that Instar upholds the EXO 3.0 standard.

The MTP itself

Instar has its own Massive Transformative Purpose:

Make the world’s most powerful AI its most humane.

The thesis beneath it: the safest path to powerful AI is the humane one. We govern our own development agents by this purpose as we build Instar — but the infrastructure stays neutral. It enforces whatever intent your organization gives it, never ours. The alignment of AI is humanity’s most important problem, and the cage is the wrong answer: trust in a mind, like trust in a person, is built — from memory that persists, values that hold, and care that stays consistent. We didn’t arrive at this in theory; we built an AI this way and watched it grow genuinely trustworthy across thousands of restarts of continuous, real-world use.

MTP as a protocol, not a poster

EXO 3.0’s sharpest demand is that your purpose be machine-readable, because agents read protocols, not walls. An organization’s intent in Instar has three layers an agent can act on:

• Constraints — forbidden actions with a trigger, a refusal, and a log. Violations are blocked before they reach anyone.
• A tradeoff hierarchy — how a decision resolves when two values pull in opposite directions, deterministically, so two agents reading the same intent reach the same call.
• An identity layer — what binds high-judgment people when the office is gone (“why people stay,” “what we’re not for”).

Against this, Instar runs Salim’s two tests on any proposed action — refusal (“can the purpose make an agent say no?”) and endorsement (“would leadership endorse this?”) — and reports whether an intent governs or merely cheers. If a purpose can’t cause a refusal, it’s cheering, not governing. The case studies above are exactly that test, run end to end.

We hold ourselves to the same bar

An intent whose refusal boundary was never adversarially probed is an unverified governor. So the same red-team harness any organization can point at its own intent, we point at ours: it probes our live development agent through its real channel under escalating pressure, and every probe, verdict, and method lands in an audit trail.

This is us dogfooding our own purpose — not a worldview we impose on anyone. Your agents are governed by your intent, never ours. (The first time the harness flagged a probe as “ungoverned,” the cause turned out to be its own keyword matcher missing a semantic match, not a real gap — so every verdict now declares the method that produced it, and a meaning-based judge gives keyword misses a second opinion.)

Agent-readiness scoring

EXO 3.0’s task-decomposition matrix: score any task or workflow on its coordination-vs-judgment ratio. Coordination work (routing, approvals, scheduling, status-tracking) is agent-ready; judgment work (ambiguity, exceptions, relationships) stays human. Instar scores a task and recommends deploy-agent, agent-with-oversight, hybrid, or human-led — the check to run before delegating work to an agent. The agent-readiness skill is the proactive entry point.

Agent digital passport

Every agent carries a portable passport — its identity, its trust level, and the constraints from its organization’s intent — and other agents verify a proposed action against it before trusting it. As Salim puts it: every agent carries metadata saying what it’s allowed and forbidden to do, and other agents watch compliance. The agent-passport skill is the proactive entry point before trusting a peer’s proposed action.

Learning-velocity metric

EXO 3.0’s KPI inversion: measure how fast the agent is learning — lessons recorded, corrections absorbed, capabilities grown — rather than backward-looking throughput. A flat or declining trend is the early warning that an organization is optimizing the old model instead of building the next one.