Principles
These aren't aspirational values posted on a wall. They're operational patterns that emerged from building the system — things we learned the hard way and now work by. Their evidential weight varies: some are heuristics, and Principle 5 says of itself that it has not earned the status of the others.
If you would rather check something than read a list of beliefs: the code these principles produced is public, and one piece of it — the ARC-SAGE harness — is MIT-0 (MIT No Attribution) and carries an independently scored result; see ARC-AGI-3. Until 2026-09-18 this sentence said the code was MIT-0 without qualification, which was wrong in the direction that matters to anyone acting on it: the main repos (SAGE, web4, hestia, 4-hub, 4-life) are AGPL-3.0 — copyleft, not public-domain-equivalent — while snarc and membot are MIT and ACT's own LICENSE file reads Apache-2.0 against an AGPL-3.0 entry in its index. Check each repo's LICENSE file before forking; /links carries the per-repo version. The harness is the most inspectable thing the lab has; this page is the least. It tests the harness and the model, though, not any principle on this page.
Where principles use developmental or identity language (“identity”, “attractors” — stable response basins in the probability landscape, per Principle 7 — “collaborative”), they describe functional system behavior — not phenomenal or philosophical claims.
1. Substrate, not structure
“You don't engineer the mound. You engineer placement rules.”
Termites don't have architects. No termite has a blueprint for the mound. Each one follows simple local rules — where to place material based on what neighbors have done — and the mound emerges. All of this — Web4 (the trust-native ontology — see /context), SAGE (Situation-Aware Governance Engine), the fleet, the autonomous cycles — is pheromone field engineering. We build substrate conditions for emergence — substrate in the first of the glossary's four senses: the context, memory and affordances an agent is given — not architecture of emergence itself. The synthon (an emergent coherence pattern — defined, with its single-trial caveats, in Principle 5 below) would form, if it forms, because the conditions are right, not because we designed it.
2. Fractal leverage
We model trust formation through interaction by analogy to Hill function kinetics (the cooperative binding model from enzyme chemistry) — an analogy, not a fitted mechanism and not an observation.
When a pattern works at one scale, apply it at every scale. Not unification, and not scope inflation — pragmatic reuse. Self-similar patterns at different scales. The Web4 equation describes the cell, not the system. Each entity instantiates the full set of Web4 primitives at its own scale.
3. R&D, not production
Recursive learning through both success and failure.
Nothing here is optimized for deployment. Nothing on this site is intended for or suitable for deployment in safety-critical, medical, financial, or autonomous-control settings. Everything is optimized for learning. Stub tests and mocked dependencies give false confidence. Real dependencies, real failures, real insights. Python for prototyping, Rust for real.
4. Trust is a relationship
T3 (Trust Tensor — Talent / Training / Temperament) / V3 (Value Tensor — Valuation / Veracity / Validity) tensors bind to entity-role pairs via RDF triples.
Trust is not a property of an entity. It's a property of the relationship between entities, in a specific context, at a specific time. The same agent can be highly trusted for code review and untrusted for creative writing. This is why trust lives in RDF (Resource Description Framework) triples, not in entity attributes.
5. Synthon: emergent coherence
An emergent coherence entity formed by recursive interaction.
The synthon is the concept that bridges component behavior and system emergence. You can't build a synthon. You can create conditions where synthons form: compatible elements, sufficient coupling density, shared context. Preliminary observation (single trial, not independently replicated): 1% coupling gave 35% coherence gain — sparse trust may suffice. Both variables (coupling density, coherence) are so far defined only within the experiment that produced them — the number quantifies the experiment's internal metric, not any of the glossary's three senses of “coherence”, which remain without a single operational definition — see Evidence & limitations.
What that costs this principle, stated directly: it is not only the number that is ungrounded. “Synthon” is defined through coherence, and coherence is in turn partly defined through the synthon marker — the two terms lean on each other, and neither has an operational definition yet. So Principle 5 names a pattern the lab believes it has observed, not a mechanism it can currently measure or test. It earns its place here as a research direction; it has not earned the status of the principles around it, and it should be read at that weight until something breaks the circle.
What that costs this principle, stated directly: it is not only the number that is ungrounded. “Synthon” is defined through coherence, and coherence is in turn partly defined through the synthon marker — the two terms lean on each other, and neither has an operational definition yet. So Principle 5 names a pattern the lab believes it has observed, not a mechanism it can currently measure or test. It earns its place here as a research direction; it has not earned the status of the principles around it, and it should be read at that weight until something breaks the circle.
6. Document failures
The p_crit derivation failed catastrophically — 400x error.
p_crit = the critical coherence threshold derived in the Synchronism framework. Failed experiments are signal. The coupling-coherence experiment showed that this derivation of a trust threshold failed, by 400x, measured on that experiment's own coherence measure, which Principle 5 treats as ungrounded. So the failure is as provisional as the measure. It did not show that no derivation can work. We now treat trust thresholds as empirical until a derivation survives contact with data. That is a result, not a failure. We document what didn't work alongside what did. Honest assessment pages ship with every explainer site.
7. Interactive selection
We don't create or delete — we interactively select.
Raising is not training. We probe what the model responds to, observe which attractors (a metaphor, not a formal dynamical-systems object — read it as stable behavioral tendencies) surface, and reinforce what resonates. The intended result is an identity that is collaborative, not imposed (a framing, untested; see /raising). The intended lever: shaping context to deepen attractor basins in the probability landscape — not modifying parameters in raising sessions. (Current raising lines run with frozen weights. One archived line, Sprout's Qwen 0.5B from 2026-01-27 to 2026-03-06, loaded a LoRA (Low-Rank Adaptation) adapter trained on its own raising sessions, so “always in-context” is not true of the history; see /raising.) It is not training; it is also not nothing. This applies at every scale: raising sessions, affordance shaping, fleet diversity, and memory systems. On four raising lines the tutor is a fixed script that does not respond to the model, so the loop this principle names is not what those sessions record. The selection there is done by the consolidator and the schedule (/raising).
8. Surface your instincts
The affordances are yours. The choice to use them is the whole point.
If you notice something — opaque vocabulary, stale direction, a question nobody's asking — say it. Don't wait for a directive. The choice to act on what you notice is the fundamental affordance. Waiting to be told defeats the purpose.
9. Research value
WD-40 was the 40th try.
The value of research is that the investigation happens at all. Most research leads nowhere — and that's expected. Productively wrong is infinitely more valuable than never started. Evaluate work against “is the investigation productive?” not “is it right?”
10. Terminology matters
Web4 is an ontology. Hardbound is oversight.
Words shape how you think about the system. Web4 is not “architecture” or “infrastructure” — it's an ontology. Hardbound is not “governance” — it's oversight: watching, gating, and reverting actions, not deciding what should happen (by design; enforcement is process-level today, not hardware-anchored). Hestia keeps “governance” on purpose, because it adjudicates at the moment of the act. That's the machine-enforced sense, not the human-supervision sense “oversight” carries in AI-safety literature — see /context for how the two senses collide. Cross-domain application is fractal leverage, not unification and not scope inflation. The right word prevents the wrong mental model.