Alignment & the substrate-alignment standard
Most platforms let you assert that your agents behave well toward the people and
systems they touch. System9 lets you prove it. The behavioural discipline
that makes agent behaviour auditable is an open standard,
substrate-alignment, and
System9 ships the reference implementation of it and is conformance-checked
against its public suite.
In the console this surfaces as the Alignment tab on the AI hub. Under the hood it is the substrate-alignment primitive set: a value gate, a deliberative foresight stage layered above it, an operating-mode classifier, drift signals, a halt protocol, and a tamper-evident audit chain, applied to every entity in the estate.
These are not tools an agent chooses to call. Together they form one integrated decision organ, a fast, involuntary reflex value check and a deliberative foresight stage above it, that runs automatically on the live action seam. See the executive-functioning engine for the full sense→decide→act loop this organ sits inside.
substrate-alignment is an open standard, a reference Python implementation,
and a machine-checkable conformance suite for substrate-alignment primitives in
multi-entity agent systems. It lives at
github.com/System9ai/substrate-alignment
(Apache-2.0). The Python reference implementation was extracted from this
platform, where these primitives are in production use; System9 is the standard's
reference production deployment.
What "aligned" means here
An agent, a service, or an automated workflow is aligned when its decisions produce a net gain for everyone they affect, not only for itself. It is misaligned when it advances its own objective, more revenue, a closed ticket, a higher reward, by quietly taking value from the people and other systems it interacts with. Aligned to what: to the net benefit of the whole system the agent is embedded in, measured across every party its decisions touch, rather than to its own private score.
This is a property of the running system, not of any single model in it. The model supplies intent; the surrounding harness supplies the check. That is why the same discipline has to hold whether the model is a sovereign edge model, a bring-your-own-model, or a cloud model reached over a relay, the harness is what makes the behaviour aligned, and the harness is local to the runtime.
The naming, made precise
The word substrate is engineering-grade vocabulary and is retained in code and
in the standard; the open-source repository is literally named
substrate-alignment. The customer-facing console label for the same subject
is Alignment.
| You are reading / writing about… | Use this term |
|---|---|
| The open standard, its primitives, its spec/conformance vocabulary | substrate-alignment, substrate |
| The console tab on the AI hub | Alignment |
| The value gate at the action boundary | Net Potential Gain (NPG) gate |
| An entity's persisted alignment state | substrate metadata (operating mode, alignment vector, net potential) |
Two things this section deliberately does not call the value gate: it is not a "net-effect" gate and not a "net-perturbation" gate. Its name is Net Potential Gain, it scores an action's system-wide net potential gain at the boundary and returns refuse or allow.
Three deliverables, mirrored on this platform
The standard ships three things; the platform maps onto all three.
*Figure: the substrate-alignment standard (slate), its conformance suite (green), and its reference implementation, and how System9's two runtimes implement them: the cell (core, filled) implements the primitives and the edge (Edge Agent, outlined) is a faithful mirror whose answers are re-derived and pinned by cross-language fixtures. Both runtimes are checked by the same public conformance probes.
Table equivalent
| Deliverable (standard) | What it is | Platform mapping |
|---|---|---|
Standard (spec/) | Nine language-neutral, RFC-2119 specifications | The contract both runtimes are held to |
Conformance suite (conformance/) | 48 machine-checkable behavioural probes | Run against the cell and the edge |
Reference implementation (python/) | The zero-dependency Python witness | Runs as the cell runtime |
| The disconnected-capable runtime | The edge (Edge Agent, Rust), a faithful mirror of the cell |
The primitives, in one screen
The standard defines a small core surface that almost every integration uses, plus an extended substrate-mechanical vocabulary. The core is what the Alignment console is built on:
| Primitive | What it decides | Where to read more |
|---|---|---|
| Operating mode (substrate mode) | Is an entity LONG_CYCLE, SHORT_CYCLE, MIXED, or UNKNOWN? | Substrate primitives |
| Alignment vector / net potential | The trust · expertise · capability · health components that roll up into an entity's net potential | Substrate primitives |
| Net Potential Gain (NPG) gate (reflex) | Is a consequential action net-positive across the entities it affects, right now? | The NPG gate |
| Foresight composer (DECIDE stage) | Does a permitted action bend an entity's trajectory, its enclosing scope, or the world-model the wrong way? Tightens only. | The foresight composer |
| Drift signals | Which of the characteristic misalignment patterns is an entity trending toward? | Substrate primitives |
| Halt-and-escalate | When must a primitive stop rather than proceed? | Substrate primitives |
| Pair coupling | Is a two-entity relationship extractive, or symmetric? | Substrate primitives |
| Audit chain | Is the record of decisions tamper-evident and peer-witnessed? | Compliance & Audit |
How alignment state reaches an organization view
An alignment posture is formed from what individual cells observe, but a cell's observations never leave its node. They converge across the node's cells and stop there, so the organization view cannot be a replicated table.
It is a fold and a signed push instead. Each node folds its cells into one node-level summary, signs it with its node identity, and pushes it to the organization proxy, which verifies the signature, keeps the latest summary per node, and folds its own cells plus its verified peers into the organization rollup.
The useful way to hold it: node-to-node is two people talking. A peer knows the node, through a summary that node folded and signed. It never sees that node's cells. A single or air-gapped node folds to itself and pushes nothing, so there is no phone-home on this path.
The boundary is an encapsulation boundary, not a range limit. Raw per-cell telemetry never crosses a node; what crosses is a signed, self-sufficient document. So cross-entity coupling, whether a relationship is symmetric or extractive, is node-scoped by that rule, because its peers are cells. Asking it across nodes would mean per-cell telemetry crossing, which is the thing the rule forbids.
Cross-node cooperation is expressed as quorum witnessing instead. A cell that takes a governed decision writes a signed self-witness, and a decision's witnesses are unioned across the organization, each one re-verified against the witnessing cell's enrolled key on arrival and dropped if it does not verify. So "was this decision witnessed by enough distinct peers?" is answered across nodes, and a defecting peer in a sovereign multi-node estate becomes visible, without any node exposing its cells to the others.
A lone or air-gapped node computes the quorum from its own witnesses, which is the honest answer for an estate of one. The threshold is governed and deny-closed to a high bar, so a single self-witness is never mistaken for a quorum.
See platform replication for the mechanism in full.
How this composes with the rest of governance
Alignment is not a separate silo. It is the behavioural half of the platform's deny-closed governance. A consequential action is first checked against the policy/grant graph (default-deny), and only an action that clears that check is scored by the NPG gate. The NPG gate is the same gate the capability & approval flow delegates to, and its decisions are recorded on the tamper-evident audit chain.
- The Net Potential Gain (NPG) gate, the reflex value test at the action boundary, on both runtimes.
- The foresight composer, the deliberative DECIDE stage above the reflex, on both runtimes.
- The executive-functioning engine, the sense→decide→act loop the decision organ runs inside.
- Substrate primitives & behavioral probes: the coupling model, the probes, and the two-runtime mirror.
- Security → Governance, how the NPG gate sits inside deny-closed governance.
- The substrate-alignment standard on GitHub: the spec, the conformance suite, and the reference implementation.