The foresight composer: the DECIDE stage

The NPG gate answers a point-in-time question: may this action pass right now, given its net effect on the entities it touches? The foresight composer answers the harder one the executive-functioning engine exists to ask: the reflex permitted this action and it is locally net-positive, but does it bend the long-horizon trajectory the wrong way?

Together they are one integrated DECIDE organ, not a toolbox of tools an agent invokes. The reflex gate is the fast, involuntary first stage; the foresight composer is the deliberative stage layered above it. Both run automatically on the tool seam, the composer fires after the reflex permits and is never a tool an agent chooses to call.

The one rule that makes it safe

The foresight composer can only tighten a permit, never loosen a refusal. Its dispositions move in one direction: Proceed → Hold → Refuse. A reflex refusal is a hard floor the composer returns unchanged; it can add caution, it can never manufacture permission.

The two stages on the seam

*Figure: the DECIDE organ is two stages on one seam. The reflex NPG gate scores the point-in-time net effect and refuses outright on a negative; only a permitted action reaches the foresight composer, which deliberates over the long horizon and can tighten the permit to Hold or Refuse but never loosen a refusal.

Table equivalent

Stage	Question	Outcomes
Reflex: NPG gate	Net effect on affected entities, now	net_positive / net_neutral → permit; net_negative → refuse (floor)
Foresight composer	Does the permitted action bend the long horizon wrong?	Proceed, Hold, or Refuse (tightening only)

The afferent arc it reasons over

The composer does not sense anything itself; it reasons over readings the platform has already gathered and provenanced. The arc runs from observation, to an entity's measured trajectory, to the deliberation:

Figure: the afferent arc into the DECIDE stage. An entity's measured trajectory (from bitemporal memory) and its enclosing topology scope feed the composer live on both runtimes (solid, green/slate). The afferent sensor fusion front and the world-model dissonance brake are drawn as candidate inputs (cyan, dotted): the fusion and signing machinery is built and parity-pinned but activates on a sensor + model binding (planned, sensor / model boundary), and the dissonance brake is live on the core while the edge's input is planned (model boundary). The composer runs and decides correctly in their absence, it degrades each missing input to an honest-absent neutral.

Table equivalent

Input	Feeds	Status
Afferent observations (fused, provenanced)	World-model	Machinery built + parity-pinned; activates on sensor + model bind (planned)
Bitemporal memory	Measured trajectory	Live, both runtimes
Measured trajectory (Improving/Stable/Declining)	Composer	Live, both runtimes
Enclosing topology scope	Composer (nested-scope roll-up)	Live, both runtimes
Semantic world-model coherence	Composer (dissonance brake)	Live on core; planned at edge (model boundary)

What the composer weighs

Once the reflex permits, the composer evaluates the proposed action against at least the abstain alternative (abstaining is always the floor) and weighs each by four disciplines:

• The temporal trajectory brake. Every affected entity carries a measured net_trajectory (Improving, Stable, or Declining) derived from the persisted bitemporal memory of its substrate posture, not a guess the composer makes. A locally-net-positive action that draws further long-horizon harm from a measurably declining entity is refused. Where the history is too short to classify, the direction is InsufficientData and the composer degrades to the bare reflex, the same fail-open posture a fresh cell takes.
• The nested-scope roll-up. Each entity carries its enclosing topology scopes, so the action's long-horizon impact rolls up the topology. An action that is net-positive at every leaf but nets negative over an enclosing scope is refused. Local good that is regional harm does not pass.
• The world-model dissonance brake. A drift signal from the semantic world-model measures whether the action points off the coherent manifold of what the estate knows. At or above the brake point (a φ-proportioned threshold (≈ 0.618)) the action is treated as trajectory drift and held or refused.
• The operating-cycle branch. A LONG_CYCLE or MIXED entity deliberates fully. A SHORT_CYCLE (transactional) entity skips the trajectory reasoning and holds only on live dissonance, the long-cycle floors above having already bound the transaction. Weighting across the horizon is φ-proportioned.

The dispositions and their reasons

A verdict carries a disposition and the leg (the specific reason) that produced it, so an operator sees not just what the composer decided but why:

Disposition	When
Proceed	Clear, or a short-cycle transaction tolerated.
Hold	A strictly better eligible alternative exists (long cycle), or live dissonance on a short-cycle transaction.
Refuse	The reflex floor; floor-harm to a protected entity; a long-cycle net-negative; world-model dissonance; a declining-trajectory entity taking further long-horizon harm; or a nested enclosing scope netting negative.

The reason legs are pinned across runtimes: the cell pins the leg, not which particular scope first went negative, so an implementation detail is never mistaken for a parity break.

Two-runtime parity: the edge and the cell decide alike

The composer is pure logic, no I/O, deterministic on identical inputs, computed on-box, so it reaches the same disposition in a sovereign single cell and in a cloud fleet. That is precisely why it belongs on the edge at full strength rather than as a degraded client of the core: an air-gapped edge cell must reach the same disposition as a core cell for the same decision.

Parity is not asserted. It is re-derived, row by row, from a contract the cell generates from its production code, by a cross-language test that fails the build on any divergence. It is the same parity precedent the platform uses for the NPG gate and the subagent grant: state the contract once, and prove both runtimes make the same decision for the same input.

The two runtimes reach the same trajectory sight by different live feeds, the core from its DAO-backed entity scopes and temporal net-change, the edge from its on-device bitemporal substrate-state history written on the live agent path, and the shared composer they both feed is what the parity fixture pins.

What is live and what is planned

• Live, both runtimes: the reflex NPG gate; the foresight composer; the temporal-trajectory brake (the edge writes and reads its substrate-state history on the live agent path; the core is DAO-backed); the nested-scope roll-up refusal; the operating-cycle branch and φ-proportioned weighting; and the cross-language decision parity.
• Live on the core; planned at the edge (model boundary): the world-model dissonance brake. The edge composer runs and tolerates the signal's absence, the trajectory brake stays active, and the dissonance input lights up when the on-device model-serving embedder and the manifold-drift feed are bound. Until then the edge decides on trajectory, scope, and cycle alone.
• Planned (sensor / model boundary): the afferent sensor-fusion front that would feed live observations into the world-model. The fusion and signing machinery is built and parity-pinned; it activates on a sensor and embedder binding.

Related topics

• The Net Potential Gain (NPG) gate, the reflex stage the composer sits above.
• Substrate primitives & behavioral probes: operating mode, the coupling model, and the two-runtime mirror.
• Knowledge (bitemporal memory) the valid-time + recorded-time history the trajectory brake reads.
• Federated Semantics, the semantic world-model whose coherence the dissonance brake measures.
• The executive-functioning engine, the sense→decide→act loop the DECIDE organ sits inside.
• Core and the edge, the two-runtime discipline that keeps the edge at full strength.