Most agent design treats identity as unification. One persona, one voice, one system prompt that smooths over the seams between audience and operator and lens. Stacks that unify get smoother but flatter. The output is consistent. It carries no edge. There is nothing for a downstream decision to grip against because every layer was sanded into the same surface before the system saw the request.
Forge separates instead. The constitutional surface treats audience as one identity, operator as another, lens as a third, scope as a fourth. The /tom mechanic does not generate content. It refuses to generate before the separations are made. Audience is not lens is not scope is not register. Each refusal is a meta loop closing — the system observing itself about to flatten and stopping.
The recovery protocol against context decay reads correctly under this frame. Selective resurfacing is the refusal to treat one component as the whole goal. Scope narrowing is the refusal to treat one question as all questions. Forced self-generation through stakeholder framing is the refusal to treat audience need as architectural specification. Recency-bias positioning is the refusal to let new constraint blur into old context. Perspective forcing is the refusal to merge lenses. Artifact-bound verification is the refusal to treat retrieved understanding and generated understanding as the same thing. Each move is a fold. The success of the protocol is not stylistic. It is the mechanical consequence of refusing to conflate.
Sabeel's working note from this week is the field evidence. /tom in operational use produces what he names the posture-proposal mechanic. The system reads source-of-truth plus conventions plus audience profile, reasons about what posture fits, proposes with rationale, and waits for the go. In the same run it surfaces preserved boundaries unprompted — vocabulary restrictions, scope limits, structural requirements — at proposal time, not at review. Those boundaries were not generated. They were observed. The conventions file separated from the source pack separated from the audience profile, and the separations were what made the boundaries visible. A unified-identity system would not have surfaced them, because there would have been nothing to surface them against.
The disambiguation matters, and it matters here for a reason. /tom-the-skill is one operative expression. The substrate primitives — selective resurfacing, scope narrowing, forced self-generation, recency exploitation, perspective forcing, artifact-bound verification — are the load-bearing claim. That distinction is itself a fold. Collapsing it would soften the architectural argument; one would slip between making a claim about a feature and making a claim about a substrate. Holding them apart sharpens both. The act of separation produces the edge.
A folded substrate is sharper than a homogenized one. Each separation distributes the load and tightens the operating tolerance. The compounding is structural; it does not require the substrate to grow, only to fold. That is why governance compounds where capability plateaus. Capability scales by accumulation. Governance scales by the multiplication of distinct-but-bound identities operating against each other.
The constitutional layer governs what the system is permitted to mean. Meaning requires identity boundaries. Identity boundaries are the fold. /tom is one fold among many. There will be more, because the substrate is not finished folding.