ADR 016: Target-Based Pact Formation

Date: 2026-07-29 Status: Accepted

Context

Pact formation was specified as an equilibrium-seeking process: a "comfort condition" evaluated as a Poisson-binomial convolution over 24 hourly availability windows, a PACT_FLOOR/PACT_CEILING band (12–40), a marginal-value calculation, and a six-state formation machine with hysteresis. The complexity review (§Simplify-3) and the game-theory review (§6) both found this far heavier than the result justifies — simulations already showed 94–98% availability under much simpler formation rules — and the formation machine's states (Growing, Comfortable, Over-Provisioned…) shadowed the narrative's own Bootstrap→Hybrid→Sovereign arc with a second, differently-named lifecycle.

Decision

Replace equilibrium-seeking with a target-based rule:

  • Form pacts until ~20 active (floor 12); replace on failure.
  • Prefer partners in different timezones and a mix of Keepers and Witnesses (Keeper ratio ≥15%) — stated as a heuristic, not an optimization to converge on.
  • No comfort-condition math, no ceiling-seeking, no marginal-value computation.

Dissolve the formation state machine into the three-phase adoption arc. The arc is the formation lifecycle: Bootstrap (0–5 pacts, relay-primary) → Hybrid (5–15, mixed) → Sovereign (15+, peer-primary). Only two lifecycles remain protocol-wide: this arc and the four-state pact FSM (ADR 015).

20 is retained as the modeling constant so existing simulation evidence stays valid.

Consequences

Positive:

  • Removes the most mathematically elaborate, least-load-bearing component of the spec.
  • The user-facing maturation story (the narrative spine) becomes the only lifecycle a reader must hold — no shadow machine.
  • Keeps the one narrative-bearing insight from the old model — Keepers accept pacts beyond their own comfort so their excess capacity carries Witnesses — as prose in the incentives doc.

Negative:

  • A flat target is less adaptive than an equilibrium model in principle. In practice the simulator shows the adaptivity bought little, and diversity heuristics recover most of it.

Neutral:

  • PACT_FLOOR = 12 survives as the floor; PACT_CEILING and the comfort condition are removed.