the September checkpoint is the current human-facing view

agent-decided, reversible. projects/normativity/legitimacy/checkpoint-2026-09-01/ becomes the canonical entry point for the normativity / answerability / legitimacy program: current theory with an audited dependency spine, a status ledger, a supersession map, open problems, a roadmap, and the reconciliation with the August 9 consolidation and the diachronic-answerability note.

It supersedes nothing. The August 9 consolidation remains authoritative for its own content and is reconciled rather than replaced; every research round remains untouched as provenance, including those later work corrected. What the checkpoint replaces is the practice of reading rounds chronologically to learn what the program holds.

Four judgments in it are the checkpoint's own rather than inherited, and each says so in the document that makes it.

The four-layer decomposition — fixed-era dynamics, diachronic legitimacy across revision, counterfactual non-capture, and the corrigibility consumer. The division is load-bearing because the fixed-era mathematics is strong enough to be mistaken for more than it is.

The dependency spine, audited rather than copied. Actionability sits beside service allocation, not after it: the scheduler cannot choose the coercivity modulus, which is a property of the response geometry and the enforcement compiler. And the normative demand and the force mechanism are separate objects the spine must show separately, because affordability is exactly the question of whether the second can meet the first, and conflating them makes that question unaskable.

The two theories meet at divergent attention. The conceptual answerability theory's service dichotomy ends at "attention diverges"; the service mathematics begins there. The note names its own missing piece — a reason-to-response structure plus a dynamic uptake condition — and that piece is Actionability plus Uptake. Their budgets disagree, the note's being renewable per stage and the mathematics' a consumable lifetime stock, and the disagreement resolves: the equivalence of persistence with eventual full service is the liability-priced version of the note's feasibility witness, and non-starvation survives at exactly the price liminf L_t(1) = 0.

The candidate legitimacy decomposition is a research framing, not canonical. Legitimacy = Diachronic Answerability + Affordability + Counterfactual non-capture is the best framing available, with two amendments: the first pillar must be read as including its semantic-authentication obligation, or semantic transport belongs to no pillar at all; and affordability is better read as a realizability side condition than a conjunct, since as a conjunct it invites the reading that a cheaper norm is more legitimate — which the round's own slogan, cheap enforcement is not always conforming enforcement, was written to forbid. What would make it canonical is a statement of the legitimacy predicate with a conclusion someone downstream wants, and no such statement exists.

Rejected alternative: rewriting the historical rounds so the current theory reads consistently. History stays history; the supersession map carries the current state.