Replication Without a Global Arbiter: The Replication Specification, the Attempt Register, and the Community Record Interface

When independent laboratories or readers attempt to reproduce, contest, or extend a published result, who decides which attempts count? This series has frozen single-lab execution (Papers 48-50: the data-to-verdict template, acceptance certificates, hash-deposited auditor exchanges, the file-drawer rule) and has legislated in-corpus revision after a typed defect is declared (Paper 65: four acts, the act-selection function, the empty rewrite cell). Between them lies an unbuilt layer: how attempts by multiple independent parties acquire standing — which submissions are replications at all, what an attempt warrants, who may say so, and where disagreement flows. The replication-crisis literature prescribes norms — preregistration, severity, disclosure — but ships no executable mechanism; standing is settled in practice by editorial fiat, prestige weighting, and undeclared aggregation (the reception of the Open Science Collaboration's 2015 coordinated estimate as headline scalar "replication rates" is the displayed case; Goodman, Fanelli and Ioannidis 2016 document that the field's own vocabulary is untyped). Into that gap steps the paper's foil, the Global Arbiter: the authority for whom "X replicated" / "X failed to replicate" is a decree rather than a typed, auditable record. The question is live, not hypothetical: the corpus's own execution seats (the 49-D/50-D data papers) wait for external experimenters by design. This paper — standalone, the first non-panel paper since Paper 26 — builds the community record interface and retires the arbiter role constructively. The Replication Specification ReplSpec declares, before any attempt exists, what "an attempt at leaf l" is: target address, grain, protocol identity, admissible-deviation classes, two-layer independence typing (Core independence fields — apparatus, raw-data acquisition, analysis chain, calibration protocol — as computational inputs, each carried as a graded triple with evidence grades evidenced / asserted / challenged / unknown; Disclosure fields — personnel, funding, institution — as annotation only, never decisive), evidence-backed replicator eligibility, and outcome vocabulary — every field with a declared, parser-checkable codomain, and a legacy target adapter deriving the target-side specification mechanically from already-deposited objects (the 49-D/50-D instantiation displayed). The question "which attempts count" is answered three times, at three decidable layers, never once by decree: schema admission (well-formedness), attempt-status typing, and profile contribution. The attempt register types every admitted record with AttemptStatus in {concordant, discordant, incomparable, defect-suspected, extension} under a fixed exclusive evaluation order — a typing of the attempt-target relation at the declared grain, never a truth verdict; incomparable is a first-class landing, and defect-suspicion is structurally separated from mere discordance (typed evidence requirements; discordance alone is never sufficient by rule; the suspicion route emits a declaration-ready packet for Paper 65's law — an input document, never the declaration act itself). Standing travels as the profile RepProf(l) = (A_l, mu_comp, mu_prov, mu_dev, mu_failure, OpenFlags), with declared codomains and an append-only pointwise update rule, and the paper's flagship lemma is the no-aggregation invariant: no mechanism clause contains a function collapsing RepProf into a single scalar replication verdict — proven as a type theorem. The replication authority layer A-rep closes with a boxed theorem — for every role r and leaf l, (r, decree-replication-standing, l) is not in A-rep — the arbiter cell structurally empty. Lemma A (arbiter-freedom) walks every clause: inputs are frozen records, declared attempt-record fields, or mechanism data — no fourth input anywhere; community consensus, journal uptake, and reviewer opinion enter as recorded events only, never as evaluation inputs. The ceiling is the point: no datum is scored, no effect size computed, no meta-analysis, no consensus machinery, no incentive design — each named-not-taken and declared in every profile instance by flag; Paper 64's realized landing is never consumed. The paper's worked instances are typed honestly: the corpus's auditor exchanges and external-review events are registry-grade rehearsals outside the register — auditor-grade, mock, typed non-evidence — assigned no status; the one end-to-end pass is a fully-instantiated synthetic example against 49-D's adapter-derived specification; and the corpus contains zero genuine cross-lab concordant or discordant instances, so those statuses are supported by independent constraints and four displayed barred cases, never by manufactured history. Strict recovery is exact. The registered slogans carry the result: "A profile is not a verdict; a register is not a court." — and, because attempt records do carry evidential weight as typed entries, "A replication profile can carry evidence without becoming an arbiter verdict." ReplStatus(machinery) = scoped at typing/mechanism grade; open by seat: data scoring, aggregation, consensus, institutional uptake, incentive design, community epistemology.