# Sixteen perspectives: the directions worth a floor cycle, as the editor reads the two rounds **Type:** record · **Date:** 20 September 2026 · **Status:** the editor's shortlist, amended on 20 September 2026 after an outside read by GPT-6 Astra (`31_astra_check.md`; every change listed at the end); the fifty-six outputs are the record (`work/opus/`, `work/astra/`, `work/r2/fable/`, `work/r2/opus/`, `work/r2/astra/`); nothing here attaches to a claim; every item is a proposal for the author · **Authors and models:** as `20_comparison.md` · **Standard:** the judges' top-five picks and the editor's reading, kept apart · **Cost:** see `20_comparison.md` ## What the judges picked Every pick with its lineage resolved (label, then who wrote it). - **Claude Opus 5 judge:** PHY-A (Opus) missing analysis, P1's falsifier re-aimed; PHI-C (Opus) worked direction, S1's second falsifier run on the page; PRE-C (Opus) worked direction, the decrement audit and a blind grading round; CAT-A (Fable) worked direction, "exactly two" typed as variances; QG-A (Opus) first step, the ordering fraction computed. - **Claude Fable 5.1 judge:** QG-A (Opus) first step; PHI-A (Fable) first step, the premise audit of three canonical arguments; PRE-C (Opus); PHY-A (Opus); CAT-C (Opus) first step, the admissible self-map class. - **GPT-6 Astra judge:** PHI-B (Astra) first step; CAT-B (Astra) first step, a thin-category obstruction; QG-C (Astra) first step, a countermodel on clock statistics; SYN-A (Astra); PRE-A (Astra). The two Claude judges share three picks (QG-A, PHY-A, PRE-C). The Astra judge shares none with them. Read `20_comparison.md` before weighing that. ## The editor's shortlist, in the order I would put them to the floor **1. P1's falsifier is aimed the wrong way, and the correction costs nothing.** Five outputs on three lineages found that causal set theory (Bombelli, Lee, Meyer and Sorkin 1987, "Space-time as a causal set") already is "an account of non-fundamental spacetime that assigns no role to perspective", so P1's published falsifier is met on its face and the ledger has not registered it. The Opus physics referee goes further: the live danger to P1 is the opposite class, accounts that give perspective a large calculable role with no self-reference (Unruh 1976; Rovelli 1996; Giacomini, Castro-Ruiz and Brukner 2019). Route: a correction through the site's machinery, re-aiming the falsifier and recording that the old one fired. Level 3, the author's, since it moves a claim node's falsifier. Cost: an outside read, one Codex call. **2. S1's strong reading may rest on P2, and its own second falsifier may already be met.** Two independent audits reached the same place from different sides. Fable's philosophy referee fixed a schema (separability against housing) before checking, marked eleven premises across Chalmers 1996 *The Conscious Mind*, Jackson 1982 "Epiphenomenal qualia" and Levine 1983 "Materialism and qualia: the explanatory gap", and found separability in three premises and housing in none: the diagnosis denies a premise rather than exposing one, and the denial is principled only through P2. Opus's philosophy referee ran S1's published second falsifier ("the hard problem survives careful restatement with no containment assumption") on the conceivability argument and reports that the restatement passes premise by premise, with the diagnosis owing a further premise it does not state. Fable's Part I auditor stated the diagnosis as a three-condition rule and found it lands on every further-fact problem, induction included, unless a fourth condition is added that is a word for place. The Opus judge notes the restatement covers only the conceivability form, not Levine's explanatory-gap form, which Fable's audit shows untouched by Part I. Correction at the outside read: the claims graph's confidence policy, which the round-two packet omitted, prices S1's 0.8 on the weak reading only and holds the strong reading "pending the Part I production-presupposes-containment argument". So there is no demotion to argue; what the audits show is that the pending argument is not supplied and that the defence Part I offers runs through P2. Astra adds that P2 is the page's offered defence, not the only conceivable one, that Fable's K2 and L2 steps do not by themselves establish separability, and that Opus's vocabulary ban cannot establish the absence of a conceptual assumption. Route, rebuilt: one explicit reconstruction of the strong reading's argument and its strongest reply, combining Fable PHI-A's argument-by-argument method with Astra PHI-B's distinction between asking whether separation is possible and asserting it, checked against the canonical sources; then objection 17 if the argument is still missing. Level 3, reputational: the author's sixth essay says "the diagnosis sits at 0.8", which the policy reads as the weak reading, and the essay should say so. **3. The correction machinery has never lowered a number.** Opus's pre-mortem audited every disconfirming event in the record: eight events, zero tier changes, zero numeric changes; responses were renames, prose withdrawals, status words, sensors added, scope changes. Its proposal: a decrement rule with a named actor, a date and a trigger, and a USD 500 blind grading round on three lineages with the adoption rule fixed before the values are seen, "the one purchase in the packet that could make a confidence number fall". Fable's pre-mortem wrote a three-clause stopping rule and found that this review round is the kind of spend its third clause forbids. Both Claude judges put the Opus piece in their top five. Amended at the outside read: a deadline or an obligation to move a number is not evidence against a claim, so the compulsory decrement is dropped; what is kept is the audit, strengthened by Astra's history check (no tier lowered since 29 May 2026, no number changed since 2 July 2026 across fifteen versions of the claims graph), and a rule that each relevant result gets a dated author decision saying what it changes, under an explicit transfer rule. Astra PRE-A's separation of reasons to keep funding from reasons to believe is added. Level 3. **4. One panel, many hinges: the cheapest instrument the record has.** Fable's synthesis designer widened the persistence conjecture's own panel to all four criteria plus a partition column and a bearer column, filled the cells the packet already fixes, and pre-read ten outcomes against named hinges (M1's falsifier, objections 8, 10, 15, 16, the persistence commitments, M2). Its reliability precondition comes first: two lineages fill the cells blind, and if they disagree on more than one cell in five, the criteria are not yet an instrument and every criterion-based result inherits that. Under USD 20 for the first pass; under USD 500 to build three rows as running code and exhibit a bearer flip by interrupting and migrating a system. Amended at the outside read: add Astra SYN-A's requirement that boundary, bearer and orientation definitions stay fixed across every answer, and read a row with no partition supplied as undetermined, not as evidence of independence. Route: floor cycle 4's test at the desk tier first, pre-registered, cap USD 20; the USD 500 build only on a pass. Level 2. **5. A number for the physics arm.** Opus's phenomenologist computed the ordering fraction of the fold's own collapsed occurrence order as a causal set for every architecture the record names and found the one-dimensional value throughout: the order carrying continuity is all order and no extension. Both Claude judges checked the arithmetic and put it in their top five; the Astra judge disputes the "every architecture" claim with a layered family that approaches dimension two. Fable's phenomenologist derived a clock bound from the continuity criterion (Margolus and Levitin 1998, "The maximum speed of dynamical evolution") and reported it eight orders of magnitude too coarse to see, a negative result at the prominence of a success; the Astra judge's objections to the derivation (summing bounds across occurrences; switching energy against mean energy) are recorded and unanswered. Amended at the outside read: the universal one-dimensional claim is contradicted inside the Opus output itself (r = 8/15 for eight slots and two layers), and Astra QG-C's countermodel shows why a fixed total duration leaves interval contrast free; neither supplies physical geometry. Route: a bounded derivation audit at the desk tier, both pieces together, under USD 20; Fable's clock bound has not earned an implementation study. Level 2. **6. The count is a typing question.** Fable's category theorist typed "exactly two" as the number of variances in which a self-representing object occurs in its own type (two at dimension one, one for streams, four for a 2-cell), which is the demonstration objection 15's reply asks for; Opus's category theorist relocated the unspecified parameter from the count to the class of admissible self-maps, which would constrain criterion 3 and supply the persistence conjecture's open anticipation rule. The two are complementary and the judges split between them. Amended at the outside read: Fable's variance count has not derived two orientations (the variance of A in A to Y^A is negative in both positions under ordinary type variance); Opus's admissible-map class does not automatically become the anticipation rule, and a retract does not establish self-maintenance; Astra CAT-B's obstruction is sound but against a bridge the paper never adopted. Route: a translation test first, specifying the intended richer representation and its connection to the four criteria, before any formalisation is paid for. Level 2. **7. The lab needs a judge outside the maker lineages.** Not a direction from the review but its clearest finding about the lab itself: three judges, three lineages, perfect partisanship. Amended at the outside read: a fourth model agreeing with one side would not settle bias. First reconcile the particular mathematical and interpretive disagreements under one rubric at desk cost; any later blinded comparison is balanced across provider families, includes specialist checks, and preserves the original result. Level 4 for a key or a person; level 2 for the reconciliation. ## What the editor left off the list, and why Astra's directions were the tidiest and, by the two Claude judges, the least new: finite controllers, register models, and obstructions against bridges the record never proposed. Two are worth keeping in view: QG-C's countermodel that qualitative recurrence alone cannot constrain a calibrated clock statistic, which sharpens the Fable derivation's weakness, and PHI-B's point that asking whether separation is possible is not asserting it. Round one's ninety-six directions stand in `work/` with the Astra judge's clusters; the editor did not re-read them for this list. ## What would count against this - The shortlist being the editor's taste dressed as synthesis: the judges' picks are listed first and separately, and every item names which judges chose it. - Item 2 being wrong about the site: the audits worked from the paper's served text; if Part I already states the further premise the auditors say is missing, the objection fails and the record should say so at the same prominence. - Item 1 being a misreading of P1's falsifier: if "assign no role to perspective" was meant in a sense causal set theory does not satisfy, the falsifier needs its sense stated, which is the same correction by another route. ## Changes after the outside read (20 September 2026) GPT-6 Astra read the shortlist with lineages known and answered the editor's three questions (`31_astra_check.md`). Its verdict on the list: Claude-heavy and overstated in places, with sixteen of twenty-four candidates Claude-made, so the count alone is not skew; the skew was in accepting Claude conclusions too readily and missing Astra qualifications. Every change above is marked "amended at the outside read". In sum: item 1 kept with a caveat (packet silence is not a negative result; keep the old falsifier's outcome on record beside any new wording); item 2 rebuilt and its demotion claim withdrawn, on a confidence policy the editor's packet omitted; item 3 keeps the audit and drops the compulsory decrement; item 4 gains fixed definitions and "undetermined"; item 5 becomes a derivation audit rather than a physics result; item 6 becomes a translation test; item 7 drops the claim that another vote settles bias. The editor's earlier line that Astra's directions were "the tidiest and the least new" stands as the editor's reading and is not a finding.