The seam face of the Radiant type-theory work — the Seam Trilemma, explored build-and-break.
Where Corona reduces an invariant to a compile-time error and
Sol proves the reduction sound, Spicule stress-tests the seams:
it constructs a seam, tries to break it, and records what survives. The
37-seam build-and-break pass is closed; the writeup lives in PAPER.md
(Tier 1+2). The monoculture control has run under TWO foreign families
(2026-08-12, GPT-5.6 Sol 6/8 and Gemini 3.1 Pro 5/8 outcome-class agreement,
each as every cold role): all four BREAKS reproduced by both with horn-exact
labels (all four share the modal horn; COSTS agreement cells are horn-inexact).
On the full-sample runs the HOLDS demotions were three horn-3 cost cells plus
one horn-1 spec-wording cell; an earlier single-seam run ruled the flagship
HOLD a BREAKS outright and was later vindicated: the v1 spec never pinned
variance — a covariance laundering path found by two families (one attested by
the disclosed session record) — nor dropck ownership (found in-family at v4),
so the recorded HOLD is contradicted as stated; the v2–v6 amendment lineage (design_seam/FOREIGN_REPLICATION.md)
ran v5 to three independent panels (COSTS [horn2] · COSTS [horn1] · BREAKS
[horn1]; fourth surface component: input provenance) and the v6 respec (input
eliminated) to three pre-registered panels (BREAKS · COSTS · BREAKS; fifth
component, latent since v1: §3's unbounded Consumer<T>: Send vs §4's bounded
stopper — the RUSTSEC-2020-0150 class). Five distinct discoveries in six specs; twelve
post-v1 panels (9 in-family + 3 foreign) never rated the flagship clean; the
clean HOLD is not recovered, the realizability table has no all-holds row, and
the lineage is closed under its pre-registered stop rule (owner-reopenable).
What the instrument left standing: the negative results (eleven broken seams,
replicated horn-exactly by two foreign families where sampled), the trilemma
and its nine-axis cost basis, the three-domain grading, and the three-streak
holds outside the flagship (spin_lock, drain, total_div-at-COSTS). The Gemini refined escalation was
re-examined line-by-line and sustained at the recorded COSTS [horn2] (the
claimed UB path targets an unvended configuration; the out-of-family
corroboration is a disclosed single draw) —
design_seam/REFINED_ESCALATION_ADJUDICATION.md. Standing caveats:
33 of 37 seam verdicts are single-gauntlet (only SPSC-v1, spin_lock,
drain, total_div ran the three-streak bar), transport deltas on the
foreign runs, and the v5/v6 tier distributions themselves.
seam/,design*/,proof/— the seam constructions and their analysis.runs*/— build-and-break run records.PAPER.md,WRITEUP.md,TUTORIAL.md,THE-SAME-WALL.md— the writeups.orchestrator/— run tooling.
Public research repo (made public 2026-08-13, after the publish-gate cold round
recorded in DEVLOG-adjacent project records). The paper is a repo-native
artifact; nothing here is submitted to any venue.
MIT. Copyright (c) ModelMiser. The full text is in LICENSE.