For anyone building on this candidate (vscode extension, TS library, validator, codegen): this is an honest map of where the candidate is most provisional and what would harden each point. Every decision carries a mizan-gated confidence ceiling; the whole ledger is
burhan-converge-clean and perturbation-robust. Treat the @0.5–0.55 decisions below as the most likely to change — design tooling so they are isolated/configurable.
| Ceiling band | Count | Meaning |
|---|---|---|
| @0.85 | ~59 | Strong: multi-aspect inherited backbone (collections-are-maps, by-name identity, dialect, paths-are-uriTemplate) |
| @0.6–0.8 | ~95 | Solid candidate decisions, sole-witness-capped (+ C030 router-delegation @0.76) |
| @0.5–0.58 | ~40 | Softest — contested shapes, Originated inventions (incl. the whole agent layer C027–C030 + cost/jobs C024–C026), provisional tooling defaults (below) |
Audit (mizan_verify_claim / burhan-converge, 2026-06-13): converge-clean, 251 claims, 7 load-bearing all witnessed, 0 contradictions. The soft claims are sole-witness, bcmea-clean, declared ≤ recommended cap — honestly capped, not inflated. They lift only with an independent witness (a second corroborating source) or SIG ratification. The agent layer (below) is the newest + most Originated band — there is no SIG prior for it at all.
| Decision | Ceiling | What would harden it |
|---|---|---|
| No declared signature-style indicator (#16) — overrides earth2marsh's stub | 0.5 | SIG ratification of the signature mechanism, or a 2nd voice corroborating uniform/implicit |
PR#183 declared signature array → optional (#16 D2) |
0.5 | the SIG merging or rejecting PR#183 |
| Tags flip array→map (#83) | 0.5 | the #67 tags-upgrade proposal landing |
| Links scoped baseline (#58), server hints (#56), order field (#20 subq4) | 0.5 | dedicated panels / SIG signal — these are thin single-voice areas |
$ref pointer-vs-name policy, inline-restriction (#73) |
0.5 | the deferred referencing fragment grammar (#26/#49/#72) |
Path inheritance shared (#116) |
0.55 | a 2nd witness; the proposal used pathItem-level params directly |
| Responses stay named-map (#83) | 0.55 | corroboration; the array camp (kevinswiber et al.) was real |
| D1: JSON-Schema out of the static matcher | 0.55 (mizan floor 0.139) | the weakest load-bearing guardrail — an independent witness, or the SIG ratifying JSON-Schema-in-matcher would overturn it |
| Callbacks/webhooks (#upgrade gap, C018) | 0.55 | Originated; SIG design or async-spec alignment |
| Tooling defaults (C019: reference/signature/deserialization grammars) | 0.55 | the deferred grammars ratifying (#26/#49/#72, C003, #100/#108) |
| Upgrade coverage 55–70% (C017) | 0.5 | building the actual transformer + measuring on a real corpus |
Cost-on-event + x-suluk-jobs (C024–C026, §15) |
0.58 | Originated vendor extension; a real fleet exercising attribution + reconciliation (the costCeiling enforcement runtime is RESERVED) |
x-suluk-agents composition (C027, §16) |
0.52 | Originated, no SIG prior; a 2nd real or nested agent beyond Conin (the recursion reopen-trigger) — today: cycle/depth-lint + schema slot only |
x-suluk-policy governance (C028, §17) |
0.52 | Originated; the costCeiling enforcement runtime + a governed fleet; the determinism-gate mandate is OPEN |
Thinking bound (C029, §17) — {maxRounds,budget}, no stopCondition |
0.55 | Originated bound; a 2nd agent that needs richer within-agent control would reopen the vocabulary |
| Model resolution + ZDR-router (C030, §17) | 0.72 | live-probe-verified (2026-06-13) that provider.zdr+openrouter/auto combine; an independent witness on strict non-ZDR-endpoint exclusion, + operator-governed-ZDR (needs a policy retention vocab) |
shared mechanism are the ones a future ratification is most likely to change. The C019 Appendix A defaults are interoperable today, but pin your tool's version.x-suluk-agents/x-suluk-policy, §16–§17) is the newest, most Originated surface — no SIG prior at all. The composition shape (skills / routes-are-$refs / cycle-linted sub-agents) is buildable today, but isolate the soft edges: recursion is one-hop-only (cycle/depth-lint + schema slot, no runtime machinery), costCeiling is declared-not-enforced, the thinking bound has no stopCondition, and model selection delegates to @suluk/models (NEEDS not frozen ids). Crucially the D1 invariant extends here: the request matcher never reads an x-suluk-agents / modelResolve / x-suluk-policy field — selection and governance are a runtime-advisory overlay, never part of routing.