{
  "version": "blind-lattice-scaling-v1",
  "master": "../../docs/basic-tiling-algorithm.md",
  "status": "Declared graph-path adaptation. No global space-time tiling, reference neighboring-pair/corona implementation, general knot-equivalence decision procedure or RL claim.",
  "selection": "Each retry includes every marked unknown from its immutable parent, unchanged, against its unchanged bank. Both lanes run those same selected inputs at identical larger aggregate budgets; all earlier lane, preparation and audit costs remain reported. Scaling predeclares first and last corpus types at each crossing count 9 through 12, three independent eight-move query scrambles each, against the complete 2977-prime bank. Those eight bank references are independently eight-move scrambled; all remaining bank references are canonical. No case rejection or replacement.",
  "blindness": "Only vertex arrays enter the classifier. Reference names, crossing counts, target indices, scramble seeds and histories stay in the evaluator. Exact embedding equality is the only positive search goal. Labels are assigned after a checked move certificate.",
  "inventory": "All incident unit coordinate faces throughout each exact lattice knot, with no patch region. Same fixed projection, single Reidemeister/routing endpoint grammar, arithmetic scope, sixteen-move horizon and 240-unit-edge bound. Unrestricted move location does not establish completeness for arbitrary lattice isotopy.",
  "pointGraph": "Full-bank reference-choice point/candidate graph before search; every expanded move domain completely authored before dead/forced/branch decisions. Unit t-values cover one choice obligation. Proved cover-budget and capacity m-values eliminate through equality mismatch and complete forward/reverse incidence. Partial domains yield unknown.",
  "scheduler": "Both lanes use the same fair round-robin resumable generic iterative depth-first search. Root redundant markings may skip proved impossible shorter contours. Exact depth dominance and recursive prefix rollback preserve bounded graph reachability. This specialized path scheduler is distinct from the reference generation-first volume-tiling scheduler.",
  "proofs": "Complete touching-face relaxed cover probes supply negative necessary continuation bounds and explicit failed-budget m-values. Positive relaxed covers do not certify lattice paths. Unknown probes never prune. Whole-discrepancy root supports and at-most-48-edge local child supports remain distinct. Every resolved label is replayed before immutable native activation, with constant-time native transfer gate and no saved model import. Catalogs persist across mixed query types only within the marked worker.",
  "memoization": "New exact optimizations union the complete face incidence once while preserving component/edge/mask order, count BigInt populations through exact hexadecimal digits, memoize target-neutral complete author domains and resolutions within each query (64-domain bound), and memoize immutable target-conditioned extension contexts (4096-context bound). Cache keys retain exact embedding, direction, region and edge cap. Cached absence comes only from complete author resolution. Both lanes receive the same author-domain cache; only the marked lane has extension contexts. No candidate shortlist, knot invariant, preferred move type or approximate key.",
  "budgets": "Each stage records its identical per-query time/node/live-state bounds in both lanes. Every root, wrong reference, failed contour, proof, activation/transfer check, cache warmup and JS certificate replay is inside query time. Cold construction and archival/checkpoints are inside lane totals. Atomic work can slightly overrun a time boundary. Deadline-interrupted replay retains only an unscored candidate certificate and unknown.",
  "verification": "Independent Python replays full scramble and identification movies, rebuilds every relaxed alphabet, checks every resolved label, complete bank and fixed seed coverage, parent hash and exact retry identity, complete root incidence/composition, catalog continuity, archive hash and node/time totals. Search expansions and cheaper proof nodes are separate units. All costs are reported; no equal-success speedup is inferred when brute force remains unknown.",
  "evidence": [
    "../../tests/test_knot_blind.mjs",
    "../../tests/test_knot_blind_memo.mjs",
    "../../scripts/verify-blind-lattice-knots.py",
    "lattice-blind-scaling-summary.json"
  ],
  "gaps": [
    "Representative queries against a complete twelve-crossing bank are not exhaustive twelve-crossing query coverage",
    "Sampled moves need not equal minimum distance",
    "Close constructed embeddings do not establish arbitrary-input classification",
    "No separate mirrors, composites, unknot, outside-bank or negative-pair score",
    "No RL or cluster proposal implementation",
    "Global root-window benefits cannot be attributed solely to local transferable markings"
  ],
  "wholeWindowContinuation": "Stages with wholeWindowChildren=true also use the native whole-window root catalog on child discrepancy components larger than the local 48-edge support. For each face-disconnected component the larger of two proved bounds may be used; these per-component bounds sum only across disjoint face alphabets. Root probe limits remain 480 support edges, sixteen horizon and 64 nodes. This is an explicitly broader necessary continuation constraint, not solely local-transfer benefit. It preserves all paths within the declared horizon and keeps unknown probes unconstrained. All new whole-window proofs and costs remain in the root archive channel.",
  "neutralPortfolio": "Stages declaring searchSeeds use fair resumable DFS contexts for each reference and neutral move-order seed, with one root candidate per reference and aggregate budgets across every context. Both lanes use identical seeds and schedule, share only exact author caches, and the marked lane shares its validated catalogs. No knot-specific ordering or seed is selected from an input label. A positive path from any context is independently checked. No per-ordering budget is hidden."
}