{
  "version": "blind-lattice-prime35-v1",
  "master": "../../docs/basic-tiling-algorithm.md",
  "status": "Declared finite graph-path adaptation; no global space-time tiling or reference neighboring-pair/corona conformance is claimed.",
  "objective": "Find a checked lattice move path from an unlabeled query embedding to any unlabeled bank embedding. Attach an evaluator label only after certificate replay. All finite failures remain unknown.",
  "inputs": "All 35 prime corpus representatives with three through eight crossings. One independently eight-move-scrambled bank embedding per type and three independent eight-move query scrambles per type. Fixed random bank and query permutations, no resampling or replacement. Only vertex arrays enter the classifier. Labels, crossing counts, seeds, provenance and correct reference indices stay in the evaluator.",
  "inventory": "Every incident unit coordinate face throughout the lattice knot; no patch bounds. The existing fixed projection (11,13,17), endpoint Reidemeister/routing grammar, arithmetic scope and maximum 240 intermediate unit edges remain restrictions. No general lattice-isotopy completeness claim.",
  "pointDomain": "A complete root reference-choice graph has one obligation; all 35 references assign unit t. Proved failed-horizon reference candidates assign incompatible m at root-budget. Every expanded path state likewise has one complete choice obligation with unit t for every raw candidate; author legality and validated capacity/extension markings determine complete forward/reverse incidence. No forced/dead claim precedes complete enumeration.",
  "scheduler": "Fair round-robin bank scheduling resumes each reference's depth-first search for one complete expansion per yield. Finite contours increase from zero to sixteen, except shorter contours proved impossible by root markings. Exact opaque-key depth dominance and persistent recursive stacks preserve bounded reachability. Minimum prefix depth dominates larger prefix depth in the memoryless graph; cycle deletion cannot destroy finite reachability. This specialized scheduler is explicitly different from reference tiling generation ordering.",
  "budgets": "Matched aggregate query budgets in both lanes: ten seconds including complete root enumeration, all failed references and contours, probes, activation/transfer checks, inventory warmup and JS positive-certificate replay; 5,000 expansions; 100,000 simultaneous visited states; sixteen move steps; 240 unit edges. Time may exceed the boundary by one atomic author/proof operation. Deadline-interrupted verification leaves unknown and retains an unscored candidate movie.",
  "rootMarkings": "The existing relaxed cover learner is additionally applied to whole finite source-target discrepancy components, up to 480 edges, with horizons zero through sixteen and 64 probe nodes. Every face touching the window is present, regardless of geometry or legality. Covering every discrepancy edge is necessary for a real path. Negative complete cover probes prove a minimum move requirement; positive relaxed covers are not knot witnesses. Components compose only when no face touches two components. This global root-window bound is distinct from local child markings and cannot prove knot inequivalence.",
  "localMarkings": "Child extension supports remain at most 48 edges with 512-node probes. Failed budgets become explicit equality-mismatch point values. Vertex capacity profiles are fully probed. Immutable native extension catalogs validate full face alphabets, resolved labels and stencils before activation; translation reuse uses an O(1) native catalog gate. Whole-window root and local-child catalogs are separate and both start empty. No saved model, knot invariant or construction route is imported.",
  "versionAndRollback": "Capacity version changes replay every retained accepted-prefix/domain graph. Extension profiles and stencils are immutable after activation and complete new context incidence is constructed before use. Resumable DFS retains exact prefix stacks; completed contours release their visited-state stores. No unproved or branch-local hypothesis is used as a hard exclusion.",
  "coverProbeControl": "After the full incident-face alphabet is enumerated, a generic union-capacity bound may prove failure before building incidence for branching. It makes no degree-zero or singleton claim on a partial graph. Otherwise complete point/candidate incidence precedes global dead checks and same-generation minimum-degree branching.",
  "comparison": "Sequential fresh marked and unmarked workers process the same 105 queries and permuted 35-reference bank. Identical DFS driver, neutral candidate seed, query-local exact diagram memoization and aggregate bounds. Marked catalogs persist only within that lane across mixed knot types. Report query subtotals, cold initialization, artifact overhead and elapsed lane time, nodes, branching/pruning counters, peak RSS, success/incorrect/unknown by type and seed. Common scramble preparation and subsequent independent proof replay are separately reported. Unknown brute-force results do not support an equal-success speedup ratio.",
  "proofArtifacts": "Gzip JSONL archive retains each new root, local or capacity proof once with channel, query index, monotone event index and SHA256 signature digest. SHA256 identifies evidence, not knot type. Independent Python rebuilds full face alphabets, replays every resolved cover label, checks root discrepancy composition, bank/seed coverage, all scramble histories and all scored spatial paths. No JS solver or learner is used by the verifier.",
  "evidence": ["../../tests/test_knot_blind.mjs: 384 exhaustive generic graph/horizon reachability checks, resumability, full-bank late target, labels ignored and budget unknowns", "lattice-blind-benchmark.json: every query, reference, cost and positive certificate", "marked-proofs.jsonl.gz: every new online root/local/capacity proof", "../../scripts/verify-blind-lattice-knots.py: independent complete experiment replay"],
  "gaps": ["Eight sampled moves need not be a minimum distance or a difficult random knot embedding", "The bank contains known close scrambled representatives; success is empirical closed-bank recognition, not arbitrary-input classification", "No separate mirror, composite, unknot, outside-bank or negative-pair score", "No RL or learned cluster proposal system", "General lattice isotopy, continuous projected events, unbounded termination and twelve-crossing classification remain open", "Global root marking benefits must not be attributed solely to local transferable markings"]
}