Skip to content

native-event-tree-kernel-trace.md

Latest commit

 

History

History
112 lines (89 loc) · 6.87 KB

native-event-tree-kernel-trace.md

File metadata and controls

112 lines (89 loc) · 6.87 KB
Raw

Supplied-source trace — event-tree calculation kernel

This note records the Session 20 trace used to replace the generic event-tree helper with the calculation path actually reached by RDP5's event reconciliation. Only the supplied VB, DNA.dll, DNA5.dll, and manual sources were inspected. No alternate RDP implementation was consulted.

Installed-5.93 runtime correction. Later DLL-boundary traces show that the active TestMoveInTreeAlt call passes Reps = 0. The bootstrap-capable kernel described below remains present, but its resampling and collapse stages are bypassed for current RDP event trees. The desktop copies each raw Tree2ArrayP2 rank matrix directly into its nominal collapsed matrix. The current browser path follows that observed call rather than the older/commented ten-replicate call.

Active call path

The supplied bootstrap-capable desktop kernel enters Module3.bas::TestMoveInTree. For a positive replicate count it:

  1. prepares the selected event-tree sequences and the inside/outside or breakpoint-flank subalignment;
  2. calls MakeNJTreesP, which builds the unbootstrapped distance/tree matrices and invokes the bundled Clearcut implementation in traditional neighbour-joining mode;
  3. calls SEQBOOT2 to create a site-major weight matrix with the original data in replicate zero and resampled replicates in columns 1 through Reps;
  4. calls FastBootDistIP6(1, Reps, ...) to create single-precision Jukes–Cantor matrices for the resampled columns;
  5. calls TreeRepsP, which runs Clearcut for each bootstrap matrix and matches its groups against the base-tree groups;
  6. converts every node count with CLng((DLen + 1) / (Reps + 1) * 100), thereby retaining the base tree as a pseudocount;
  7. calls CollapseNodes with a cutoff of 50; and
  8. uses the raw and collapsed topology-distance matrices in the later phylogenetic-correlation and recombinant-role decisions.

The manual describes six event subalignments and bootstrapped NJ with collapse below 50%. That description and the older source path do not match the installed 5.93 active zero-replicate call.

There is a manual/source divergence worth making explicit. The later supplied VB path builds the whole inside/outside FAMat/SAMat through the bootstrapped tree kernel, but builds its four breakpoint DMatS inputs from MakeBPosLR(VSN=60) plus direct distances. The manual instead says that all six subalignments receive bootstrapped NJ trees and that each breakpoint flank spans 20 variable positions. RDP Web follows that documented six-tree workflow and applies the traced kernel to every family, using a separate 20-informative-site boundary pass; it does not silently reuse the 60-site distance-correlation flanks. Authorized golden comparison should therefore report both the current desktop matrices and the manual-intended six-tree result.

Observable numerical details retained

SEQBOOT2

SEQBOOT2 calls srand(seed), discards two rand() values before the draw loop, and then iterates sites outside replicates. The supplied Windows build therefore exposes the Microsoft CRT 15-bit rand() stream. Session 20 implements that LCG directly. It keeps replicate zero at weight one for every site and uses the same configured BSRndNumSeed for each regional call, as the desktop does.

FastBootDistIP6

Only coordinates where both states are non-missing contribute. A pair with at least one comparable site uses single-precision Jukes–Cantor correction; identity at or below 0.25, or no comparable sites, produces the source saturation distance 10. The earlier browser helper incorrectly imposed a ten-comparable-site floor on this event-tree path.

Bundled Clearcut

The port retains the observable behavior of the supplied Clearcut copy rather than substituting a generic double-precision NJ routine:

  • matrices and row transforms are single precision;
  • the minimum transformed-distance scan is ordered and uses strict <, retaining the first tie;
  • the final initial r2 slot remains zero-initialized on the first join, matching NJ_init_r;
  • the packed-matrix collapse order follows NJ_decompose/NJ_collapse;
  • the final two clades each receive the remaining distance from the NJ_LAST path; and
  • the Newick path suppresses a negative branch sign and truncates the magnitude to five decimals before Tree2ArrayP reads it back.

The graph endpoint exposes those serialized branch lengths. They are useful for a faithful visual tree, but they are not the matrix values used by RDP's analytical role decisions.

Midpoint-ultrametric ranked analytical distances

Tree2ArrayP parses branches at four fractional digits, clamps them to [0, 1], and first builds leaf-to-leaf path lengths. TreeMidP selects the first maximum-distance pair, locates its midpoint, and UltraTreeDistP lengthens each terminal path so all leaves are equally distant from that root. MakeTreeArrayXP2 rounds the resulting ultrametric distances to four decimals, sorts distinct levels ascending, and replaces each distance with rank / 1000. CollapseNodesXP3 does not rerank the matrix: it promotes a weak node's pairs to an existing parent-node rank, retaining any gaps in the original rank scale. Session 20 now supplies those raw and collapsed rank matrices to TreePhPr, CollapsedTreePhPr, TreeSubPhPr, TreeSubDist, triplet-tree scoring, phylogenetic-correlation set construction, and the existing downstream consensus/recheck stages.

The supplied Newick writer omits a length after one internal root child, while TreeToArrayP scans forward for the next decimal without checking that it belongs to that node. Reproducing that out-of-token read would make an internal edge depend on the first numeric token in the opposite subtree. Session 20 deliberately repairs this handoff: every actual Clearcut edge is serialized, parsed at the source's four-digit precision, clamped, and then passed through the same mathematical midpoint/ultrametric/rank transformation. This can differ from a desktop result that triggers the parser defect; it is an explicit numerical-correctness fix, not hidden parity drift.

Deliberate browser bounds

The existing event-tree panel cap remains 100 working sequences and the current event workflow uses zero bootstrap replicates. The six matrices are built once per reconciled event and reused across all downstream role metrics. The lazy graphical endpoint transfers only the already-saved bounded edge lists; it does not rebuild a tree or copy alignment rows.

Validation boundary

The linked host regression fixes the seed-3 SEQBOOT2 weight stream, matrix dimensions, rooted Clearcut-shaped node/edge counts, five-decimal branches, the allowed pseudocount support percentages, 50% collapse metadata, rank/1000 analytical matrices, determinism, and the one-comparable-site rule. It does not claim native-vs-WASM golden parity: exact supplied-binary fixtures remain an authorized validation phase.