Architecture

ChromoSort is a reference-guided genome assembly curation toolkit. It consumes existing whole-genome alignments, assembly graph records, and optional cross-evidence tables; it does not assemble reads, run aligners, polish bases, or infer variants. Its core responsibility is to convert heterogeneous evidence into deterministic, auditable contig-level decisions: order, split, cut, scaffold, review, plot, and optionally fill graph-supported gaps.

The implementation is intentionally organized around command boundaries. Each command writes explicit FASTA, TSV, HTML, or plot artifacts rather than sharing hidden state. This makes the pipeline inspectable by users and reviewers, and it makes sequence-changing steps depend on documented inputs and reports.

Design Principles

1. Evidence first, sequence edits second. Alignment metrics, graph context, gap decisions, and manual-review choices are written as reports. Sequence changes occur only in commands whose purpose is to write a new FASTA: sort, clean, fix, cut, manual apply, scaffold, and gapfill --apply.

2. Conservative defaults. The default behavior favors retaining ambiguous evidence for review over making irreversible edits. Examples include split-candidate protection in sort, breakpoint smoothing in fix, report-only graph evidence in sort, fix, and scaffold, and ambiguous-path refusal in gapfill.

3. Simple, explicit data models. FASTA records, alignment segments, merged intervals, assignment rows, graph nodes/edges, split pieces, scaffold gaps, and fill plans are represented with small dataclasses or TSV schemas. The project avoids an internal database and uses files as the public contract between stages.

4. Format normalization before decision logic. MUMmer show-coords and minimap2 PAF records are normalized into a shared Segment representation before coverage, identity, orientation, and ordering decisions are made. GFA graph support is similarly normalized into oriented node and link records.

5. Graph evidence is guarded. GFA, GAF, Hi-C-like contact tables, and reference-placement PAF can refine review and gap-filling decisions, but sequence insertion requires a validated graph path. Complex or unknown GFA overlaps are preserved as uncertainty rather than converted into trim lengths.

6. Reviewer-readable provenance. Important decisions are surfaced as status codes, reason strings, support counts, interval lengths, risk flags, and summary tables. The goal is to make each accepted or rejected decision reconstructable from input evidence.

System Map

The commands can be used as a pipeline, but the architecture does not require a single prescribed run order. Users can stop at any report, inspect it, edit manual inputs, re-align edited FASTA records, and resume with later commands.

Command/API Responsibilities

The public command surface is the chromo dispatcher in src/chromosort/cli.py. pyproject.toml also exposes compatibility scripts: chromosort, chromosort-clean, chromosort-fix-contigs, chromosort-scaffold, and chromosort-gapfill.

Algorithm And Data Model Activation Map

This section makes explicit where each documented algorithm or data model is used. “Activation” means the command path, mode, or parameter that calls the logic; it does not imply that the evidence source is allowed to change sequence by itself.

Evidence Activation And Authority

Formal Data Model

Coordinate Conventions

ChromoSort accepts two alignment coordinate systems:

• MUMmer show-coords rows, which are read as 1-based closed endpoints.
• PAF rows, whose 0-based half-open endpoints are converted into equivalent 1-based closed endpoints by adding one to the start coordinate and preserving the end coordinate.

Most interval arithmetic then converts endpoints to half-open intervals $I = [a,b)$ with length

For reverse-strand rows, interval endpoints are normalized with min and max for arithmetic, while the original strand is retained as an orientation field. This separation keeps overlap and coverage computations independent of strand direction.

Alignment Segment and Match Metrics

The shared Segment model records:

• reference name $r$ and query/contig name $q$,
• reference and query endpoints,
• aligned reference and query lengths,
• percent identity $p_s$,
• sequence lengths $L_r$ and $L_q$ when provided by the alignment format,
• orientation, MAPQ, and secondary-alignment status when available.

For a query $q$ and reference $r$, let $S_{q,r}$ be the set of passing alignment segments after identity, MAPQ, secondary-PAF, and length filters. For segment $s$, let $Q_s$ be its query interval, $R_s$ its reference interval, and $l_s = |Q_s|$ its query-aligned length. ChromoSort aggregates those segments as:

The resulting MatchMetrics object stores raw aligned bp, merged query bp, merged reference bp, query coverage, reference coverage, weighted mean identity, reference start/end, reference midpoint, dominant orientation, segment count, and merged reference intervals.

The best-reference share used by sort is:

where $r^*$ is the best-ranked reference for query $q$.

Graph Model

graph.py reads the subset of GFA needed by ChromoSort:

• S records become GraphNode objects with name, length, optional sequence, tags, and source line number.
• L records become GraphEdge objects with source node, target node, orientations, raw overlap CIGAR, parsed overlap bp when exact, tags, and line number.
• AssemblyGraph indexes outgoing and incoming edges by oriented key $(node, orientation)$ where $orientation \in {+,-}$.

The parser is strict for missing linked segments and inconsistent LN:i lengths when sequence is present. It only converts simple overlap CIGARs made of M, =, and X operations into exact trim lengths. * or complex CIGARs containing insertion, deletion, clipping, skipped, or padding operations return unknown overlap (None) so downstream sequence-changing code cannot treat them as validated trim lengths.

Command-Level Records

The command modules layer additional records over these primitives:

• Assignment records store the selected reference, status, best and second match, overlap-filter annotations, split-candidate annotations, output name, and orientation outcome for sort.
• QueryBlock, BlockGroup, SplitPiece, and ContigPlan represent alignment blocks, smoothed breakpoint groups, emitted pieces, and planner status for fix.
• ScaffoldMember, GapRecord, GraphGapRecord, and ScaffoldRecord represent ordered contigs, inferred gaps or overlaps, graph evidence at adjacent contig junctions, and emitted scaffold records for scaffold.
• FillPlan records graph path status, support counts, risk flags, fill sequence, right-flank trim length, reviewed acceptance, and applied status for gapfill.
• ReviewEvent records use schema chromosort-review-event-v1 and carry task type, candidate action, target, default acceptance, status, confidence, reason, notes, and task-specific evidence fields. They are written by chromo eval fix/scaffold/gapfill, embedded by chromo manual fix/scaffold/gapfill, and consumed by fix, scaffold, or gapfill through --reviewed-plan.
• Manual recipes use schema chromosort-manual-v1 and store 1-based inclusive source slices, output names, strand choices, scaffold labels, and removal flags.

Long-Read Evidence Records

longreads.py parses long-read-to-assembly PAF records into read-level alignments keyed by read name and target contig. These records are not a general-purpose aligner replacement. They are used only by chromo eval modes to summarize breakpoint and contig-end support, and by chromo manual to label the optional long-read PAF evidence panel when review-table fields are present.

gaf.py parses long-read-to-graph GAF rows into oriented node paths. The shared summary functions are used in four places:

• chromo eval fix --gaf summarizes reads touching a candidate graph node.
• chromo eval scaffold --gaf compares support for selected and alternate GFA paths between adjacent contigs.
• chromo eval gapfill --gaf and chromo gapfill --gaf summarize candidate fill-path support from the same traversal logic.
• chromo manual --gaf displays gaf_* review-table fields for the selected event; it records the evidence input but does not independently recompute GAF support in the browser generator.

Core Algorithms

Alignment Normalization and Metric Construction

The parser layer streams coords or PAF rows, skips malformed rows, applies filter thresholds, and yields normalized Segment objects. The aggregation layer groups segments by (query, reference), merges query and reference intervals independently, and computes the formal metrics above. This design is appropriate for fragmented whole-genome alignments because it avoids double counting overlapping alignment rows while preserving raw aligned bp and segment count for review.

Reference Assignment and Duplicate Filtering

chromo sort evaluates every assembly FASTA record against its reference matches:

1. For each query, matches are ranked by merged query bp, raw query bp, and mean identity.
2. The best match is rejected as below_min_aligned_bp, below_min_query_cov, or ambiguous_ref_match if it fails the configured support thresholds.
3. A strong multi-reference record can be protected as kept_split_candidate when at least two references meet the split-candidate bp and query-fraction thresholds and the best-reference share is not too high. This keeps likely chimeric contigs available for chromo fix.
4. A large alignment can be rescued as kept_large_alignment when it narrowly fails query coverage but satisfies the large-alignment thresholds.
5. Kept records are passed to the duplicate-overlap filter unless --no-overlap-filter is used.

For each reference, the duplicate-overlap pass ranks candidate assignments by merged reference bp, merged query bp, query coverage, mean identity, and query length. Higher-ranked contigs claim reference intervals first. Candidate intervals are either:

• a broad first-to-last reference span (--overlap-mode span), or
• exact merged alignment intervals (--overlap-mode alignment).

Let $I$ be the candidate interval set and $C$ be the already covered interval set on the same reference. Novel reference support is:

A lower-ranked contig is kept when the configured absolute and fractional novel reference thresholds pass. If not, split-candidate protection and terminal overlap rescue are considered before the contig is labeled terminal_overlap or duplicate_overlap.

Terminal-overlap rescue is intentionally narrower than general overlap acceptance. It applies only when the novel interval is a one-sided extension of a single candidate interval:

This distinguishes plausible dovetail extensions from contained duplicate fragments.

Breakpoint Planning for chromo fix

chromo fix starts from filtered alignment segments and converts them into query-ordered QueryBlock records. Adjacent blocks with the same target reference and orientation can be merged when they are separated by at most the configured query gap. Candidate split signals are then planned differently by mode:

The important implementation distinction is that comprehensive is not implemented as “conservative plus extra calls.” It runs the smoothed planner with orientation included in every target signature. That can split the evidence into different candidate pieces or reject a plan for weak-piece support even when conservative would split the same contig using reference-only signatures.

The smoothed planner uses identity-weighted bp:

for block $i$ with aligned length $l_i$ and identity $p_i$. For an interval of blocks $G$, support is accumulated by signature $\sigma$ (reference alone or reference plus orientation, depending on mode). The discordance cost of keeping that interval as one piece is:

Dynamic programming chooses a partition of the block list that minimizes:

where $\lambda$ is --breakpoint-penalty-bp. The planner reports a score equivalent to the reduction in discordance after paying breakpoint penalties:

After partitioning, terminal groups that fail support thresholds can be merged into neighbors, all remaining groups must pass minimum aligned-bp and query-fraction thresholds, and a per-contig breakpoint guard can reject overly fragmented plans. Accepted breakpoints are placed midway between adjacent alignment blocks, then clamped to the contig sequence length. This produces piece coordinates that are deterministic and easy to audit, while smoothing small discordant blocks and local structural-variant-like noise by default.

Explicit Cutting

chromo cut is deliberately not evidence-driven. It accepts repeatable CONTIG:POS[,POS...] cut specifications, a cut file, or a single-contig shorthand. Cut positions are 1-based and mean “cut after this base.” The planner rejects missing contigs, duplicate cut positions, terminal positions, and any plan that would create pieces shorter than --min-piece-bp. It also checks that emitted FASTA IDs will remain unique.

Manual Review and Application

chromo manual computes the same best-match metrics used by sort, embeds filtered alignment segments, optionally embeds graph node context, and writes a browser-based dashboard. Initial pieces cover each source contig completely, with the scaffold label initialized to the best reference or unplaced. Task-specific manual fix, manual scaffold, and manual gapfill modes load the shared review-event table with --review-table, open a focused event queue, and render selected-event evidence panels for alignment, GFA, long-read PAF, and GAF inputs when those data are available.

chromo manual apply is schema-driven. It validates the recipe schema, verifies that requested source contigs and slices exist, reverse-complements pieces whose strand is -, omits removed pieces, and optionally groups active pieces into scaffolds joined by fixed N gaps. The recipe is therefore the sole authority for manual edits.

Eval Tables and Reviewed Execution

chromo eval is the table-first counterpart to task-specific manual review. It does not introduce a separate decision algorithm; it reuses the algorithm that the matching executor already exposes, then serializes candidate decisions as ReviewEvent rows with enough evidence fields for spreadsheet or dashboard review.

• eval fix activates the fix breakpoint planner and adds optional graph_*, gaf_*, and longread_* fields. Accepted rows are executable only through fix --reviewed-plan, which bypasses alignment re-planning and applies the reviewed source slices.
• eval scaffold activates scaffold gap/overlap calculation and optional graph junction evidence. Accepted rows are executable only through scaffold --reviewed-plan, where they override matching gap lengths but do not reorder, orient, or trim contigs by themselves.
• eval gapfill activates the gapfill planner and writes the same path, support, validation, and fillability concepts as a review-event table. Accepted rows are executable only through gapfill --reviewed-plan --apply, where the current plan must still match the reviewed path.

Scaffold Construction

chromo scaffold groups kept assignment rows by assigned reference and uses the ordered FASTA from the same sort run. For adjacent contigs $a$ and $b$ on the same scaffold, the reference-inferred gap is:

If $g \ge 0$, the scaffold inserts either $g$ Ns or a fixed configured gap. If $g < 0$, the adjacent reference spans overlap by:

Overlap handling is policy-controlled:

• zero-gap writes no Ns and keeps both sequences unchanged.
• warn does the same but emits warnings.
• trim-reference trims the right-contig prefix by the reference-inferred terminal overlap.
• trim-sequence trims only when the right-prefix/left-suffix sequence identity over the overlap meets the configured threshold.
• graph-overlap-policy confirm can trim a terminal overlap when an oriented direct GFA edge confirms the junction.

Nonterminal overlaps are reported but not trimmed by the terminal-only trimming policies.

Graph-Supported Gap Filling

chromo gapfill plans fills between adjacent sorted contigs using oriented GFA paths. Path enumeration is breadth-first, orientation-aware, bounded by --max-path-edges, bounded by --max-candidate-paths, and cycle-aware. Missing graph nodes produce missing_node, absent paths produce no_graph_path, and multiple candidate paths remain ambiguous_paths unless exactly one path is selected by support evidence.

Support evidence is computed as follows:

• GAF support counts read paths that contain the candidate oriented path or its reverse-oriented counterpart as a contiguous subpath.
• Hi-C support sums contact counts across adjacent node pairs on the candidate path.
• Reference-placement PAF support sums the best identity-weighted query bp for intermediate graph nodes whose placements match the scaffold reference, orientation, and expected reference gap window.

If different evidence sources uniquely select different paths, ChromoSort reports the conflict and refuses to fill automatically. When a path is selected, sequence reconstruction still must pass validation:

• the GFA path must include sequence for both flank nodes and all intermediate nodes,
• the oriented flank node sequences must exactly match the adjacent FASTA records,
• every applied GFA link overlap must have a known simple bp length,
• overlap lengths must be valid relative to sequence lengths,
• the resulting inserted sequence must not exceed --max-fill-bp.

Only fillable plans can be applied. With a reviewed plan, the accepted row must still exist in the current plan, the current path must match the reviewed path exactly, and the current status must still be fillable.

Plotting

chromo plot draws direct visualizations from existing alignment rows. It constructs cumulative offsets for reference and query FASTA records:

and maps normalized alignment endpoints into a whole-genome or per-reference canvas. Forward and reverse alignments are rendered as line segments with different colors. Optional assignment input reorders the query axis by kept ChromoSort assignments without re-running placement. Optional selected-reference input filters the reference axis, alignment rows, and --per-ref panel set without changing the underlying FASTA or alignment file. The dot-plot guide describes how these segments map to common review patterns such as clean collinearity, reversed contigs, internal inversions, chimeric candidates, duplicate signal, and alignment gaps.

Evidence, Scoring, Ranking, or Decision Logic

ChromoSort does not currently implement a probabilistic confidence model. Instead, it uses deterministic scores and thresholded evidence quantities that are visible in reports.

The gapfill branch-complexity score is additive and intentionally simple:

where $P$ is the number of candidate paths, $H$ is the set of high-degree nodes, $S$ is the set of self-loop nodes, $U$ is the set of unsequenced nodes, $C$ is the number of avoided cycles during path search, and $X$ is the set of extra warning flags such as weak, tied, conflicting, or sequence-validation failures. The score is a review aid, not a calibrated probability.

Region, Window, Graph, or Table Construction Logic

Interval Regions

All alignment-derived region logic is built from merged half-open intervals. This applies to coverage, duplicate-overlap filtering, terminal-overlap classification, and reference span reporting. A candidate’s overlap with prior claims is computed by two-pointer interval intersection after sorting; novel intervals are computed by subtracting the covered set from the candidate set.

Reference Gap Windows

Scaffold gaps and reference-placement gapfill support use adjacent assignment coordinates. For same-reference gapfill candidates, the expected placement window is bounded by the adjacent assignment edges:

Intermediate graph-node placements from --ref-paf support a candidate path only when their reference midpoint falls inside this window and the orientation matches the candidate path.

Graph Paths

Graph path construction uses oriented keys. Starting keys are determined from the left assignment orientation when known, otherwise both orientations are allowed. Target keys are built the same way from the right assignment. The BFS state contains the current oriented node, the path of edges taken so far, and the seen oriented nodes. Revisiting a seen oriented node is counted as an avoided cycle and is not expanded.

Output Tables

TSV reports are not secondary decorations; they are command contracts. For example:

• sort assignment tables feed scaffold, gapfill, and ordered plots.
• eval review-event tables feed task-specific manual dashboards and the matching fix, scaffold, or gapfill reviewed execution path.
• fix, cut, and manual apply reports describe every emitted piece.
• scaffold_gaps.tsv records raw inferred gaps, final gap sizes, overlap class, policy, action, trim length, and optional sequence identity.
• graph_gaps.tsv records direct GFA adjacency, shortest reported path, and missing-node status for scaffold junctions.
• gapfill_plan.tsv records selected paths, alternative support, risk flags, fill status, fill length, right trim, acceptance, and applied status.

Cross-source or Multi-stage Reconciliation

ChromoSort reconciles evidence sources through explicit names and stage-local tables:

• Alignment evidence is keyed by reference and query FASTA names. PAF and coords inputs are interchangeable only after normalization into Segment records.
• Graph evidence resolves contig names conservatively. Scaffold and gapfill junctions try original contig names, ChromoSort new_name values, and FASTA record names when mapping records back to GFA segments.
• Review-event rows are task-scoped. manual fix/scaffold/gapfill rejects rows whose task does not match the selected mode, and executors consume only accepted rows for their own task.
• Manual-review recipes are validated against the original assembly FASTA; they do not inherit implicit state from the dashboard.
• Reviewed scaffold gaps are checked against the current scaffold, left contig, and right contig names before application.
• Reviewed gapfill plans are checked against the current graph path before application, preventing stale accepted paths from being silently applied.
• Sequence-editing stages can change contig IDs and coordinates. The intended workflow is to re-align edited FASTA records before using downstream alignment-dependent commands.

This reconciliation strategy favors reproducibility over convenience. It avoids implicit cross-stage mutation and forces each downstream stage to consume the current FASTA and the corresponding current reports.

Implementation Architecture

The source tree is a small Python package under src/chromosort/.

Most dependencies are from the Python standard library. Pillow is used for PNG plot output. Jekyll-related files support the documentation site and are separate from runtime package behavior.

Complexity and Scaling

The following bounds describe the implemented algorithms, not biological limits. Let $S$ be the number of passing alignment segments, $V$ graph nodes, $E$ graph edges, $Q$ query contigs, $B_q$ fix blocks for contig $q$, and $G$ adjacent scaffold gaps.

The most performance-sensitive code paths are interval merging, the fix planner’s quadratic DP over alignment blocks, graph BFS at scaffold/gapfill junctions, and large plot rendering. The exposed depth, candidate, segment, and format filters exist to bound those costs for larger assemblies.

Validation Strategy

The repository contains a pytest/unittest suite configured in pyproject.toml. The tests are synthetic and deterministic; they are best interpreted as unit and integration evidence for the implemented decision rules, not as biological benchmarking.

Validation coverage currently includes:

• CLI dispatch and mutually exclusive argument checks.
• coords and PAF parity for sorting and fixing.
• best-reference assignment, duplicate-overlap rejection, split-candidate protection, terminal-overlap classification, terminal-extension rescue, and large-alignment rescue.
• graph assignment reports and graph guard warnings for sort/fix cases.
• breakpoint smoothing, sensitive/comprehensive/conservative fix modes, inversion handling, noisy small-event rejection, contig-subset parity, and breakpoint guardrails.
• review-event schema reading/writing, accepted/rejected value parsing, duplicate-id rejection, and task mismatch rejection.
• eval fix, eval scaffold, and eval gapfill table generation, including reviewed executor handoff cases.
• long-read PAF breakpoint and contig-bridge summaries.
• GAF parsing, MAPQ filtering, node summaries, selected-path support, alternate support, and mixed GFA/PAF/GAF scaffold evidence.
• explicit cut validation for duplicate, terminal, and file-based cut inputs.
• manual dashboard ordering, optional sequence embedding, graph context, task-event queues, modular evidence panels, recipe slicing, reverse-complementing, removal, and scaffold grouping.
• plot generation for PDF, SVG, PNG, per-reference plots, and assignment-based query ordering.
• scaffold inferred/fixed gaps, negative reference-gap overlap reporting, reference trimming, sequence-confirmed trimming, and graph-confirmed trimming.
• GFA parsing of orientations, simple versus complex overlap CIGARs, missing links, and length mismatches.
• gapfill unique path application, reviewed-plan acceptance/rejection, stale path rejection, ambiguous branch refusal, GAF/Hi-C/reference-placement branch resolution, conflict refusal, risk flags, and review HTML embedding.

The tests/data/graph_gotchas/ fixture documents intentionally difficult graph cases: an unambiguous path, an ambiguous branch, a cycle, orientation-specific links, disconnected mapped nodes, unsequenced segments, and repeat/duplicate warning contexts.

Runtime validation is also part of the architecture. Many user-facing errors are raised as ValueError and converted to command exit status 2 with a clear message. Examples include mutually exclusive alignment inputs, malformed GFA records, missing assignment-table columns, stale reviewed gapfill paths, invalid cut positions, missing source FASTA records, and unsupported graph safety flags without --gfa.

The current GitHub Actions workflow deploys the documentation site from docs/. At the time this document was written, test execution is configured locally but is not represented as a repository CI workflow.

Non-goals and Limitations

• ChromoSort does not run MUMmer, minimap2, GraphAligner, Hi-C processing, or assembly tools internally. It consumes their outputs.
• ChromoSort does not polish base calls, call variants, estimate haplotypes, or validate biological correctness beyond the supplied evidence.
• The ranking and risk scores are deterministic heuristics, not calibrated probabilities or statistical confidence intervals.
• Reference-guided decisions inherit reference bias. Contigs absent from, highly diverged from, or structurally incompatible with the reference can be under-supported or classified for review.
• Duplicate-overlap filtering is per reference and based on reference interval novelty. It is not a global assembly graph optimization.
• GFA support is limited to segment (S) and link (L) records. GFA path/walk records are not used by the core graph parser.
• Sequence-changing graph fills require segment sequences and simple exact link overlap lengths. Missing sequences, complex overlap CIGARs, flank mismatches, invalid overlaps, overly long fills, or ambiguous/conflicting branches prevent automatic filling.
• Scaffold overlap trimming is policy-limited and terminal-focused. Nonterminal overlaps are reported for review rather than broadly edited.
• After FASTA-changing steps, downstream alignment-dependent commands should be run with fresh alignments against the edited FASTA.
• The package metadata currently classifies the project as alpha, so reviewers should treat command behavior as implemented and tested, but still evolving.

Reviewer’s Reading Guide

For code review, start with the command boundary and then follow the evidence models into the algorithms:

1. src/chromosort/cli.py for public command routing.
2. src/chromosort/reference_order.py for alignment normalization, interval math, match aggregation, sort assignment, and duplicate-overlap logic.
3. src/chromosort/fix_contigs.py for breakpoint planning, smoothing, mode behavior, and emitted split pieces.
4. src/chromosort/evaluate.py and src/chromosort/review.py for review-event table generation, schema constraints, and reviewed handoff fields.
5. src/chromosort/graph.py, src/chromosort/gaf.py, and src/chromosort/longreads.py for graph and read evidence constraints.
6. src/chromosort/scaffold.py and src/chromosort/gapfill.py for junction-level gap, overlap, graph path, and fill behavior.
7. src/chromosort/manual.py, src/chromosort/cut.py, and src/chromosort/plot.py for review, explicit edit, and visualization surfaces.
8. tests/test_reference_order.py, tests/test_fix_contigs.py, tests/test_eval.py, tests/test_review.py, tests/test_longreads.py, tests/test_gaf.py, tests/test_scaffold.py, tests/test_gapfill.py, tests/test_graph.py, tests/test_manual.py, tests/test_cut.py, tests/test_plot.py, and tests/test_cli.py for executable examples of the expected behavior.

For a methods review, the most important invariants to check are:

• interval normalization and merged-bp accounting,
• conservative handling of ambiguous or duplicate alignments,
• breakpoint smoothing and support thresholds,
• explicitness of every sequence-changing operation,
• review-event task scoping and accepted-row validation,
• graph path validation before sequence insertion,
• distinction between report-only evidence and evidence allowed to select or apply a sequence-changing path,
• stale-plan prevention for reviewed gap fills,
• report schemas used as contracts between stages.