# Tree-decomposition comparison This directory contains a static interface for comparing tree-decomposition solvers. Each instance is one graph. Optional source, kind and collection fields record where the graph came from without changing how it is compared. The interface has no build step or external web dependency. Its stylesheet takes the solver page's palette, in both colour schemes, and its type (`bindings/wasm/styles.css` in goatd), so the two pages read as one site; `logo.png` is the same logo. Published, the page sits under `comparison/` on goatd's `gh-pages` branch, beside the solver page at the root, and the two link to each other. `results.fixture.json` contains invented values for interface development. It must not be presented as a benchmark result. ## Local preview Serve this directory with any static file server. For example: ```sh python3 -m http.server 8000 ``` Without a browser on the host, `scripts/render-comparison.sh [light|dark]` (from the repository root) writes `tmp/render/.png` with headless Chrome, in either colour scheme. The page loads `results.json` by default. Load the synthetic interface fixture explicitly when developing the view: ```text ?data=results.fixture.json ``` Filters update both the aggregate comparison and the detailed results. Tree components are omitted. By default, a graph is excluded only when the pinned NetworkX min-degree run returned a counted width below 30. A missing, invalid, timed-out or no-reduction min-degree result is retained. The control can remove or change that floor. Adding a solver to the comparison therefore does not change the selected graph population. Sorting applies only to the detailed records, which are paged so a large corpus does not create every row at once. For benchmark accounting, a result must pass validation, contain more than one bag, and have width below the graph's one-bag width `|V| - 1`. A result that does not improve that bound is counted like a timeout. The raw result remains available in the detailed view. ## Aggregate statistics The summary reports, for every displayed solver: - counted decompositions; - exact ties with the best width observed for each graph; and - counted widths within one and four of the best observed width. The canonical result set contains one goatd row: the portfolio from current upstream main. Historical portfolio checkpoints stay in experiment records and git history; they are not part of the displayed field or its observed minima. Tied minima count as best for every tied solver. The quality--coverage curve reports the count with a counted width at most 0, 1, 2, 4 or 8 above the best observed width. Missing, invalid, timed-out and no-reduction results do not satisfy a profile threshold. Hovering or focusing a metric label explains its denominator and interpretation. Aggregate metric labels appear once in the header. Each curve has a wider invisible pointer target; hovering or focusing it identifies the solver, while its marked thresholds report exact values. The detailed view lays out every solver in a responsive grid under its graph; it does not require horizontal scrolling. Connected graphs with `|E| = |V| - 1`, whose exact width is one, are omitted. Graph views with no nontrivial component are absent from the export. Generate the public-README preview from the same solver field and default filters: ```sh node export_markdown.js results.json README-table.md ``` `presentation.js` is the shared source for the displayed field, the pinned min-degree filter and metric definitions. `README-table.md` is generated; do not edit its numbers by hand. ## Result format Schema version 2 is independent of a particular graph corpus, solver or runner. Files use this top-level shape: ```json { "schema_version": 2, "dataset": { "title": "Display name", "description": "Optional description", "synthetic": false }, "run": { "generated_at": "2026-08-26T18:30:00Z", "timeout_seconds": 10 }, "solvers": [ { "id": "stable-id", "name": "Display name", "version": "Optional version" } ], "source_instances": [], "instances": [] } ``` `source_instances` is optional. It has the same result-map shape as `instances`, plus component counts, and supplies the source-level aggregate rows described above. Each instance supplies one graph and one result map: ```json { "id": "graph-pair-sha256:digest/primal", "label": "example / primal", "source": "collection/example.cnf", "kind": "primal", "group": "collection", "alias_count": 3, "aliases": ["optional/example-copy.cnf"], "vertices": 100, "edges": 450, "sha256": "graph-file-digest", "results": { "stable-id": { "status": "ok", "budget_reached": false, "width": 12, "total_bag_size": 700, "bag_count": 90, "elapsed_ms": 250, "wall_elapsed_ms": 250 } } } ``` Only `id`, `vertices`, `edges` and `results` are required. `label` is the human-readable graph name. `source`, `kind` and `group` are optional provenance and filter fields. For the CNF-derived evaluation, a label such as `example / primal` names the graph while `source` retains the canonical CNF path. The interface does not pair that graph with an incidence graph or give either kind special treatment. Solver IDs in `results` refer to entries in the top-level `solvers` array. Solver columns are not fixed; the page renders every displayed entry in a responsive grid. If a solver has no entry for a graph, the page shows that cell as unavailable. `aliases` lists other source names represented by the graph. `alias_count` is the number of additional names and takes precedence in the display when both fields are present. Listed aliases are included in text search. A completed result uses status `ok` and supplies all four metrics. Other statuses may be `timeout`, `error`, `invalid`, `skipped`, `unavailable` or `incomplete`; their metrics are `null`. `elapsed_ms` may be present for a non-completed run, and `message` may give a short reason. `budget_reached` distinguishes a result emitted in response to the fixed deadline from a solver that exited early. Deadline results remain eligible for decomposition-metric comparisons but not elapsed-time rankings. Width, total bag size, bag count and elapsed time are minimized. Every rank is computed among counted decompositions for one graph. Large exports use schema version 3. The named JSON file contains the same dataset, run, and solver metadata plus `instance_count` and `instance_files`; each listed file contains an `instances` array. The view loads both forms.