Status
35 open roadmap items, 2 open pull requests and 1 open issue, read at Oct 30, 2026, 05:54 UTC from dc61877.
The build reads the roadmap file and GitHub again after each commit and on a schedule.
Open work
Each row of a table of ROADMAP.md with an ID and a status that is not done. The build reads the file, so this list follows each commit.
35 of 35 open items
Later 35
| ID | Item | Status |
|---|---|---|
| BUILD-QUEUE | Up to 3 builds at a time take the ideas from the top of the build queue. Each build takes an idea with pnpm queue --take, skips a blocked idea and commits the new queue. | later |
| INTERIOR-ROUTE | Slice 2: inner box to inner box. Each edge between two nodes of one frame routes inside their nearest common frame, around the other nodes. Slice 1 (the plugs of the walls) shipped. | later |
| FIX-PORTS-RESULT | fixPorts also gives the chosen plugs (port, point, side) as data, so a consumer does not read them back from the ports of the graph. The return type of today stays as it is. | later |
| INNER-ONE-CALL | One call routes the outer edges and the inner edges of the frames. Today the inner routes need a second call (splitFrames, then frameRoutes) after the outer route. | later |
| SEARCH-CROSS | The default score of search follows ink, R1 and R4 together. SEARCH-OBJECTIVE found that crossingWeight 0.3 with half penalties follows ink and R1, but its fit to R4 falls from 0.53 to 0.11. Step 1: find the term that makes B1 go against R4 (ρ -0.71). Add an R4 term or correct that penalty, and screen until the median ρ is 0.5 or more for the three measures. Step 2: a work counter for each new term. Step 3, the gate of the default: the seeds 6 to 10 on screen-v3 and the dense window of 500 boxes. Gains on ink and R1 with Holm, and R4 drops beyond the floor on 2% of the inputs or fewer. The claim gives the median and p10 of each measure, the share of the inputs that move and the time ratio. A pass is a MINOR. | later |
| PLUG-ROW-GAP | Two plugs of one wall land a few pixels apart, so the second row of a card has no room. A consumer anchors 0 of 48 such row ends. Step 1: a bench measure, the fraction of the plugs whose row fits, and the minimum gap of the plugs of one wall in row heights. Step 2: the plug gap of spread knows the row height of each port. The gate: the measure goes up on a made-up board of tall rows, with no drop of the other measures. | later |
| INCREMENTAL-SEARCH | Each plan that search measures routes and scores all the edges, but a move changes a few highways. This item routes and scores again only the edges of the changed highways. The gate: the scores and the route hashes are equal to the full evaluation on each screen input. Then the speedup against the work. The baseline is the lane count of one evaluation: 3.1 billion on 1000 boxes (SEARCH-BUDGET). The crossing term of SEARCH-OBJECTIVE counts pairs too, so it needs an incremental form. | later |
| MOVE-OTHERS | A measure next to Z11: when one box moves, do the routes that do not touch it stay where they are? A second number judges the routes with an end on the moved box, against the move of the box. Each router gets the measure on the same 5 member sets, with predictions before the run. | later |
| GROUP-BUNDLING | Bundle purity is a measure of the bench: the share of the edges of a bundle that are in one group. Each bundling strategy gets a group term in its compatibility, from 0 to 1. A bundle splits when the group rule weighs more than the ink that it saves. The atlas shows purity against ink as a front. | later |
| PASS-GRID | On screen-v2, each highway strategy has too many pass-throughs on some packed-grid seed, so the front of the highway screen is empty. Find the cause on the worst seeds: the plan, follow, or the floor. | later |
| FAMILIES-WEEKLY | New input families each week, with no ceiling: dumps of made-up data, adversarial and worst cases, scaling ladders and benchmarks from the literature. The sliders of the site get each one at once. | later |
| SPACE-COVERAGE | One number on the site, with a weekly trend: the share of strategies, options and input dimensions where each measure has a fitted model, a held-out error and Sobol indices. | later |
| SOLVED-ATLAS | The atlas shows each solved region. A region is solved when its leader is within a measured bound of a proven lower bound, or when one leader holds for 4 weeks. | later |
| QUEUE-REFILL | pnpm queue tells when the open ideas are fewer than one week of build slots. Then a literature survey and a round of hybrids add ideas to the queue. | later |
| POINT-AT | The export pointAtFraction(route, t) gives the point and the tangent at a fraction of the length of a route. Then a renderer that puts a badge at half the length has no copy of that math. | later |
| OPT-IN-LIST | The core exports OPT_IN_STRATEGIES and STEP_DEFAULTS. The strategy cards, compare and the bench read them, and no page names the opt-in routers by hand. | later |
| IDEA-GUARANTEES | An optional field guarantee on each idea record. Today 72 of 79 strategy cards say "none stated", because the cards find a guarantee by a keyword match. | later |
| SHAPE-SUPPORTS | The options of routeSpace declare supports: { shape: ["rect"] }, because each router takes box obstacles only. | later |
| LANE-WALK | The lane order of a dense run without a walk for each compare. After ROUTE-SCALE, partBefore and partAfter take approximately 55% of a shared route on the dense board. The lane order stays the same. | later |
| BENCH-STEPS | Cheaper measures on a dense board: exactStretches (a landmark guide for distancesTo), measureQuality and checkProperties. The values stay the same. | later |
| STAGE-DATA | The stages of the routers and their options as data (strategy-space/1 of layout-contract 0.2.10): routeSpace, routeConfig and each router of today as a named point with an equal route hash. | later |
| STRATEGY-CARDS | Part 2 of the strategy cards: the curves of one option at a time and the Sobol indices of each strategy, from the nightly sweep. | later |
| TIME-PAIR | The time test runs each candidate and its default in one job, in alternate runs, so both use the same CPU. | later |
| SELECT | A selector for each input from its features, or from a cheap trial run of a few strategies. The oracle bound is large: 0.352 of gap closed on exact-v1, and 39% or more of B1 on 4 of 4 boards. A choice by class keeps none of it, so the selector needs finer features and more seeds of each board. A portfolio with a score of the output alone keeps 6% on exact-v1 (PORTFOLIO), so the selector needs a score that knows the rule of the run. PORTFOLIO-RULE gives that score: B1 among the runs that obey the rule of the caller, at one cap. | later |
| HWC-E | Room for the plug stubs, and the spacing rise of the bends: most arcs under the smallest radius sit on short plug stubs or on legs along one box. | later |
| MEAS-FLAT | Z2 is 0 on each record, and B8 runs is degenerate where shared has 0. Correct each definition, or mark it inactive with its reason. | later |
| REBUILD-SILENT | ACTIVATION found 15 strategies whose own move does not fire on screen-v1. For each, a small board on which the move must fire gives the cause. When the move fires there, the corpus has a gap, and such boards go to CORPUS-SEEDS. When it does not fire, the trigger is wrong, and the strategy is built again. | later |
| CURV3 | More curve types from CURV-R, each with its options: Hobby curves, B-splines, biarc corners and superellipse corners. | later |
| GRP3 | More grouping rules from GRP-R: agglomerative, leiden and pack, and the options refine and band of the rule cut. | later |
| TUNE | The tuning harness: it sweeps the options of each strategy from its declaration. It gives the Pareto front and the best setting for given weights. | later |
| PLAY | The playground page: choose a strategy for each step and move the options. The measures change live, and a link keeps the setting. | later |
| EX3 | More example boards of made-up data: charts of teams, pipelines, network maps, timelines and dense grids, each with its page. | later |
| SHARED-SMOOTH | The routers shared and shared-all read their option smooth, or the schema loses it. Today the control does nothing. | later |
| COMPARE-PIPELINE | Compare picks a whole pipeline from routeSpace. Waits for STAGE-DATA. | later |
| BUILD-W | Build and screen each idea of the catalog, in waves by family and by expected value. Each kept idea gets its folder, its tests and a row in the benchmark. The next waves aim at two holes of the portfolio first. The first hole is the rule collection at the cap 1.1 on large boards: there the base shared wins on 18 of 20 test inputs. The second hole is the rule all at a tight cap (1.1, 1.25) on large boards, where the medians of gap closed are negative. After them, the waves aim at the rule collection, because there the best single strategy reaches 0.021 of gap closed, against an oracle of 0.378. | later |
No open item matches the filters.
No item is done.
Pull requests
2 open pull requests, the most recently updated first.
Issues
1 open issue, the most recently updated first.