# Seet multicpu in Assembly Explorer

Select **Seet multicpu** on the
Molecule tab. It can run alone or alongside the serial Seet baseline and other
methods, with separate pathways, timing, and upper-bound curves.

The underlying method is independently implemented from
[Seet et al. (2025), Rapid Exploration of the Assembly Chemical Space of Molecular Graphs](https://doi.org/10.1021/acs.jcim.5c01964).
No authors' implementation code is incorporated. Parallel fragment preparation,
exact matching, shared-frontier scheduling, and bounded shared state caching are
Assembly Explorer execution extensions. They preserve the duplicate-removal
branches and the paper's conditional bound (Eqs. 3–4). The serial baseline remains
separately selectable. Both require the same element/bond-labelled input graph
for meaningful comparison.

## Resources and guarantees

The web worker launches the fixed `native/bin/assembly-seet-parallel` executable
as one process with OpenMP threads. A single-method job receives the configured
allocation (currently up to 32 threads and 64 GiB); simultaneous comparison
methods divide those limits. Sequential comparison gives each method the full
per-job allocation. The native cache budget follows the assigned memory share.
Network isolation, process-group cancellation, CPU/wall-clock deadlines, output
caps, and aggregate cgroup limits remain in force.

All transmitted pathways are independently checked. An optimality result also
requires complete fragment enumeration, no unfinished search branches, a clean
worker exit, and completion within the deadline. Interrupted runs retain only
checked upper bounds. Native search telemetry is advisory and cannot establish
optimality.

## Progress

Generated fragments and duplicatable occurrences describe fragment enumeration.
Assembly states count duplicate-removal search work across workers. Queued
continuations describe the current unfinished frontier, not the total number of
paths remaining. The progress bar continues to show time budget used.

The optional CLI estimator is enabled for this web backend. It independently
adapts the weighted-backtrack estimator from
[Kilby et al. (2006), Estimating Search Tree Size](https://users.cecs.anu.edu.au/~thiebaux/papers/aaai06.pdf)
to multiway search and parallel leaf accounting. Its search-tree node count
includes immediately pruned children and differs from assembly-state counts.
Estimated remaining time is heuristic, uncalibrated, and not guaranteed; it may
change substantially. Warm-up, underestimated, numerically unusable, and
interrupted estimates are not shown as completion forecasts. Estimates never
affect pruning or the exact-result flag.

## Deployment and HPC

Build the OpenMP target with the project's CMake configuration, test it, and
install `assembly-seet-parallel` into `native/bin/` before restarting the API.
Startup validates its algorithm identity, OpenMP backend, and molecular input
contract. The HTTP request accepts an algorithm selection and time limit, not
executable paths, checkpoint files, MPI ranks, or arbitrary native flags.

The same search has an MPI build with distributed continuations and checked
catalogue distribution. MPI launch and checkpoint/restart are CLI facilities;
see `references/SEET_HPC_RELEASE.md` and `hpc/README.md` in the source repository.

## Deployment verification — 2026-09-22

The installed OpenMP executable matches the tested HPC release:
SHA256 `490e8b530e58d161c0cc00b496c3c2c703430841a8900112abfec80a90de6a93`.
The API exposes thirteen molecular methods, including serial and parallel Seet.
Its confinement check includes the new executable's read-only permissions.

- 197 targeted Python regression tests passed; 18 opt-in experimental tests were
  skipped. Coverage includes real 1/4/32-thread runs, independent pathway checks,
  comparison resource shares, invalid proof evidence, timeout/cancellation,
  strict advisory telemetry, history, and confinement.
- Mock browser checks passed for method eligibility, descriptions/citations,
  separate colours/bars, advisory estimates, final counts, and catalogue state.
- Public-browser checks passed for eleven string, thirteen molecule, and nine
  custom-space methods. Downloaded pathways were independently verified; the
  exact methods agreed on indices 4, 5, and 3 for their respective small examples.
  Six-hour admission with immediate owned-job cancellation, proxy-error recovery,
  molecular pathway display, and mobile layout also passed.
- A standalone production parallel-Seet cysteine job proved index 5, reported
  complete enumeration and an empty frontier, and received 32 threads with a
  64-GiB address-space ceiling. This is a functional check, not a speedup study.

The comparison renderer now releases molecule controls after its viewing flag
clears, so opening a completed 3D pathway leaves Build Pathway usable.