outputs.md

Output files

ensemble-mtd-aggregate --outdir rcng writes the following into rcng/. The names are unchanged from earlier versions of this pipeline so existing analysis scripts keep working.

The ones you will actually read

aggregate_reacnet_species.tsv

One row per collapsed species, reactants first (in the order they were first seen), then everything else by prevalence.

Column	Meaning
species_id	S1, S2, … Stable within one aggregation, not across aggregations
species	The representative SMILES for the group
collapsed_inchi	The fixed-H InChI the group is keyed on. This, not the SMILES, is the stable identity
representative_smiles	Same as species; kept for older readers
member_raw_species_count	How many distinct raw ReacNet SMILES collapsed into this group
display_run_support	max(observed_run_support, first_frame_run_support) — for display only
observed_run_support	$N_i^\mathrm{runs}$. The number to quote
observed_run_support_json	The same count from ReacNet's JSON species list, which --maxspecies truncates
support_delta_json	The gap between the two. Non-zero means the JSON list undercounted
support_source	Which source observed_run_support came from (timeline by default)
pct_runs_observed	$\Pi_i$ in percent. The number to quote
is_initial	1 if present in the first analysed frame
first_frame_run_support	Runs in which it was in the first frame
initial_frame_count	Summed first-frame multiplicity
involvement_events	Total reaction events this species took part in, either side

aggregate_reacnet_species_run_membership.tsv

The presence matrix: for each species, the names of the runs it appeared in, semicolon-separated. Everything else stores only counts, so this is the file needed for any per-run or resampling analysis — including ensemble-mtd-convergence, which reads it. The run sets here are the same ones pct_runs_observed is computed from, so the counts reproduce exactly.

aggregate_reacnet_reactions.tsv

Collapsed reaction records, ordered by run support then total events.

Column	Meaning
reaction	lhs => rhs, with stoichiometry written as 2*X
run_support	Distinct runs containing this record
pct_runs	The same as a percentage
total_events	Summed event count across the ensemble

These are the unnetted, directional weights. --clean affects only the drawn network, never this table.

aggregate_reacnet_network.dot

The consensus network as Graphviz source. Nodes are S# (map them through aggregate_reacnet_network_nodes.tsv or the species table); reactants are outlined in indigo, a highlighted route in red. Every edge is labelled e=<event weight>, r=<run support> and its tooltip carries the full SMILES. Render with:

dot -Tpdf rcng/aggregate_reacnet_network.dot -o network.pdf

aggregate_reacnet_network.svg is written too if dot is installed. It is not required: the .dot file is the artefact, and it can be rendered anywhere later — which is also what --rebuild-from is for.

aggregate_reacnet_summary.json

Provenance and run accounting. The fields worth checking every time:

• n_input_runs, n_success_runs, n_excluded_runs, n_failed_runs — the denominator of every prevalence, and whether anything went missing.
• excluded_runs — runs dropped because a frame reported zero energy, i.e. a collapsed SCF.
• run_stats — per run: frames, atoms before and after Li removal, species and reaction counts, and the exclusion reason if any.
• n_raw_species, n_collapsed_species, collapse_ratio — how much the InChI grouping merged.
• obabel_version — the InChI grouping depends on the Open Babel build, so this is part of the provenance of every species label.
• reacnet_settings — what ReacNetGenerator was actually told, nohmm included.
• graph — the thresholds applied, the edge and node counts, and the highlighted route if one was drawn.
• warnings, failures — same content as the two TSVs below.

Supporting files

File	Contents
aggregate_reacnet_inchi_alias_map.tsv	Every raw ReacNet SMILES and the collapsed group it was folded into. Where to look when a species label surprises you
aggregate_reacnet_network_nodes.tsv	S# → SMILES for the nodes actually drawn, with their support
aggregate_reacnet_mechanistic_path.tsv	The highlighted route, one row per step. Header only when --pathway-mode none
aggregate_reacnet_pathlength_scan.tsv	Written only with --scan-pathlength: the best route of each exact length
aggregate_reacnet_report.html	Standalone page: run list, top reactions, top species. No dependencies
aggregate_reacnet_reacnetstyle.html	ReacNetGenerator's own interactive shell, with the aggregate substituted in and run-support labels added to each row
aggregate_reacnet_rngdata.json	The payload behind that page
aggregate_reacnet_failures.tsv	Per-run errors. Written only if something failed
aggregate_reacnet_warnings.tsv	Excluded runs and JSON-vs-timeline support gaps. Written only if there are any

Exit codes

Code	Meaning
0	At least one run was aggregated
2	Bad arguments, no matching inputs, or the InChI collapse failed
3	Ran, but no run succeeded (or, for --rebuild-from, something in the rebuild failed)

A single bad trajectory does not stop the ensemble: it is recorded in aggregate_reacnet_failures.tsv and the remaining runs are aggregated. An ensemble is worth reporting with 19 of 20 members — but check the denominator.

Intermediate files

Per-run ReacNetGenerator output lives in a temporary directory that is removed on exit. --keep-temp (with --tmp-root <dir> to control where) keeps it, one subdirectory per run, which is the way to inspect a per-run .species timeline or a ReacNet log. --save-noli-dir <dir> keeps just the Li-removed trajectories.