15_data_dictionary.md

15. Parquet data dictionary

Consolidated column reference for every artifact the MuMDIA engine writes. Every row here was taken from the write_table/Col construction in the stage source, not from prose. Citations are file:line.

Type legend

Columns are built from the Col enum in rust/mumdia/crates/mumdia-io/src/table.rs. Each variant maps to a fixed Arrow DataType and nullability in Col::field (table.rs:87-100):

Col variant	Arrow type	nullable
I64	Int64	no
I32	Int32	no
U32	UInt32	no
F64	Float64	no
F32	Float32	no
Bool	Boolean	no
Str	Utf8	no
OptF64	Float64	yes
OptF32	Float32	yes
OptI32	Int32	yes
OptStr	Utf8	yes
ListF32	List<Float32>	yes (list + item)
ListF64	List<Float64>	yes (list + item)
LargeListF32	LargeList<Float32>	yes (list + item)

All non-Opt, non-List columns are written as Arrow non-nullable (nullable = false). Below, "nullable" means the Arrow field nullability, not whether a value is semantically optional. Note that mumdia-io integer and bool getters do not check is_null (safe only because the non-nullable columns are never null in practice).

Non-Parquet sidecars written alongside artifacts (JSON manifests, the feature schema companion, the PIN, the human-readable TSV reports) are documented in their own sections at the end.

---

Stage 0: convert (stages/convert.rs)

spectra_ms1 (convert.rs:171-179)

column	Arrow type	nullable	units	meaning
scan_index	UInt32	no	-	monotonic 0-based scan index assigned in read order
rt_seconds	Float64	no	s	scan start time, converted from mzdata minutes
mz	List<Float32>	yes	m/z	centroided, m/z-sorted peak m/z values
intensity	List<Float32>	yes	counts	peak intensities aligned to mz

spectra_ms2 (convert.rs:198-213)

column	Arrow type	nullable	units	meaning
scan_index	UInt32	no	-	monotonic scan index (shared axis with MS1)
id	Utf8	no	-	vendor spectrum id string
rt_seconds	Float64	no	s	scan start time
window_id	UInt32	no	-	dense id of the distinct isolation window ((lower,upper) bits), matches isolation_windows.window_id
window_target	Float64	no	m/z	isolation-window target m/z (0.0 for AIF/all-ion)
window_lower	Float64	no	m/z	isolation-window lower bound (0.0 for AIF full-range fallback)
window_upper	Float64	no	m/z	isolation-window upper bound (1.0e6 for AIF full-range fallback)
precursor_mz	Float64	yes	m/z	selected precursor m/z, null when absent
precursor_charge	Int32	yes	-	precursor charge, null when absent
mz	List<Float32>	yes	m/z	centroided fragment m/z values
intensity	List<Float32>	yes	counts	fragment intensities aligned to mz

isolation_windows (convert.rs:215-226)

The distinct-window column is window_id (verified: convert.rs:218).

column	Arrow type	nullable	units	meaning
window_id	UInt32	no	-	dense window id (0-based first-seen order)
target	Float64	no	m/z	window target m/z
lower	Float64	no	m/z	window lower bound
upper	Float64	no	m/z	window upper bound

ms2_to_ms1 (convert.rs:228-234)

column	Arrow type	nullable	units	meaning
ms2_scan_index	UInt32	no	-	MS2 scan_index
ms1_scan_index	Int32	no	-	scan_index of the most recent preceding MS1 scan, or -1 if none

---

Stage A: digest (stages/digest.rs)

peptides (digest.rs:286-298)

column	Arrow type	nullable	units	meaning
id	UInt32	no	-	peptide row id (targets and decoys interleaved)
peptide	Utf8	no	-	stripped amino-acid sequence
protein	Utf8	no	-	;-joined protein accessions; decoys carry a DECOY_ prefix
start	Int32	no	-	0-based start offset in the source protein
end	Int32	no	-	0-based end offset (exclusive)
label	Utf8	no	-	target or decoy
target_id	Int32	no	-	for a decoy, the id of its paired target; -1 for a target
decoy_strategy	Utf8	no	-	lowercased decoy strategy name (e.g. reverse, scramble)

---

Stage A2: peptidoforms (stages/peptidoforms.rs)

peptidoforms (peptidoforms.rs:253-265)

column	Arrow type	nullable	units	meaning
id	UInt32	no	-	peptidoform row id (unique per mod/charge expansion)
peptide_id	UInt32	no	-	source peptides.id
base_peptide_id	UInt32	no	-	target id for grouping (a decoy points at its target); competition/FDR key
peptide	Utf8	no	-	stripped sequence (carried from peptides)
peptidoform	Utf8	no	-	ProForma-lite string with UniMod mod names in brackets
charge	Int32	no	-	precursor charge state
label	Utf8	no	-	target or decoy
protein	Utf8	no	-	protein accessions (carried from peptides)

---

Stage C: predict-frag (stages/predict_frag.rs)

fragment_library precursors (predict_frag.rs:197-211)

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	contiguous id, sorted by ascending precursor_mz (the fragment-index build key)
peptidoform_id	UInt32	no	-	source peptidoforms.id
base_peptide_id	UInt32	no	-	grouping key (target/decoy family)
peptidoform	Utf8	no	-	ProForma-lite peptidoform
charge	Int32	no	-	precursor charge
precursor_mz	Float64	no	m/z	precursor m/z from the shared mass model
predicted_irt	Float32	no	iRT units	predicted indexed retention time (native or DeepLC)
label	Utf8	no	-	target or decoy
protein	Utf8	no	-	protein accessions
n_fragments	Int32	no	-	number of fragments kept for this candidate after top-N

fragment_library fragments (predict_frag.rs:212-223)

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	precursor candidate_id this fragment belongs to
mz	Float64	no	m/z	theoretical fragment m/z
predicted_intensity	Float32	no	relative	predicted fragment intensity (native heuristic or MS2PIP)
name	Utf8	no	-	fragment name, e.g. b3, y7, b3^2
ion_type	Utf8	no	-	ion-type symbol (b/y)
ordinal	Int32	no	-	ion ordinal (residue position)
frag_charge	Int32	no	-	fragment charge state
cardinality	Int32	no	count	distinct library precursors sharing this fragment m/z (0.01 Da bin); low = clean/quantification-friendly ion, high = interference-prone. Precomputed at library build

---

Stage S: search-seed (stages/search_seed.rs)

seed_psms (search_seed.rs:233-250)

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	library candidate id
peptidoform	Utf8	no	-	ProForma-lite peptidoform
charge	Int32	no	-	precursor charge
precursor_mz	Float64	no	m/z	precursor m/z
base_peptide_id	UInt32	no	-	grouping key
protein	Utf8	no	-	protein accessions
label	Utf8	no	-	target or decoy
score	Float64	no	-	best Sage-lite hyperscore for this candidate
spectrum_q	Float64	no	-	target-decoy q-value (best per candidate)
observed_rt	Float64	no	s	RT of the best-scoring scan
predicted_irt	Float32	no	iRT units	library predicted iRT (carried)
matched_peaks	Int32	no	-	matched fragment count at the best scan
scan_index	UInt32	no	-	scan_index of the best-scoring MS2 scan

<seed>.masscal.json (search_seed.rs:217-227)

Per-run fragment mass recalibration sidecar consumed by extract.

field	JSON type	units	meaning
frag_ppm_offset	number	ppm	median systematic fragment mass offset (0.0 when < 20 calibrants)
frag_tol_ppm	number	ppm	learned fragment tolerance (95th percentile of centered deviations x1.5, floor 5.0); falls back to the configured tolerance when too few calibrants
frag_ppm_sigma	number	ppm	local mass-uncertainty estimate (equal to frag_tol_ppm)
n_dev	integer	-	number of matched-fragment ppm deviations collected
cal_passes	integer	-	calibration passes run (0 = fallback, 1 = single, 2 = robust two-pass)

---

Stage B: rt-im-train (stages/rt_im_train.rs)

run_windows (rt_im_train.rs:266-277)

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	library candidate id (one row per library candidate)
rt_pred_cal	Float64	no	s	predicted iRT calibrated to this run's observed RT
rt_lo	Float64	no	s	lower RT window bound (rt_pred_cal - width)
rt_hi	Float64	no	s	upper RT window bound (rt_pred_cal + width)
im_pred_cal	Float64	yes	1/K0	calibrated ion mobility; always null in the 3D MVP
im_lo	Float64	yes	1/K0	IM window lower bound; always null
im_hi	Float64	yes	1/K0	IM window upper bound; always null

<run_windows> cal.json (rt_im_train.rs:290-302)

RT-calibration sidecar (written to out_cal, not part of the Parquet contract).

field	JSON type	meaning
method	string	loess, linear, or unavailable (no calibration fit ran)
slope	number or null	linear fit slope (iRT -> s); null when calibration is unavailable
intercept	number or null	linear fit intercept (s); null when calibration is unavailable
w_rt	number	global RT half-window width (s)
p_rt	number	residual percentile used for the window
multiplier	number	RT window multiplier
n_train	integer	number of confident anchors used
calibration_status	string	loess, linear, or fallback_fixed

---

Stage D: extract (stages/extract.rs)

psms_extracted (extract.rs:1474-1495, base columns)

One row per accepted candidate (the selected apex PSM).

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	library candidate id
apex_rt	Float64	no	s	selected apex retention time
apex_im	Float64	yes	1/K0	apex ion mobility; always null in the 3D MVP
apex_intensity	Float32	no	counts	summed observed intensity of all matched fragments at the apex scan
n_matched_fragments	Int32	no	-	count of distinct matched predicted fragments
n_predicted_fragments	Int32	no	-	number of predicted fragments in the library
coelution_run	Int32	no	scans	longest consecutive-scan co-elution run
rt_pred_cal	Float64	no	s	calibrated predicted RT (from run_windows)
precursor_mz	Float64	no	m/z	precursor m/z
charge	Int32	no	-	precursor charge
label	Utf8	no	-	target or decoy
base_peptide_id	UInt32	no	-	grouping key
peptidoform	Utf8	no	-	ProForma-lite peptidoform
protein	Utf8	no	-	protein accessions
predicted_irt	Float32	no	iRT units	library predicted iRT
contested_frac	Float64	no	fraction	fraction of contested intensity lost to better co-eluters (0 when the two-pass path is off)
ms1_isom1	Float64	yes	counts	MS1 intensity at mono minus one isotope spacing; null when no MS1
ms1_mono	Float64	yes	counts	MS1 monoisotopic intensity near the apex; null when no MS1
ms1_iso1	Float64	yes	counts	MS1 +1 isotope intensity; null when no MS1
ms1_iso2	Float64	yes	counts	MS1 +2 isotope intensity; null when no MS1

Conditional columns, appended only when extract.emit_contested_features (extract.rs:1498-1501):

column	Arrow type	nullable	units	meaning
contested_count_frac	Float64	no	fraction	fraction of contested fragment-peak instances lost
apportioned_frac	Float64	no	fraction	fraction of contested intensity kept under proportional apportionment

Conditional columns, appended only when extract.emit_gate_diagnostics (extract.rs:1504-1509). All four gate scores are computed together only when this flag is set (extract.rs:1224-1229); when it is off they are not computed and the columns are not written, so the default schema stays byte-identical:

column	Arrow type	nullable	units	meaning
gate_apex	Float32	no	correlation	single-apex-scan intensity Pearson
gate_peak_spectral	Float32	no	correlation	peak-integrated spectral Pearson
gate_coelution	Float32	no	correlation	temporal co-elution score
gate_spectral_entropy	Float32	no	similarity	apex spectral-entropy similarity (sqrt)

chromatograms (extract.rs:1512-1526)

One row per predicted transition per accepted candidate, plus MS1 isotope XIC pseudo-rows (frag_name = ms1_mono/ms1_iso1/ms1_iso2) when MS1 + window grid are available.

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	library candidate id
frag_name	Utf8	no	-	fragment name (or ms1_mono/ms1_iso1/ms1_iso2)
frag_mz	Float64	no	m/z	theoretical fragment m/z (precursor m/z for MS1 rows)
frag_obs_mz	Float64	no	m/z	intensity-weighted observed m/z; falls back to frag_mz
predicted_intensity	Float32	no	relative	library predicted intensity (0.0 for MS1 rows)
rt	LargeList<Float32>	yes	s	XIC retention times (empty for a never-observed transition)
intensity	LargeList<Float32>	yes	counts	XIC intensities aligned to rt

rt/intensity use LargeList (64-bit offsets) because the total list-value count can exceed the 32-bit ListArray offset ceiling on large candidate sets.

<psms>.peaks.parquet, top-K peak retention (extract.rs:1532-1543)

Written next to the PSM table only when extract.retain_top_peaks > 1; one row per (candidate, peak).

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	library candidate id
peak_rank	Int32	no	-	peak rank by co-eluting fragment breadth-area
apex_rt	Float64	no	s	peak apex RT
start_rt	Float64	no	s	peak start RT
end_rt	Float64	no	s	peak end RT
evidence_count	Float64	no	-	distinct co-eluting fragment count at the apex
area	Float64	no	-	breadth-profile area of the peak

---

Stage E: features (stages/features.rs)

features.parquet (features.rs:913-932)

Bookkeeping columns followed by the active feature columns. Every feature column is Float64.

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	library candidate id
label	Utf8	no	-	target or decoy
base_peptide_id	UInt32	no	-	grouping key
peptidoform	Utf8	no	-	ProForma-lite peptidoform
protein	Utf8	no	-	protein accessions
apex_rt	Float64	no	s	selected apex RT
elution_lo	Float64	no	s	detected elution-peak lower RT bound
elution_hi	Float64	no	s	detected elution-peak upper RT bound
precursor_mz	Float64	no	m/z	precursor m/z
prelim_score	Float64	no	-	preliminary composite score (feature-derived; rescorer warm-start)
...feature columns...	Float64	no	varies	one column per name in active_features(set) (see below)

The active feature list depends on features.set (features.rs:210-231): Minimal (14), Rich (14 + 30 = 44), or Extended (14 + 30 + extended battery + 6 psms-derived; 381 columns total, the value of active_features(FeatureSet::Extended).len() confirmed by the runtime features=381 log. The feature_sets_sized test (features.rs:1590-1605) asserts the structural formula 14 + 30 + extended_names().len() + 6 and name uniqueness, not a literal count).

Minimal feature columns (features.rs:158-173): rt_error_abs, rt_error_rel, n_matched_fragments, coelution_run, log_apex_intensity, frag_corr, frag_cosine, spectral_angle, coelution_mean, coelution_best, n_coelution_above, charge, peptide_length, n_proteins.

Rich-extra feature columns (added for Rich and Extended, features.rs:176-207): library_norm_manhattan, library_rmsd, xcorr_coelution, xcorr_shape, sum_b_intensity, sum_y_intensity, diff_by_intensity, n_b_ions, n_y_ions, weighted_mass_error, mean_mass_error, isotope_corr, ms1_isom1_ratio, log_mono_ms1, has_ms1, log_sn, n_observations, base_width_rt, seed_score, seed_identified, matched_fraction, profile_cos, ref_corr, best_ref_corr, low_frag_coel, evidence, contrast_min, resid_corr, coel_clean, shadow_frac.

psms-derived Extended extras (added only for Extended, features.rs:220-228): peak_contested_frac, peak_contested_count_frac, peak_apportioned_frac, n_charge_states, charge_multi_flag, cross_charge_intensity_log.

Extended battery families. The 15 family modules under stages/features/ are concatenated in the registry order at features.rs:52-68. Each family exposes a NAMES: &[&str] array; the registry drops any name colliding with a Minimal/Rich name and any cross-family duplicate, keeping the first appearance (features.rs:84-99). The raw per-family names are listed below (raw total 335; the deduplicated survivors are the Extended battery columns).

• similarity (similarity.rs:18-96, 64): spectrum_cosine_matched, spectrum_cosine_sqrt, spectrum_cosine_log, spectral_angle, spectral_angle_sqrt, spectral_angle_matched, pearson_intensity_matched, pearson_intensity_log, spearman_intensity, spearman_intensity_matched, kendall_tau_intensity, dot_product_raw, dot_product_norm, library_recall_intensity, manhattan_sim, manhattan_sqrt, rmsd_norm, mae_norm, mse_log, mae_weighted_pred, abs_diff_q3, max_positive_residual, chebyshev_dist, minkowski_p3, bray_curtis, bray_curtis_sqrt, canberra, canberra_matched, wave_hedges, chi_square_pearson, chi_square_symmetric, divergence_distance, bhattacharyya_coef, hellinger, squared_chord, harmonic_mean_sim, jaccard_presence, dice_presence, intensity_weighted_pearson, regression_slope, gini_diff, wasserstein_mz, footrule_norm, rank_overlap_top3, top1_frag_match, top1_predicted_observed, frac_top3_predicted_observed, count_strong_predicted_absent, frac_predicted_absent, cosine_area, pearson_area, spectral_angle_area, cosine_fullwindow, stein_scott_weighted_dot, log_dot_product, spectral_log_evidence, scribe_score, log_dot_product_area, spectral_log_evidence_area, scribe_score_area, cosine_high_ordinal, cosine_robust_trim1, cosine_robust_trim2, cosine_robust_trim3.
• entropy (entropy.rs:13-32, 18): spectral_entropy_similarity, weighted_spectral_entropy_similarity, spectral_entropy_similarity_sqrt, spectral_entropy_similarity_topk, spectral_entropy_similarity_area, jensen_shannon_divergence, jeffreys_divergence, kl_obs_pred, kl_pred_obs, cross_entropy_obs_pred, obs_spectrum_entropy, pred_spectrum_entropy, entropy_diff, entropy_ratio, obs_normalized_entropy, normalized_entropy_diff, residual_spectrum_entropy, entropy_weight_obs.
• coelution (coelution.rs:21-60, 38): frag_ref_corr_mean, frag_ref_corr_obsweighted, frag_ref_corr_min, frag_ref_corr_std, frag_ref_corr_sq_mean, frag_ref_corr_topk_weighted, n_frag_ref_corr_above_0_9, frac_frag_ref_corr_above_0_8, frag_ref_corr_mean_full, full_vs_peak_corr_gain, pairwise_coelution_weighted, pairwise_coelution_min, pairwise_coelution_median, pairwise_coelution_std, pairwise_coelution_frac_negative, pairwise_coelution_hi, pairwise_coelution_lo, coelution_hi_lo_contrast, coelution_corr_entropy, xcorr_shape_mean, xcorr_shape_min, xcorr_shape_std, xcorr_lag_mean_abs, xcorr_lag_std, xcorr_lag_iqr, xcorr_lag_frac_zero, xcorr_lag_max_abs, xcorr_lag_entropy, ref_xcorr_lag_mean, ref_xcorr_shape_mean, observed_sum_vs_template_corr, frag_loo_ref_corr_mean, frag_loo_ref_corr_min, frac_frags_apex_aligned, top3_frag_ref_corr, by_cross_coelution, by_cross_lag_mean, charge_cross_coelution.
• interference (interference.rs:17-44, 26): explained_variance_ref, profile_residual_fraction, n_interfered_fragments, corrected_vs_raw_cos, corrected_vs_raw_ratio, ifs_removed_count, ifs_removed_intensity_frac, ifs_corr_gain, ifs_retained_frac, matched_frac_after_ifs, peak_to_full_area_ratio_profile, peak_to_full_area_ratio_frag_mean, peak_to_full_area_ratio_weighted, out_of_peak_intensity_frac, profile_corr_full_vs_peak_delta, frac_frag_ref_corr_below_0_5, explained_apex_intensity_frac, apex_purity, interference_apex_residual_fraction, dominant_frag_ref_corr, explained_variance_ratio, second_component_fraction, profile_second_peak_ratio, n_competing_peaks_in_window, matched_pred_intensity_fraction, top_pred_frag_matched.
• chromatographic (chromatographic.rs:20-64, 43): gaussian_fit_r2, gaussian_cosine, emg_fit_improvement, apex_prominence, profile_peak_snr, fwhm_seconds, fwhm_to_window_ratio, width_at_10pct, width_ratio_10_50, hwhm_asymmetry, tailing_factor_usp, asymmetry_factor_10pct, apex_sharpness, apex_curvature, apex_to_boundary_ratio, apex_dominance, zigzag_index, jaggedness, roughness_2nd_deriv, n_local_maxima, modality, rt_skewness, rt_excess_kurtosis, rt_std_seconds, mean_mode_offset, fraction_area_within_fwhm, triangle_area_similarity, baseline_fraction, peak_completeness, apex_centering_offset, intensity_score, total_xic_log, frag_fwhm_cv, frag_fwhm_mean, frag_apex_rt_dispersion, frag_apex_rt_dispersion_weighted, frag_apex_offset_from_profile_mean, frag_gaussianity_mean, frag_gaussianity_weighted, frag_zigzag_mean, sumtrace_unweighted_gaussian_r2, reference_profile_rt_entropy_peak, reference_profile_rt_entropy_ratio.
• mass_accuracy (mass_accuracy.rs:19-38, 17): median_abs_frag_ppm, signed_mean_frag_ppm, ppm_std, ppm_iqr, ppm_range, max_abs_frag_ppm, intensity_weighted_abs_ppm, intensity_weighted_signed_ppm, intensity_weighted_ppm_std, lib_weighted_abs_ppm, frac_frag_within_half_tol, high_ppm_intensity_frac, ppm_intensity_anticorr, mass_error_mz_trend, mean_abs_mz_error_da, mass_evidence_gauss, mass_log_evidence.
• ion_series (ion_series.rs:19-54, 34): n_matched_b, n_matched_y, frac_matched_b, frac_matched_y, by_count_balance, by_intensity_ratio, by_ratio_agreement, by_ratio_consistency, longest_b_run, longest_y_run, longest_run_max, longest_run_frac_length, series_coverage_b, series_coverage_y, sequence_coverage, series_gap_fraction, by_complement_count, by_complement_mz_consistency, by_complement_coelution, ordinal_intensity_concordance_y, ordinal_intensity_concordance_b, series_coelution_y, series_coelution_b, spectral_angle_b, spectral_angle_y, pearson_b, pearson_y, cosine_charge1, cosine_charge2, charge_corr_balance, mean_matched_ordinal_norm, by_ion_contiguous_intensity, by_ion_contiguous_lib_frac, both_series_present.
• ms1 (ms1.rs:18-49, 25): ms1_isotope_cosine_apex, ms1_isotope_spectral_angle_apex, ms1_isotope_chi2_apex, ms1_isotope_manhattan_apex, iso_ratio_1_0, iso_ratio_2_0, iso_plus1_ratio_dev, iso_plus2_ratio_dev, iso_minus_one_fraction, iso_overlap_flag, log_ms1_mono, ms1_total_isotope_log, has_ms1_signal, ms1_isotope_apex_entropy_3, ms1_m1_entropy_contribution, ms1_ms2_time_corr, ms1_ms2_envelope_time_corr, ms1_iso_coelution, ms1_ms2_apex_rt_delta, ms1_iso_ratio_stability, ms1_mono_gaussianity, ms1_ms2_fwhm_ratio, ms1_isotope_corr_xic, ms1_envelope_over_time_corr, ms1_isotope_xic_shape_consistency.
• rt (rt.rs:12-26, 13): rt_error_signed, rt_error_abs, rt_error_squared, rt_error_signed_norm_gradient, rt_error_abs_norm_gradient, observed_rt_raw, predicted_rt_raw, observed_rt_fraction, predicted_rt_fraction, rt_error_over_peak_width, rt_error_over_fwhm, rt_diff_profile_apex, predicted_rt_in_gradient.
• novel (novel.rs:16-29, 12): log_seed_hyperscore, seed_hyperscore_per_matched, seed_identified, peptide_length, precursor_charge, charge_is_2, charge_is_3, charge_is_4plus, precursor_mass, log_total_matched_intensity, n_matched_frags, n_predicted_frags.
• nonzero (nonzero.rs:17-30, 12): frag_corr_peakmax, frag_cosine_peakmax, spectral_angle_peakmax, frag_corr_matched_nz, frag_cosine_matched_nz, peakmax_apex_gain, n_frag_present_inpeak, frac_frag_present_inpeak, coelution_mean_bothpos, coelution_mean_summpos, ref_corr_nz, profile_cos_nz.
• order_consistency (order_consistency.rs:27-36, 8): rank_corr_vs_apex_mean, rank_corr_vs_apex_std, rank_corr_adjacent_mean, kendall_vs_apex_mean, top1_frag_persistence, top2_order_persistence, argmax_frag_entropy, self_cosine_vs_apex_mean.
• peak_scans (peak_scans.rs:16, 2): n_peak_scans, peak_window_degenerate.
• apex_dispersion (apex_dispersion.rs:15-29, 13): frag_apex_rt_std, frag_apex_rt_mad, frag_apex_max_dev, frag_apex_mean_dev, frag_apex_agree_frac, precursor_frag_apex_delta, peak_symmetry, peak_tailing, peak_n_local_maxima, peak_shoulder_score, peak_fwhm_scans, peak_truncation, apex_frac_of_window.
• mass_uncertainty (mass_uncertainty.rs:18-29, 10): frag_mass_err_median, frag_mass_err_abs_median, frag_mass_err_std, frag_mass_err_iqr, frag_mass_err_max_abs, frag_mass_err_range, effective_frag_count, evidence_concentration, frac_top3_pred_observed, frac_top5_pred_observed.

<features>.schema.json (features.rs:936-942)

Feature schema companion carried forward to compete/rescore so the classifier input is reproducible.

field	JSON type	meaning
feature_columns	array of strings	ordered active feature list
schema_id	string	blake3 hash of the joined column list

PIN (features.rs:1537-1573)

Percolator input, tab-separated text (written to out_pin). Header: SpecId Label ScanNr ExpMass CalcMass <feature cols...> Peptide Proteins.

field	meaning
SpecId	cand_<candidate_id>
Label	1 for target, -1 for decoy
ScanNr	candidate_id
ExpMass	precursor m/z (5 decimals)
CalcMass	precursor m/z (5 decimals; same value)
feature columns	active feature values (6 decimals), one per feature
Peptide	-.<peptidoform>.-
Proteins	protein-accession string

---

Stage F: compete (stages/compete.rs)

psms_competed (compete.rs:121-151)

Bookkeeping columns plus the feature columns carried from the schema (each Float64). A <out>.schema.json companion is copied forward (compete.rs:153).

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	library candidate id
label	Utf8	no	-	target or decoy
base_peptide_id	UInt32	no	-	grouping key
peptidoform	Utf8	no	-	ProForma-lite peptidoform
protein	Utf8	no	-	protein accessions
apex_rt	Float64	no	s	apex RT
elution_lo	Float64	no	s	identified elution lower bound
elution_hi	Float64	no	s	identified elution upper bound
precursor_mz	Float64	no	m/z	precursor m/z
prelim_score	Float64	no	-	preliminary composite score
...feature columns...	Float64	no	varies	each feature_columns entry from the schema

<out>.compete_audit.parquet (compete.rs:167-202)

Conditional, written only when compete.emit_competition_audit; one row per candidate removed by within-group competition.

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	removed candidate id
label	Utf8	no	-	target or decoy (same label as its winner)
peptidoform	Utf8	no	-	removed peptidoform
winner_candidate_id	UInt32	no	-	candidate id that outcompeted it
loser_prelim	Float64	no	-	loser preliminary score
winner_prelim	Float64	no	-	winner preliminary score
rejection_reason	Utf8	no	-	OUTCOMPETED_BY_TARGET or OUTCOMPETED_BY_DECOY

---

Stage F: rescore (stages/rescore.rs)

psms_scored (schema v3, rescore.rs:447-476)

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	library candidate id
peptidoform	Utf8	no	-	ProForma-lite peptidoform
charge	Int32	no	-	precursor charge
label	Utf8	no	-	target or decoy
protein	Utf8	no	-	protein accessions
base_peptide_id	UInt32	no	-	grouping key
apex_rt	Float64	no	s	identification apex carried through for quant
elution_lo	Float64	no	s	identified elution lower bound
elution_hi	Float64	no	s	identified elution upper bound
score	Float64	no	-	rescorer discriminant score (higher is better)
q_value	Float64	no	-	PSM-level q-value (pooled target-decoy or entrapment null)
peptide_q_value	Float64	no	-	peptide-level q (best PSM per base peptide); losers get 1.0
protein_group	Utf8	no	-	protein-accession-set string (decoys carry DECOY_)
pg_q_value	Float64	no	-	protein-group-level q
global_q_value	Float64	no	-	byte-identical alias of pooled q_value (backward-compat)
prelim_score	Float64	no	-	preliminary composite score carried forward
source	UInt32	no	-	index into --competed inputs (run identity); all-zero single-run
run_psm_q	Float64	no	-	per-run PSM q (TDA within each source)
experiment_psm_q	Float64	no	-	pooled PSM q (equals q_value)
precursor_q	Float64	no	-	per-precursor q (grouped on peptidoform + charge)

Which null the q-values use depends on the classifier: target-decoy for NativeTda/Mokapot/NnTorch/Percolator, entrapment for the Entrapment rescorer. The q-value branch is at rescore.rs:289-297; the qmode selector defaults to target-decoy (rescore.rs:145) and switches to entrapment only inside the Entrapment classifier arm (rescore.rs:252-273).

---

Stage G: quant (stages/quant.rs)

peptide_quant (schema v2, quant.rs:487-502)

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	library candidate id
base_peptide_id	UInt32	no	-	unique base-peptide rollup key
peptidoform	Utf8	no	-	ProForma-lite peptidoform
charge	Int32	no	-	precursor charge
protein_group	Utf8	no	-	protein-accession-set string
quantity	Float64	yes	intensity x s	sum of top-N positive finite fragment areas; null when not quantifiable
quant_status	Utf8	no	-	quantified or explicit missing-quantity reason
n_fragments_used	Int32	no	-	number of positive finite fragments summed
integration_apex_rt	Float64	yes	s	apex actually used for integration
integration_lo_rt	Float64	yes	s	integration lower bound actually used
integration_hi_rt	Float64	yes	s	integration upper bound actually used

protein_group_quant (schema v2, quant.rs:519-527)

column	Arrow type	nullable	units	meaning
protein_group	Utf8	no	-	protein-accession-set string
quantity	Float64	yes	intensity x s	rollup over unique positive base-peptide representatives
quant_status	Utf8	no	-	quantified or no_quantifiable_peptide
n_peptides	Int32	no	-	unique positive base peptides before Top-N truncation

fragment_quant (optional, quant.rs:558-568)

Written only when out_fragment is set (ion-level directLFQ input).

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	library candidate id
peptidoform	Utf8	no	-	ProForma-lite peptidoform
charge	Int32	no	-	precursor charge
protein_group	Utf8	no	-	protein-accession-set string
fragment_name	Utf8	no	-	fragment name
quantity	Float64	no	intensity x s	per-fragment trapezoid area over the peak window

peak-bounds diagnostic (optional, quant.rs:421-429)

Written only when out_peak_bounds is set and bound_peak is on.

column	Arrow type	nullable	units	meaning
candidate_id	UInt32	no	-	library candidate id
lo_rt	Float64	no	s	integration window lower RT
hi_rt	Float64	no	s	integration window upper RT
width_s	Float64	no	s	hi_rt - lo_rt

quant-lfq cross-run output (quant.rs:720-728, run_lfq_combine at quant.rs:657)

Long-form MaxLFQ (peptide-level) or directLFQ (ion-level) matrix.

column	Arrow type	nullable	units	meaning
protein_group	Utf8	no	-	protein-accession-set string
run	Int32	no	-	input run index (0-based)
quantity	Float64	no	intensity	LFQ profile value for that protein in that run
n_features	Int32	no	-	number of features contributing to the protein

---

Stage D2: align (stages/align.rs)

Experiment-level cross-run stage that prepares MBR. It puts every run on a common RT coordinate: the first input seed is the reference, and for each run it fits a monotone LOESS mapping from the run's observed RT to the reference RT over confidently shared peptides, then emits that mapping on a grid_n-point grid. The reference run (or a run with fewer than four shared peptides) uses the identity mapping. This artifact is written with an inline alignment (v1) ArtifactReport (align.rs:141-153); it is not registered in the frozen schema.rs artifact table.

alignment (align.rs:130-138)

One row per (run, grid point).

column	Arrow type	nullable	units	meaning
run_id	UInt32	no	-	0-based input run index (position in the seed list); run 0 is the reference
source_rt	Float64	no	s	grid point in this run's own observed RT coordinate
reference_rt	Float64	no	s	RT mapped onto the reference coordinate; identity for the reference run or when fewer than four peptides are shared with the reference
residual_spread	Float64	no	s	95th-percentile absolute LOESS residual on shared peptides for this run; 0.0 for the reference or an unfit run; bounds how tight an MBR window can be

---

audit (stages/audit.rs)

candidate_audit.parquet (audit.rs:180-200)

One row per library candidate (the full search space), with per-stage survival flags and the earliest rejection reason.

column	Arrow type	nullable	units	meaning
run_id	Utf8	no	-	run identifier stamped on every row
precursor_id	UInt32	no	-	library candidate_id
modified_sequence	Utf8	no	-	ProForma-lite peptidoform
charge	Int32	no	-	precursor charge
target_decoy_label	Utf8	no	-	target or decoy
entrapment_label	Boolean	no	-	protein contains the entrapment substring
candidate_generated	Boolean	no	-	in the search space (always true)
traces_extracted	Boolean	no	-	produced fragment traces in extract
peak_generated	Boolean	no	-	an accepted peak exists (equals traces_extracted at artifact resolution)
peak_selected	Boolean	no	-	selected as the reported peak (same resolution)
variant_selected	Boolean	no	-	survived within-group competition
target_decoy_winner	Boolean	no	-	present in the scored table
passed_precursor_fdr	Boolean	no	-	q-value <= threshold
passed_peptide_fdr	Boolean	no	-	peptide-q and precursor gate both pass
reported	Boolean	no	-	passed the precursor FDR gate
rejection_reason	Utf8	no	-	earliest-loss reason code (e.g. REPORTED, NO_PEAK_GROUP, OUTCOMPETED_BY_TARGET, FAILED_PRECURSOR_FDR)

<audit>.metrics.json (audit.rs:206-216)

field	JSON type	meaning
run_id	string	run identifier
q_threshold	number	FDR threshold used
search_space	integer	total library candidates
extracted	integer	candidates with accepted peaks
competed	integer	competition survivors
reported	integer	candidates reported at threshold
trace_recall	number	extracted / search_space
waterfall	object	map of rejection-reason code to count

---

Cross-stage JSON sidecars

<artifact>.report.json (mumdia-io/src/report.rs:10-24)

Written next to most primary stage Parquets, but not every Parquet or stage output. Optional diagnostics, schema/PIN files, TSV reports, and some experiment-level/Python-written outputs have partial or no coverage.

field	JSON type	meaning
logical_name	string	artifact logical name (e.g. spectra_ms2)
schema_name	string	frozen schema name
schema_version	integer	schema version (from mumdia-core/src/schema.rs)
stage	string	producing stage (e.g. convert, extract)
rows	integer	row count
content_hash	string	blake3 hash of the artifact file
params	object	resolved parameters the stage actually used
stats	object	summary key distributions / metrics
model_identity	string or null	predictor/rescorer model identity where applicable
elapsed_ms	integer	stage wall-clock time (ms)

manifest.json (mumdia-core/src/manifest.rs:9-31)

Written by the run orchestrator (stages/run.rs:500-501). Records per-artifact provenance for reproducibility. Top-level Manifest:

field	JSON type	meaning
mumdia_version	string	crate version
config_json	string	fully-resolved config JSON
config_hash	string	hash of the resolved config
model_identities	object	map of role (rt_predictor, fragment_predictor, rescorer, feature_schema_id) to identity string
artifacts	object	map of logical name to ArtifactRecord

Each ArtifactRecord:

field	JSON type	meaning
logical_name	string	artifact logical name
path	string	output path
format	string	file format (Parquet)
schema_name	string	schema name
schema_version	integer	schema version
rows	integer	row count
content_hash	string	blake3 hash
producing_stage	string	stage that wrote it
config_hash	string	config hash at write time

---

Not covered here

report writes peptides.tsv and proteins.tsv (stages/run.rs:459-461), which are human-readable TSV, not Parquet, and are outside this Parquet data dictionary. Sidecar worker I/O in the working directory (the Mokapot/NN PIN, the entrapment GBM in/out Parquet, MS2PIP/DeepLC exchange files) are transient contract files, not engine artifacts.

Schema-version registry

mumdia-core/src/schema.rs freezes the (logical name, version) pairs. psms_competed, peptide_quant, and protein_group_quant are v2; psms_scored is v3; the remaining registered artifacts are v1.