Bring docs current with v3.1.0: remove stale q/P_adj, add bg_fdr glossary entry
The q / P_adj / P_adj_lo / P_adj_hi columns were removed from score_info.iic in
v3.1.0, but three pages still described them as live and load-bearing:
- Glossary "q / P_adj (legacy)" claimed they "still feed adjusted_score/rel_score/
type_id, so the chain remains load-bearing" — now a removal note, with the
type_id 38 -> 35 position change called out.
- Technical-algorithm said they were "retained only for backward compatibility".
- Both pages gave adjusted_score = 100*P_adj; it is computed directly as
100*q_eff*P_motif.
Also adds a bg_fdr glossary entry (it was referenced from three pages but never
defined), and updates the version strings on Home, Glossary, Technical-algorithm
and Training-data-and-PWMs from 3.0.0 to 3.1.0.
The illustrative CLI banner in Example-usage still reads v3.0.0; the surrounding
text already disclaims version-specific output, so it is left as-is.
Document the v3.1 low-k escape hatch; broaden INCONCLUSIVE
The strength gate requires >= 3 calls, and z_excess is bounded below the loss
ceiling at low call counts by construction (one call is its own call-core, so
z_excess <= 0; two calls give <= 1). A genome with one or two strong calls could
therefore not reach DETECTED by any path and fell through to NOT_DETECTED, which
zeroes every score. Since v3.1 such a genome resolves to INCONCLUSIVE when its
strongest call clears min(bg_fdr) <= 3e-3, so its scores survive for downstream
corroboration.
- Technical-algorithm: hatch added to the gate pseudocode in evaluation order,
plus a section on why it exists, why INCONCLUSIVE rather than DETECTED, and why
the threshold is a trust line rather than a separator.
- Glossary: INCONCLUSIVE now covers both routes, and states explicitly that the
label no longer implies 2.60 < z_excess < 5.50.
- Output-files: bg_fdr is still report-only per intron, but the per-genome minimum
now feeds the hatch, so "never contributes to any call" was inaccurate.
Remove "IPA" jargon; use plain "comparative genomic" wording throughout
"IPA" (Intron Position Analysis) is internal shorthand. Delete the redundant
Glossary "IPA" entry (its content is covered by the existing "comparative
genomics labeling" entry), redirect the links there, and replace inline
"IPA-conserved" / "Intron Position Analysis" uses across Training-data-and-PWMs
and Technical-algorithm with "comparative genomic" phrasing.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
wiki: trim deprecated-metric clutter + fix v2.4.x/blockquote + CDS-relative frac_pos
- Minimize scattered references to removed v2.5-2.7 mechanisms (mode-separation,
two-pass/126-model ensembles, continuous-discount parentheticals); keep one
high-level z-normalization-removal statement (Technical Stage 3) and the single
pre-zstack-removal note. Legacy q/P_adj output columns preserved.
- Training-data: drop the confusing "v2.4.1" version stamps and unwrap the top
blockquote (v3.0.0 model / PWMs / corpus) into normal paragraphs.
- Update frac_pos everywhere for the new CDS-relative definition: description
(Output-files, Glossary, About), and example rows 0.035 -> 0.039 (Output-files,
Quick-start). Exon-only (UTR) introns are now NA.
wiki: reduce em-dash overuse + trim residual redundancy
Editorial pass across all pages to cut an AI-tell writing pattern (heavy
em-dash use) and tighten repetition. Em-dash count roughly halved
(~169 -> ~75): "dramatic pause" and all-purpose-connector dashes converted to
periods, commas, colons, or parentheses; term:definition list separators
normalized to colons. Punctuation-only (75 insertions / 75 deletions);
no facts, numbers, thresholds, formulas, code, tables, example rows, image
links, DOIs, or headings/anchors changed.
Redundancy:
- Technical: the U12-absent loss-FP statistic (19 genomes, median 3/max 15)
was stated in full twice; the second instance now cross-references the
Evaluation section.
- Overview: the "Brief method summary" re-described the full architecture
(ensemble, P_motif, adjudicator, corpus) already given under Classification
method just above; trimmed to a cross-reference.
Left in place (judged page-appropriate, not redundancy to strip): the P_motif
formula in the self-contained Glossary; short architecture blurbs on each
landing page; the streaming/in-memory "bit-identical" note where each page is
a plausible entry point.
wiki: decouple from DB framing + backtick tool/code refs
- Reframe the two "snRNA/database layer" hand-offs in the adjudicator docs:
intronIC has no database, so NOT_DETECTED's motif-silent bearers are now
described as needing orthogonal evidence (minor-snRNA presence / phylogenetic
context) rather than being "deferred to the database layer". snRNA and
comparative-genomics evidence remain where they legitimately apply
(anchor calibration, corpus labeling).
- Backtick `intronIC` consistently in prose (35 occurrences) and `sklearn`,
per standard technical-doc style. The H1 project title is left plain.
wiki: v3 editorial pass — glossary, citations, terms, neutral tone
Second-pass edit of the v3 wiki (follows the accuracy audit in the prior
commit). Editorial + reference corrections, not an architecture rewrite:
- New Glossary.md (62 terms / 6 sections) + _Sidebar entry; every page now
links technical nomenclature to it at first use.
- Define terms at first use: bearer, U12-loss, hard negative, EVT/Gumbel,
Poisson significance, AUC, ECE, IPA, leave-clade-out, singleton-decay
filter, PPT; unify decision_function/decision distance → "ensemble margin".
- Document magic-number derivations inline (Platt 2.796/-1.178, loss_ceiling_z
2.60, bearer_floor_z 5.50, p_gumbel_p95 0.01, cs_p95 5.0, min_u2 200).
- Fix three DOIs pointing at the wrong paper: Alioto gkl842->gkl796,
Sheth gkj456->gkl556, Mercer gr.182477.114->gr.182899.114; add Coles 2001
(EVT) and the Larue & Roy 2023 IPA-labeling citation.
- Consolidate repeated content (human benchmark table, bit-identical note,
P_motif formula, rel_score, call rule, corpus stats) to one canonical home
each + links elsewhere.
- Neutral academic tone throughout (drop marketing adjectives, exclamations,
product-launch headings); adhere to "U12-type"/"U2-type" for bare nouns.