TEagle 3.8.1

TEagle 3.8.1 is a corrective release for whole-genome landscape integrity, provenance, custom-library behavior, installer cleanup, legacy server mode, and export metadata.

Highlights:

• Computes all-repeat and TE coverage with contig-wide interval unions, so overlaps between repeat names/models are not double-counted. Whole-genome results now separate RepeatMasker repeat name/model (column 10) from class/family (column 11), treat PLE calls as TE, and display independent, non-additive union bars.
• Revalidates cached genomes, chunks, and completed outputs before reuse and finalization; binds the verified genome SHA-256 and length into the provenance seal; and prevents custom-FASTA runs from claiming Dfam evidence.
• Constrains clean-install and uninstall cleanup to the valid TEagle-recorded WSL distribution, with explicit partial-failure handling. Legacy loopback server startup and custom-library pre-run messaging are also corrected.
• Removes the optional Pillow dependency from PNG metadata checks and enforces minimal author-only export metadata.

The version-locked v3.8.0 benchmark evidence remains archived separately at https://doi.org/10.5281/zenodo.21848814. Corrected interval-union semantics can change whole-genome coverage totals relative to 3.8.0.

Uploaded asset: TEagle-Setup-3.8.1.exe only (40,545,260 bytes).

SHA-256: D731BD3066136F6E540F6A10256D328FDB37321613206CBD58CFFA29B2A86375

The Windows installer is unsigned, so SmartScreen may warn or managed devices may block it. Linux-backed workflows use the optional WSL2 environment.

TEagle 3.8.0

TEagle 3.8.0

Release date: 9 August 2026
Platform: Windows 10/11, 64-bit
License: GNU Affero General Public License v3.0 or later (AGPL-3.0-or-later)
Benchmark evidence: https://doi.org/10.5281/zenodo.21848814

TEagle 3.8.0 provides an evidence-traceable integrated workflow for examining one transposable-element sequence, assessing structural and domain evidence, planning a PCR assay, screening predicted products and recording the inputs, versions, thresholds and checksums behind the result. The component algorithms are established methods. The contribution lies in their integration, explicit evidence contracts, reasoned withholding, sealed provenance and native Windows interface; this release does not claim a new general-purpose classifier.

Release status

Version 3.8.0 supersedes version 3.7.0, which was retracted on 7 August 2026. One bundled development accession had entered the corpus described as held out, and two thresholds that affected calls were absent from the provenance manifest. The v3.7.0 benchmark figures and release assets should not be used. Version 3.8.0 removes the development accession from that corpus and seals the two omitted thresholds.

Every analysis-manifest hash changes in v3.8.0 because the sealed parameter set grew by two. Release review also corrected a LINE corroboration guard: a raw endonuclease hit no longer supports a LINE call unless it is N-terminal to reverse transcriptase on the same strand. Primer rules did not change between v3.7.0 and v3.8.0.

Changes in v3.8.0

• The optional WSL2 backend now requests exact releases of RepeatMasker, minimap2, isPcr, NCBI Datasets, faToTwoBit and ViennaRNA. Status and integrity checks reject version drift, and the micromamba bootstrap verifies its versioned archive before extraction.
• Backend cleanup now removes only TEagle-owned environments, caches, logs, staging files and locks. It preserves the shared micromamba executable and environments belonging to other applications.
• Rebuilding the Windows bundle now handles read-only metadata generated by an earlier PyInstaller build without changing files outside the generated bundle.
• Result-table context menus distinguish copying the selected cell from copying the complete row and use the same machine value written to exports.
• Saved or test primer queues remain readable when optional display metadata is absent; the interface states that the value is missing.
• The provenance manifest now records the standalone terminal-inverted-repeat length threshold and the terminal-repeat seed-search window. An automated coverage check now fails if an applied numeric threshold is absent from the manifest.
• The LINE guard now accepts endonuclease as ORF2p corroboration only when its position and strand form the expected N-terminal EN–RT arrangement; downstream and opposite-strand hits no longer license a LINE call.
• The held-out validation corpus excludes all five bundled development examples. It now contains 43 records carrying 36 distinct literal organism annotations.

The changes introduced in v3.7.0 and retained here include stricter treatment of an uncorroborated reverse-transcriptase hit, corrected benchmark accounting, explicit scoring of abstention, 3′-anchored matching of primers with non-templated 5′ tails, complete sealing of in-silico PCR defaults, retrospective reconstruction of published assays and a paired comparison with TEsorter 1.5.1. The changelog gives the complete change history and separates current v3.8.0 evidence from superseded historical results.

Interpretation and limitations

• The benchmark corpora are curated and non-random. Their results describe the included cases and do not estimate population-level accuracy for arbitrary transposable elements.
• Accuracy among answered cases excludes abstentions. Answered-case accuracy must therefore be read with the corresponding abstention denominator and case count.
• In the paired comparison corpus, neither TEagle nor TEsorter made an incorrect order call. TEagle withheld some calls that TEsorter answered correctly, so this corpus measures a cost of withholding but cannot show that withholding prevents errors.
• Retrospective replay of published primer pairs assesses the plausibility of in-silico product prediction for established assays. It does not provide prospective wet-lab validation of primers designed by TEagle and does not establish de novo primer-design success.
• The evidence-traceable classification and assay workflow analyses one locus at a time. The optional landscape module applies a repeat library to a whole genome, but the single-locus benchmark does not validate genome-scale classification performance. TEagle does not perform de novo family discovery, library construction or consensus inference.
• The bundled domain-profile panel is retroviral-tuned. Completeness calls for plant elements may therefore be conservative.
• The installer and executable are not code-signed. Windows SmartScreen may warn on first launch, and managed institutional systems may require approval from local information-technology staff.
• The frozen executable and GUI were exercised on the development Windows host. Clean-machine installation, uninstallation and SmartScreen behaviour were not independently tested for this release.

Distribution

The manually uploaded GitHub release assets are the version-locked Windows installer and the controlled user manual. The release page records the SHA-256 checksum of each asset. The sanitized benchmark evidence is a separate immutable Zenodo record at the DOI above; it is not a software-source archive.

SHA-256 checksums

• TEagle-Setup-3.8.0.exe: A25FDECF27749BF3DDF9D9B2A6C3587F09F8913825BD7B1E999D581FC6D38F39
• TEagle-User-Manual.pdf: 738A96EAEFA44D988872BAC7555013EAE8F8D364AA4457C69CFDAA91959CF746

TEagle v3.6.0

Two development cycles reach release together: 3.5.0 was an internal step and was never tagged, so this release carries both. Provenance manifest hashes change, and results change for elements whose diagnostic structure was not detected. No element loses a call and no superfamily assignment changes.

Classification: four order-level defects, found by benchmarking against an external classifier

A chromodomain now names the LTR order. Chromoviruses are a Ty3/Gypsy lineage and non-LTR elements do not carry the domain, so its presence is informative on its own.

A lone reverse transcriptase is no longer called a LINE. The LINE branch read an absent integrase as evidence of absence rather than as no evidence, which is a different claim.

A terminal inverted repeat can now name a non-autonomous Class II element. Requiring a transposase excluded exactly the elements that do not encode one.

Over the benchmark corpus, order-level accuracy rose from 0.855 to 0.980 and negative-control abstention from 72.7% to 90.9%. Current figures after corpus restratification are 0.986 (0.922-0.997) over 94 gradable cases; benchmarks/raw/scores.json is the authority.

Assay path

A primer carrying a non-templated 5' tail could bind for the first time. The matcher required the whole oligo to match the template, so any pair with a restriction site, adapter or promoter tail predicted no product at all. Binding is now anchored at the 3' end.

In-silico PCR sealed only the parameters the caller typed, so a run left entirely at defaults recorded almost nothing. It now seals the resolved values.

Provenance

The seal recorded twenty of the twenty-one thresholds that decide a call: the upper bound on terminal inverted-repeat length was applied but never written, so a run could not be reconstructed from its own seal. It is now sealed, and a benchmark enumerates every numeric default across the seven detectors the seal draws from and fails if any is unaccounted for.

The stated lower limit of terminal-repeat detection was wrong. The card claimed k-mer seeding stops below roughly 72% identity. That came from a single rough test and is not a constant, because the limit depends on repeat length. Measured over 600 controlled runs on a 574 bp LTR, candidates were still reported at 65% and none at 60%. The card now says so.

Benchmarks

A reproducible suite under benchmarks/ covers the divergence boundary (600 runs), amplicon prediction against 17 published assays from 8 sources, a paired comparison against TEsorter on both database modes, seal stability and coverage, and a sixteen-tool capability matrix in which every cell is sourced. Every figure and table regenerates from raw output with one command; no value is typed by hand.

633 tests across 39 files: 632 pass, one skipped, none fail.

Known issue

The documentation screenshots carry a v3.5.0 header badge rather than 3.6.0. A capture run against this build is still owed.

TEagle v3.4.0

This release is mostly about answers TEagle no longer gives. Two rounds of review went looking for places where a result claimed more than the evidence behind it, and every one of them now withholds the claim, qualifies it, or shows the limit that was previously invisible.

Results may differ from 3.3.0 for three classes of element. In every case the earlier answer was the confident one and the new answer is the honest one.

Calls that claimed more than their evidence

A Copia/Gypsy call is no longer made from an unreadable arrangement. Copia and Gypsy are told apart by whether the integrase sits before or after the reverse transcriptase in translation order. That order is recovered by negating the minus-strand coordinate — correct within one strand, meaningless across two. Measured: with the integrase and RT on opposite strands, moving the integrase from far upstream to far downstream of the RT produced the identical verdict. The comparison had collapsed to "which of the two is on the minus strand." Such an element is now reported as superfamily undetermined — which is what the manual already said the tool did.

A protein-domain scan that crashed no longer reports "not detected". The empty result was passed to the classifier, which then listed every unfound domain as absent. An environment failure was indistinguishable from a finding about the sequence.

The search budget is disclosed. Only the twelve longest open reading frames are searched. On a record with more, the shorter ones were never looked at, yet still fed the visible "not detected" list. The record now says how many were left out.

A Helitron is no longer filed as a retroelement. A helicase hit fell through to a generic branch and emerged classed as Class I retro, labelled "transposase not detected" on a record where the diagnostic domain was found. The class is left unassigned. No Helitron superfamily call is made — that standing decision is unchanged.

A terminal repeat rejected on identity leaves a trace. The detector accepts a pair at ≥80% identity and previously discarded anything below it silently. LTR copies are identical at insertion and diverge with time, so that floor is an age limit: the detector is blind to older elements by construction, and reported that blindness as absence. A pair failing only the identity test is now listed as an advisory candidate carrying its measured identity — and cannot be mistaken for a confirmed one. Below roughly 72% identity the k-mer seeding stops finding the pair at all, which no threshold change recovers; that is stated too.

The confidence tier follows the evidence ledger. A terminal inverted repeat that does not enclose the transposase was already refused credit by the completeness ledger, yet still lifted the badge out of Candidate. Two records with identical ledgers no longer get different badges.

In-silico PCR says when it stopped early, instead of letting a truncated search on a repetitive template read as a clean specificity result.

New

A coding-structure line beside the completeness tier. The tier is a domain-ledger verdict: it reports partial whenever an expected model was not detected. Read alone that says decayed — which is why other tools call the same element "intact". The card now states separately whether the domains that were detected sit in one uninterrupted reading frame, so "a module we did not detect" is never confused with "a broken copy". The tier value itself is unchanged.

Genome-anchored GFF3 and BED export. When the specimen came from a coordinate fetch, the annotation can be written against the chromosome with every feature offset into genome space, so bedtools intersect and a genome browser find it. A minus-strand fetch is refused rather than shipped subtly wrong — remapping those needs a flip, not an offset.

The thresholds that decide the result are sealed. The analysis manifest sealed one parameter while every decisive number — HMMER E-value, the LTR and TIR identity floors, the PPT purine fraction, the TSD window, the ORF budget — lived only as a function default. Twenty are now sealed, read from the functions that apply them, so a changed default changes the seal. The manifest hash for identical input therefore differs from 3.3.0.

Exports carry their caveats. The scope notes shown beneath a table travel into the file, and coordinate columns are exported under headers that spell out the convention.

Fixed

• A finished whole-genome annotation is shown. It previously completed with no banner and no results window, while discarding an annotation announced a fabricated completion.
• Deleting a cached genome is refused while an annotation is reading it, and the refusal is visible rather than silent.
• Every export is written atomically, so an interrupted write cannot leave a complete-looking half-file over a good one. A failed export explains what to do.
• File-open and backend failures report in plain language instead of an errno with a full filesystem path.
• The ViennaRNA cross-check distinguishes "ran, found no structure" from "engine not installed".
• Degenerate IUPAC bases are reported — a sequence a third R/Y/W read "valid" beside "N content 0.0%".
• The TSD method text states the 2–12 bp window it actually searches.
• Numeric table columns align so digits line up; scope prose moved out of the data typeface; cards ease open and closed, honouring the Windows "show animations" setting.
• At high UI scale in a narrow window the figure toolbar was clipped and unreachable.

Documentation

Thirty claims across the user manual and the technical report were audited against the code and corrected. All 44 figures were re-captured from this build through the pinned harness. Both PDFs were rebuilt and inspected page by page — 50 of 50 pages, no defects found.

Note on the installer

The installer is not code-signed, so Windows SmartScreen will show "Windows protected your PC". Choose More info → Run anyway. On a managed laptop the prompt may be blocked outright, in which case your IT team has to approve the executable.

TEagle v3.3.0

Two scientific additions and a rework of the optional Linux (WSL) backend installer.

Added

• LTR cis-elements. The polyadenylation-signal motif, counted only when a GU/U-rich downstream element follows it, plus the seven verified non-canonical terminal motifs. Both are advisory and reported with their limits — a motif, never a located cleavage site.
• Whole-genome transposable-element annotation. RepeatMasker over a genome you have already downloaded, processed in contig chunks so progress is visible, resumable, with the cost stated before the run starts. Transposable elements are counted separately from tandem and non-TE repeats, because merging them overstates TE content.
• Flanking sequence from any feature — copy it, export FASTA, or design primers on it. Amplifying across an insertion from its flanks is how an insertion is genotyped.
• A BACKEND button in the window header opens the backend installer from anywhere.
• A Library choice on panel 03. RepeatMasker searches curated Dfam families unless it is asked for both, so the optional uncurated partitions were unreachable from single-sequence naming — you could download 3.9 GB and get the identical blank result. The setting is sealed into the run's provenance. Measured on yeast Ty1 (M18706): curated-only finds nothing; including uncurated finds an LTR/Copia family.

Fixed

• The poly(A) downstream-element window was measured from the wrong point (+10–30 instead of +20–60), so it judged the wrong stretch of sequence. Results may differ from 3.2.1 for LTR elements.
• A false claim about curated Dfam coverage. The app said the curated library holds 9 families for Drosophila melanogaster and lacks copia, gypsy, hobo and mdg1. Measured against famdb: 399 models, and it does contain Copia_I, Copia_LTR, Gypsy_I, Gypsy_LTR, hobo and MDG1_I/LTR. The 9 was the yeast figure attached to the wrong organism.
• A Dfam download showed no progress — curl draws its meter with carriage returns and no newline, so a 40–60 minute transfer was one unreadable line. Progress is now reported as whole lines every 15 seconds.
• An interrupted multi-gigabyte download was deleted by the checksum gate, despite the panel promising it resumes.
• The free space a partition needs was understated by roughly half — the uncurated eukaryote partition unpacks to 22.6 GiB, not the ~14 GB claimed.
• micromamba could refuse to create its environment permanently if a Dfam partition was installed before the environment existed.
• The methods panel claimed a 21-model Pfam panel after it had grown to 30.

Performance

Startup import graph 537 ms → 151 ms.

Verification

580 tests. README, manual and report audited claim by claim against the build; all 44 figures re-captured; both PDFs rebuilt and inspected page by page. Eight specimens across six organisms checked against published architecture — 6 of 8 match, both mismatches documented.

---

TEagle-Setup-3.3.0.exe — SHA256 112e0b9226089fe648a84d610b7ba08d1a7b4b607d9e9fcdff4491f52796c747
Per-user install, no administrator rights. The optional Linux backend installs from inside the app.

TEagle v3.2.1

A fix release. No new feature — one scientific-correctness fix and a set of robustness and documentation corrections surfaced by a full review of 3.2.0.

Classification — a reported target-site duplication can change

Structural detection runs before the superfamily is known, so it selected the longest exact repeat flanking the element. A coincidental longer repeat in an AT-rich flank could outrank the diagnostic short duplication — a Tc1/mariner element's 2 bp TA in particular — which then read as incongruent (discrediting genuinely complete termini) and exported the wrong duplication coordinates.

Once the superfamily is resolved, the duplication is now re-detected with its literature target-site length preferred. The correction is monotonic: it can only shorten a coincidental repeat to the diagnostic length when that length genuinely flanks the element, never lengthen or fabricate one. Effect on real specimens: none for a bare pasted element (no flanks, no duplication to reassess); a change is possible only for a flanked element of a superfamily with a literature target-site length whose insertion happens to sit in a coincidental longer exact repeat.

Reliability and connectivity

• Adding a custom organism no longer races on its store. Two concurrent resolves could each read the assembly store and the second write clobber the first — an organism the app reported as added would silently vanish. The read-modify-write is now serialised.
• The "Add" control is disabled for the duration of its network resolve. Only the text field was disabled before, so a double-click or Enter-then-click queued duplicate NCBI lookups; a single-flight guard now survives a manager close/reopen.
• The "Resolving… against NCBI" status shows as a Notice, not an Error — it had inherited the default error styling (an Error-titled, focus-stealing dialog) for a routine in-progress message.

Documentation

• Report: corrected the mariner Mos1 benchmark accession (M14653 → X78906, the record actually benchmarked) and removed an unsubstantiated Tam3 specimen.
• app/README.md: provenance cross-reference pointed at the wrong panel (05 → 07); the superfamily list predated the 3.2.0 panel growth (now includes ERV, DIRS, CACTA, MULE and IS4-like/piggyBac); XLSX was missing from the table-export formats.
• The LTR congruent-TSD evidence sentence was rewritten to state the observed length; the DNA "no attributable expected length" case now reads neutrally instead of implying congruence.
• Every figure recaptured at 3.2.1 — the app renders its version in the header badge and provenance card, so figures go stale on any bump.

---

Installer SHA256 B5731FE1D7036B9BFC9979205EA0FEC2E4546AFDC3EFD9F27CAE1405D779393A. Frozen-bundle self-test passes (primer3 2.3.0 · ViennaRNA optional, not bundled · pyhmmer 0.12.1 · HMM 57f2b7881f35, 30 profiles).

TEagle v3.2.0

TEagle is now free software under the GNU Affero General Public License, version 3 or later.

Added

• GFF3 / BED annotation export — verified Sequence Ontology terms, gated on the completeness tier so an evidence-incomplete call degrades to a generic term rather than asserting an unsupported subclass.
• Self-similarity dot plot in its own resizable window — zoom/pan, selectable direct/inverted mark colours (colour-vision-safe default), export to SVG, PNG, and vector PDF.
• Five bundled real published TE examples for a first-run that isn't a synthetic null.
• Optional in-app WSL ViennaRNA primer-QC engine (same ΔG as the in-process route).
• Pfam panel grows 21 → 30 models — LINE ORF1p/ORF2p-endonuclease, a tyrosine recombinase (new DIRS-group class), the Helitron helicase, and the CACTA / MULE / IS4 transposases.

Changed

• Relicensed proprietary → AGPL-3.0-or-later.
• A locus with reverse transcriptase + a tyrosine recombinase and no DDE integrase is now classified as DIRS-group (3.1.0 read it as a LINE).
• Terminal-repeat (LTR/TIR) lengths measured by X-drop diagonal extension across the whole sequence.

Results may differ from 3.1.0 for DIRS-group calls and terminal-repeat lengths — re-run any record whose classification or terminal-repeat length you have on file.

Assets

• TEagle-Setup-3.2.0.exe — per-user installer (~40 MiB), no admin required.
  SHA-256 a1c11d85f36d6d55cfd6e6e43d57162cf8f7161d785ce3206f9fc7390df33b5d
• TEagle-User-Manual.pdf — the complete illustrated user guide.

Full history in CHANGELOG.md.

TEagle v3.1.0

Classification — displayed completeness tiers can change

The Class II (DNA transposon) branch of the completeness assessment now derives its tier from a real present/missing ledger instead of asserting intact / autonomous-consistent whenever a transposase and a TIR pair merely co-occur. A TIR pair now counts only if its arms bracket the transposase — an inverted repeat elsewhere in the record is not that element's termini (Wicker 2007).

Effect on real specimens: none. Tc1, Activator Ac, hobo, mariner Mos1 and Tam3 all keep intact / autonomous-consistent, and the 22-specimen retroelement panel is untouched. What changes is the case the previous logic got wrong — a non-enclosing TIR pair now reads partial (transposase present), as do 5′/3′-truncated copies and internal fragments.

The tier vocabulary is unchanged, and every existing hedge — including the "scoped to the tested Pfam domain panel" caveat — is preserved verbatim. classify() only receives structural and domain evidence, so it still cannot see record-level truncation; the new evidence line says so in hedged terms rather than implying otherwise.

Figures

• GAG #7A7A7A → #6438FC. The old grey sat only 5.63 CAM02-UCS from the #888 unknown-domain fallback — and since both are achromatic, colour-vision-deficiency simulation moved that separation by 0.00. A named GAG domain looked unnamed. The new hue was chosen for the best worst-case CVD separation (18.41) while clearing 3:1 contrast against both the dark and light track backgrounds.

Documentation

• Every figure recaptured at 3.1.0 — the app renders its version in the header badge and provenance card, so figures go stale on any bump.
• Two manual figures produced by neither capture harness are now regenerated, and the orphan sweep covers report/manual_figures/ (it previously audited only docs/img/).
• README installer size corrected: 40 MiB is the download; 148 MiB was the unpacked bundle.

---

Installer SHA256 FE66DFA25EA9901F04D2B2228580695E91711BE842E89604DFF9BE8CFCB463AD. Frozen-bundle selftest passes (primer3 2.3.0 · ViennaRNA 2.7.2 · pyhmmer 0.12.1 · HMM 197579b7a891, 21 profiles).

TEagle v3.0.0

81% smaller installer — 203 MiB → 40.4 MiB (bundle 776.8 → 147.6 MiB, TEagle.exe 71.2 → 4.4 MB). A single collect_submodules("openpyxl") call pulled a pandas-interop shim, dragging in PyTorch, scipy, transformers, sklearn, pandas, PIL, lxml and numpy that the code never imports. The build now fails if an unlisted top-level package enters the bundle.

Scientific presentation

• ENV protein domains render in their own measured hue instead of the unknown-domain grey.
• The gel distinguishes on/off-target by shape as well as colour (colour-blind safe).
• Feature labels too narrow to draw in place are surfaced via leader lines — the PBS tRNA-identity hedge is no longer silently dropped.
• The poly-A tail is no longer drawn as a terminal repeat.
• Self-priming products are their own call, so derived off-target counts can no longer go negative.
• The whole-genome scan table states isPcr's real 1-based inclusive convention.
• A 3′-truncated record no longer also reads a bare "intact" completeness tier.

Honesty

• Deleting a cached genome now asks for confirmation.
• The real disk requirement (≥8 GB free, ~3.8 GB unpack peak) is disclosed before download instead of surfacing as a post-hoc failure.
• Backend failure is announced in words, not colour alone.
• Both export entry points propose the same filename.

Accessibility

• Keyboard focus is visible on every control, tables show a current-cell cue, citation links are keyboard-reachable, and accessible names were added.

Design system

• One source of colour truth, integer-only type scale, spacing snapped to a real scale, and distinct loading / error / empty states.

---

Verification — installer SHA256 5317352B43ECF9CB260C713E3DBE790E55BDE3DCC35678523FBE139E62E4663B; the frozen-bundle selftest passes (primer3 2.3.0 · ViennaRNA 2.7.2 · pyhmmer 0.12.1 · HMM 197579b7a891, 21 profiles). Note that the selftest gate in the build script never actually executed before this release — & on a GUI-subsystem executable does not block, so its exit code was never read. It is now genuinely enforced.

TEagle v2.12.0 — dialogs, phase navigation, fixed-size genome manager

Post-overhaul refinements from direct use: the tool now carries you to the next step, warnings are proper dialogs you can dismiss, the genome manager shows everything at a fixed comfortable size, and download failures read plainly instead of as a crash. Driven by a user-directed fix loop with per-round triple-reviewer verification.

Added

• The tool moves you to the next phase. "Design primer here" now brings the primer panel up with a busy cue and lands you on it (no more staring at the previous region's stale table); "send to in-silico PCR" and "send to splice" scroll their panel into view.
• Progressive disclosure in the classification banner. The full domain-panel methodology folds behind a Scope and methods toggle; the one-line completeness caveat stays visible, so the card is less crowded without hiding the honest limitation.

Changed

• Warnings and confirmations are dismissable dialogs, centred over the window, instead of a banner wedged above the panels. Errors and warnings persist until closed; confirmations auto-dismiss without stealing focus; the message text is selectable so an accession or seal can be copied.
• Manage genomes is a fixed-size panel sized to show every organism, column and Download button — wide enough that names never truncate, at a comfortable row height that keeps every button fully legible.
• The backend installer opens larger so each component row is comfortably readable.
• Genome downloads retry with exponential backoff (five attempts, 5→10→20→40 s) so a transient NCBI rate-limit or dropped transfer recovers instead of exhausting a short fixed retry.

Fixed

• A genome download or scan that fails now shows one clear warning with the real reason, not a red "unexpected error" crash dialog stacked on top of the status line.
• The light-theme dialog and confidence-chip text is darkened to meet WCAG AA (≥ 4.5:1) against its tinted background.
• The Manage genomes Download/Delete button no longer renders clipped or squeezed — the label is centred at a comfortable height in every row, at every UI scale.
• The genome-manager dialog no longer collapses to a sliver when it rebuilds while open (rescale, add-organism, or a download completing).

Documentation

The complete illustrated TEagle-User-Manual.pdf is attached, refreshed for these changes (the classification banner's Scope-and-methods disclosure, the fixed-size genome manager, the phase-navigation, and the dialog-based messages). Every screenshot across the manual and README is freshly captured at v2.12.0.

Install

Download TEagle-Setup-2.12.0.exe and run it (per-user, no admin). The scientific core is bundled; the optional Dfam/splice/genome-scan backend installs from within the app.

TEagle-Setup-2.12.0.exe SHA-256:
3636ebeae47df3c57fe6c12134f68a8faa65912ae22bceeb7a880fa84f3f27df