Product updates

What’s changed.

Published improvements to DeltaFold, with what changed, why it matters, and how it was checked.

36 published releasesLatest: 21 November 2026

Newest first.
Publication dates use Pacific Time.

Download release notes
Usability

See what a protein score can tell you

The public example explains how to choose a generated protein sequence to inspect and which question still needs a lab test. The Workbench carries that distinction through saved results and recorded experiments.

What changed

  • Design explains how letter patterns affect a score, what Use changes, and where to find the original batch. If a response is interrupted, the example tells you how to check for an existing run.
  • Experiments explains which untested versions receive a rank. Campaigns distinguishes planned samples from recorded outcomes. Disease Research identifies which evidence applies to a target, and Molecular Dynamics explains what the archive says about each displayed frame.
  • Updated labels, guides and recovery messages identify the current controls, the source of a result and what survives reopening.

Why it matters

Researchers can explain why they chose a version for closer inspection, find that choice again and tell which questions still need a measurement.

Checks & evidence

Software verification

  • Scripted signed-out desktop and phone-width visits generated four candidates, compared letters, used one candidate and reopened the saved run. An initial phone visit returned a server error; a later visit completed the task.
  • Disposable local accounts exercised ranking, saved rounds, guest copying and recovery with synthetic data. The updated Help instructions and public examples were checked on the deployed site.

Scientific evidence

These checks verify software behavior and wording. They provide no new laboratory measurements or evidence of drug activity. Independent visitor comprehension and next-day recall have not been measured.

Reliability

Keep tiny readings from becoming zero during project recovery

Restoring an Experiment project now stops if a nonzero number in its input data would be saved as zero. The message asks you to check the original value and units.

What changed

  • Restore Project Data preserves the numbers as written in the backup until DeltaFold checks them. The check applies to older project exports and newer exports with a saved project snapshot.
  • A rejected restore leaves the file selected and creates no project. A genuine measured zero and small values the software can represent still reopen correctly.

Why it matters

Researchers can recover a project for the next round of testing without a tiny reading being recorded as a measured zero. For a zero already present in an older record, consult the original lab data.

Checks & evidence

Software verification

  • Using disposable local accounts, the previous release accepted all 16 invalid API examples as zero. The update rejected all 16 before creating a project, and all 18 valid control readings reopened exactly.
  • Desktop and phone-width browser checks exercised file selection, rejection, correction, reopening and account isolation. The deployed client, signed-out account gate and recovery Help were verified on the public site.

Scientific evidence

These checks used synthetic data to verify recovery behavior. They add no evidence about assay accuracy, protein activity or predictive performance.

Usability

Updated welcome text and a simpler protein heading

The welcome page has a revised introduction. Each rotating protein now appears under its name, with its structure details in the caption.

What changed

  • The opening paragraph now begins “Put simply, DeltaFold helps you choose what drug to make.”
  • Removed the “Experimental reference” label next to the protein name.

Why it matters

Visitors can read the protein name at a glance and find its source in the caption below the animation.

Checks & evidence

Software verification

  • The homepage source differed only in the introduction and the removed label. Live checks at desktop and phone widths confirmed the new paragraph and the protein captions. The label stayed absent as the animation moved from Ubiquitin to SUMO-1.

Scientific evidence

These presentation edits add no new scientific or clinical evidence.

Usability

New wording in About DeltaFold

The About DeltaFold section on the welcome page has a new tagline.

What changed

  • The line reads “One letter can make all the difference.” The word “all” appears in bold.

Why it matters

Visitors see the revised message above the About section’s explanation of the research workflow.

Checks & evidence

Software verification

  • Compared the welcome page with the preceding release and verified the exact sentence and bold emphasis on desktop and phone layouts.

Scientific evidence

This change is editorial and adds no scientific evidence.

Usability

A new introduction on the welcome page

The welcome paragraph now addresses visitors directly and introduces the sequence comparison example.

What changed

  • The opening paragraph uses the revised wording and emphasizes the word “you.”

Why it matters

Visitors see the new introduction above the link to try the Design example.

Checks & evidence

Software verification

  • The published paragraph matched the requested text on desktop and phone layouts, with “you” in bold. The page source comparison found a single paragraph edit.

Scientific evidence

This copy edit adds no scientific or clinical validation.

Usability

See why you might use each Workbench tool

Short explanations under Workbench headings connect each tool to a choice you can make, from spotting a gap in a structure to choosing which protein versions to test.

What changed

  • Sequence, Evolution, Structure, Molecular Dynamics, Disease Research and Model now give a specific reason to use the result on screen.
  • Steps inside Experiments and Campaigns explain how to match lab readings to samples, choose untested versions, count space for repeats and compare later measurements with an earlier decision.

Why it matters

A newcomer can see when a sequence pattern suggests an idea, when a score can guide a choice, and when a lab measurement can answer the question.

Checks & evidence

Software verification

  • Browser checks covered 25 headings and workflow locations at desktop and phone widths. All 12 revised descriptions matched the reviewed text and fit the available width.
  • The public Workbench headings, the Structure handoff, the Dynamics viewer and Model Help were checked on the live site. Private project and campaign steps were opened with disposable local accounts.

Scientific evidence

This wording update adds no new laboratory evidence or predictive validation.

Usability

Find the exact letter to fix in a sequence

Design now identifies invalid starting letters by position. Model points to the affected training record and the positions to correct. The public Design example also gives you a way back if its files fail to load.

What changed

  • Read the correction beside Starting letters in Design. In Model, use the record number and positions to find the letters in a rejected sequence. Positions start at one and skip spaces.
  • If the homepage example stops loading, choose reload the example to restore the four-candidate setup. Choose Generate when you are ready to create a saved run.

Why it matters

Researchers can correct the original entry in place. A guest can recover the public example after an interrupted load and decide when to generate it.

Checks & evidence

Software verification

  • Desktop and phone-width checks dismissed the input error, navigated away and back, and read the same invalid positions. Correcting the training text left training under the visitor’s control.
  • In disposable browser sessions, controlled failures of the application and example files each left a working reload link. Reload restored the example; one explicit Generate action created one saved run.

Scientific evidence

The checks used synthetic inputs and tested software behavior. No laboratory or predictive validation was performed.

Usability

See what was measured and recover from an input error

Experimental shortlists now put each prediction beside its measurement status on phones. Workbench results identify the run or structure they came from, and Help can take you back to a saved sequence.

What changed

  • Sequence shows which letters need correction. Design identifies the previous run if a new attempt fails. Evolution labels the download that contains the best sequence across all generations.
  • The GammaFold example opens with the outcomes of its six fictional samples. Saved molecular plans show which revisions contain measurements before you open them.
  • Structure points guests to public ubiquitin. Dynamics labels the starting structure separately from the current GROMACS snapshot and the UNRES schematic.
  • Model clears its empty-input message when you fill the example. Account recovery offers a way back to the guest workflow, and copying guest work lists the runs and model copied.

Why it matters

Researchers can see whether a candidate has a lab measurement, identify earlier work on return, and correct an input while keeping their previous result available.

Checks & evidence

Software verification

  • Signed-out desktop and phone-width checks exercised invalid input, correction, saved-run reopening, Help, and all eight archived frames. Fresh visits completed the demo after intermittent loading failures.
  • Disposable local accounts exercised the phone shortlist, same-name molecular revisions and copy receipts after reload. The deployed interface files matched the tested versions.

Scientific evidence

These software checks used public structures and synthetic examples. The release adds no lab measurements, estimates of simulation time, or new evidence about predictive accuracy.

Usability

Try four sequences and find your selected result again

The homepage example opens with four candidates ready to generate. Sequence identifies the protein in the editor, and Library brings the reference you inspect into view.

What changed

  • Choose Generate 4 candidates once the example is ready. Open Edit settings to use the regular Design form.
  • After using a version from Evolution, reopen its saved run and generation from Sequence. Guests can copy their letters before signing in to save to Library.
  • Inspect a Library search hit to see its accession and sequence choice. Return to the same hit with your filters and place preserved.

Why it matters

A visitor can compare short sequences and a returning researcher can find the version they chose without repeating the calculation.

Checks & evidence

Software verification

  • Signed-out desktop and phone-width checks generated four candidates, compared them, used one and found the saved run. A delayed startup check exercised early and repeated input.
  • Reopening the same browser recovered the recorded Evolution version. Keyboard, mouse and touch checks opened the selected Library reference and returned to the originating search hit.

Scientific evidence

These checks used public and synthetic examples. The release adds no laboratory measurement or predictive validation.

Research workflow

See which letters changed in each generation

Choose a round in Evolution to see how its top sequence compares with the protein you started with.

What changed

  • Select a starting letter to see its position and amino acid name. After a run, choose a chart point or generation to read its highest and average scores. Underlined letters show where the two sequences differ.
  • Inspect one sample in Campaigns to see why it was chosen and which outcome was recorded. In Runs, expand a calculation to read its saved inputs and results.

Why it matters

You can see that two different sequences earned the same score and decide which version to inspect next.

Checks & evidence

Software verification

  • Signed-out desktop and phone-width walkthroughs ran the synthetic example, selected a generation, read a changed letter and reopened the same saved run in under two minutes.
  • Keyboard and touch checks covered the comparison. A disposable test account exercised measured, missing, failed and censored campaign outcomes using the published interface.

Scientific evidence

The demonstration uses synthetic sequences and existing scoring rules. The interface checks add no experimental evidence about how the proteins behave.

Usability

Compare two protein sequences letter by letter

Open the Design example from the homepage to see how researchers choose a protein sequence for a closer look.

What changed

  • Generate four short examples. The starting letters stay aligned as you switch between two candidates, and the first difference is spelled out below them.
  • Read how each sequence earned its score, take a candidate into Sequence with the Use button, or reopen the batch from Runs.
  • Across the Workbench, the guides explain what each tool gives you, from a position in a sequence to a recorded simulation frame or a measurement from the lab. The compact phone menu and storage details are available as the page opens.

Why it matters

You can inspect several candidates before choosing which one to work with, and revisit that choice later.

Checks & evidence

Software verification

  • Signed-out, scripted walkthroughs on the live site at desktop and phone widths generated the example, identified a letter difference and opened the selected sequence in under two minutes.
  • Browser checks covered selecting by keyboard, reviewing saved results, guest workspace isolation and reduced motion.

Scientific evidence

These checks used synthetic sequences and existing scoring rules. They provide no new experimental evidence about protein behavior.

Usability

Watch Ubiquitin come into view

The homepage gives you a moment to find the protein panel before bright blue and cyan pieces come together to form Ubiquitin.

What changed

  • The frame and caption appear first. Once the panel is in view, the molecule area stays empty for about two seconds. Ubiquitin then takes about two seconds to come together.
  • You get the usual 7.5 seconds to see the whole protein before the move to SUMO-1. The opening plays once per page load.
  • Visitors who use reduced motion or whose browsers cannot play the animation see the complete bright structure.

Why it matters

You can follow the first entrance even if you have to scroll down to reach the panel.

Checks & evidence

Software verification

  • Recordings of fresh and returning visits at desktop and phone widths showed the empty panel, the full reveal and the switch to SUMO-1. A separate recording started with the panel below the screen.
  • Software checks covered the static fallback, playback controls and the timing of later transitions.

Scientific evidence

The panel illustrates the existing experimental Ubiquitin structure, PDB 1UBQ. These checks measured screen timing and rendering.

Reliability

Keep every record when you choose zero validation

The local dataset builder now keeps every accepted sequence in training when you request a validation fraction of zero. Previously, it still set one aside if you had two or more.

What changed

  • Both validation files are empty when you request zero, even when you rebuild in a folder that held an earlier split.
  • New manifests record the fraction you requested alongside the number of training and validation records actually written. Model Help explains where to find those counts.

Why it matters

You can tell exactly how many records are available for training and whether you have a separate set for validation.

Checks & evidence

Software verification

  • The published command passed all eight zero-fraction checks. Across 24 positive-fraction checks, all four split files matched the previous release byte for byte.
  • On the live site, a signed-out visitor could read the Model guide at phone width and return to an unsaved draft using the keyboard.

Scientific evidence

These checks measured how records were assigned to files. No model accuracy evaluation or laboratory experiment was performed.

Usability

A brighter welcome and clearer Workbench tabs

Ubiquitin opens at full detail on the homepage. Each Workbench tab now explains what you can do there, and Molecular Dynamics shows how to read its drawings and recorded frames.

What changed

  • The first Ubiquitin image matches the interactive view. The bright opening works on desktop and phone, with a complete static image for reduced motion.
  • Each Workbench heading has a short introduction. Structure and Runs keep their source and storage information on separate lines.
  • In Molecular Dynamics, highlight residues and switch to the UNRES schematic to see how those same positions are represented. Frame controls explain simulated time and computer time. Source run details distinguish graphics processors, independent runs and replicas.

Why it matters

You can recognize the opening protein immediately, choose the right tab for your task, and tell which details in a simulation file are known.

Checks & evidence

Software verification

  • Live desktop and phone-width browsers checked the first frame through the first protein change, returning visits, reduced motion, and keyboard playback. A slow phone-width load displayed the complete opening protein.
  • All eleven Workbench descriptions and Molecular Dynamics were checked on the live site. Both example structures reopened byte for byte, the selected residues stayed in view, and no external requests occurred during inspection.

Scientific evidence

The schematic illustrates the published UNRES representation. The eight GROMACS frames come from an existing archive. This release adds no simulation or experimental measurement.

Research workflow

Prepare a structure for dynamics

Open Molecular Dynamics from Structure to inspect existing coordinates, save a preparation package, or explore an archived trajectory. The examples work without signing in.

What changed

  • Choose experimental ubiquitin, predicted SUMO1, or a saved fold from your workspace. Select a chain or residue range to see which positions have coordinates.
  • Export the original PDB with its source and checksum, then reopen the package. Preparation checks identify missing or conflicting positions.
  • Compare source coordinates with a labeled UNRES schematic or step through eight archived GROMACS ubiquitin frames.

Why it matters

You can see exactly which coordinates you would hand off and reopen the same source later. Inspection takes place in your browser.

Checks & evidence

Software verification

  • 279 automated tests passed. A staged account-isolation check confirmed that another guest and a signed-in account could not fetch a guest?s saved structure.
  • Signed-out desktop and phone-width browsers on the live site reopened both example packages byte for byte, stepped through all eight frames, and identified a missing residue. Neither browser made an external request during inspection.

Scientific evidence

The viewer displays previously published coordinates and simulation frames. This release adds no new simulation, predictive evaluation or laboratory measurement.

Reliability

Model training shows which record to fix

If a submitted sequence is invalid, Model identifies the record and leaves your active model in place.

What changed

  • Model and the training command now stop on the first invalid record in a FASTA, JSONL or plain-text submission. Valid duplicate sequences still count once.
  • The dataset command reports why it skipped a file when the selected coordinate chain contains an unsupported residue, has conflicting identities at one position or exceeds the length limit. A FASTA file with an invalid entry is skipped in full.
  • The Model Help topic and tutorial transcript explain how to correct an entry and submit again.

Why it matters

You can correct the source before training and see why the dataset command left a file out.

Checks & evidence

Software verification

  • 258 software tests passed, including checks for invalid records, duplicate handling, saved-model preservation, account boundaries and Help.
  • On the live site, a synthetic submission identified the invalid second record. The active model and saved run list were unchanged after rejection.

Scientific evidence

This release verifies software behavior on synthetic inputs. It adds no measurements of predictive accuracy or laboratory performance.

Usability

Smoother protein and Workbench motion

Desktop protein handoffs retain the phone presentation, and Workbench navigation moves at a more deliberate pace.

What changed

  • Keep the fine blue/cyan transition texture consistent on larger protein views, with unchanged reference geometry, labels and 7.5-second settled viewing periods.
  • Use smoother directional tab slides and an opaque mobile-menu reveal and collapse. Rapid selections take effect immediately, with no animation queue.
  • Keep every menu destination, including What’s new, on one document scroll surface. Preserve reduced-motion behavior, Help and working inputs.

Why it matters

Follow changes between structures and research tasks with clearer visual continuity while retaining the current workspace.

Checks & evidence

Software verification

  • 279 release-source checks passed. Fixed-pose browser comparisons matched 84 settled renders and 35 phone transition frames exactly; protected decorative animation assets stayed unchanged.
  • Live Chromium checks covered complete seven-reference cycles, directional navigation, rapid selection, menu reversal, Help and input retention. Responsive viewport, enlarged-text, reduced-motion, loading-failure and lifecycle checks passed.

Scientific evidence

Presentation and rendering changes only. The seven reference coordinate assets and scientific calculations are unchanged. This release adds no predictive or biological validation.

Usability

A complete narrated Help walkthrough

Watch the research workflow from inputs to saved decisions, or jump to the chapter you need.

What changed

  • Add a 24-minute 1080p tutorial with 19 workflow chapters, English captions, playback speed controls and a searchable timestamped transcript.
  • Connect the chapters to written Help, including the public recorded GammaFold demonstration. Keep the guided walkthrough and its progress available.
  • Load video and captions only after Play or a chapter or timestamp selection. Offer native fullscreen where supported and a separate Expanded view that retains playback position and captions.

Why it matters

Learn at your own pace, find a specific task quickly and return to written instructions without running an analysis or changing research.

Checks & evidence

Software verification

  • 143 player, Help, navigation and route checks passed against the release source. Full video and caption review was completed with no corrections.
  • Live Chromium checks verified playback, all 19 chapter seeks, captions, transcript search, speed controls, native fullscreen, Expanded view and written Help. Responsive viewport checks and unsaved-input preservation passed; the published movie matched the reviewed file byte-for-byte.

Scientific evidence

Instructional software release. Public and synthetic examples distinguish measurements, predictions, heuristics, geometric checks, reference annotations and interpretations. This release adds no predictive or biological validation.

Usability

A smoother mobile menu without nested scrolling

The Workbench menu uses more deliberate motion and keeps its destinations on one page scroll surface.

What changed

  • Remove the menu's internal scrollbar and height cap. On short screens, ordinary page scrolling reaches every destination, including What's new.
  • Use slower, full-brightness opening and closing motion while keeping rapid toggles responsive. Retain immediate behavior for reduced motion or unavailable animation support.
  • Settle the menu before tab or browser-history navigation so a deeply scrolled page does not leave a temporarily displaced navigation bar.

Why it matters

Reach menu destinations without managing a second scrollbar, with clearer motion and stable navigation between research tasks.

Checks & evidence

Software verification

  • 106 menu, workbench, Help and route checks passed against the release source. Browser checks covered narrow and landscape layouts, deep scrolling, rapid toggles, focus, Back navigation, input preservation and Help restoration.
  • The deployed menu and Help were checked in Chromium at phone viewport widths. Separate local browser checks verified reduced motion and unavailable animation support.

Scientific evidence

Interface refinement only. Scientific calculations, models and research interpretation are unchanged; this release adds no scientific or laboratory validation.

Reliability

Clearer handling of unverified structure confidence

Structure now withholds confidence interpretation when returned B-factor values have an ambiguous scale.

What changed

  • Show Scale unverified when all valid available displayed C-alpha B-factors lie between 0 and 1. Withhold the pLDDT mean, confidence categories, chart and confidence colors instead of assuming a scale.
  • Keep raw B-factor values inspectable and preserve the original PDB, coordinates and stored run summaries without automatic rescaling.
  • Explain the distinction in Structure and Help. Saved run summaries retain their stored numerical values without presenting them as newly verified confidence.

Why it matters

Researchers can inspect the returned structure without mistaking an unresolved numerical scale for established model confidence.

Checks & evidence

Software verification

  • 109 web, Help and route checks and 33 folding, storage, response-admission and backend route checks passed against the staged source. Fixtures cover fractional, zero, missing, invalid and legacy values without changing raw data or historical summaries.
  • The deployed interface reopened an existing public ubiquitin result through the supported backup-restore workflow. Its downloaded PDB hash and stored summary were unchanged. Desktop and phone-width presentation, Help keyboard exit, input preservation and logout clearing were checked without issuing a new prediction.

Scientific evidence

Software interpretation safeguards only. No new confidence calibration, structure-accuracy evaluation, biological validation or therapeutic evidence. Other values retain the existing scale assumption; range alone does not establish a confidence scale.

Appearance

A balanced, smoother mobile menu

Consistent navigation tiles and brief, responsive motion make the mobile workbench menu easier to use.

What changed

  • Use centered, consistently sized navigation tiles. Unpaired Disease Research, Model and Runs tiles sit in the middle of their rows and retain the same width as neighboring tiles.
  • Open and close the compact menu with brief, full-brightness motion. Rapid toggles respond immediately without queuing animations.
  • Remove closing menus from keyboard and assistive-technology interaction immediately. Escape restores focus to the Workbench button.
  • Keep reduced motion, unavailable animation support and Help previews immediately usable, while retaining inputs, routes and desktop navigation.

Why it matters

The compact workbench feels more cohesive and responsive while keeping research context and controls intact.

Checks & evidence

Software verification

  • 91 checks passed against the exact staged menu, workbench and Help sources. Actual local and deployed browser checks covered 320 and 390 pixel widths, rapid toggles, reduced motion, animation fallback, Escape, keyboard navigation, Back, unsaved input preservation and desktop resizing. Live Help returned to the original route and input without running a task.

Scientific evidence

Presentation and navigation only. Scientific calculations, research records and model behavior are unchanged; no new scientific validation is claimed.

Usability

Smoother navigation between research tools

Short directional slides connect workbench pages while keeping content bright and navigation responsive.

What changed

  • Use a consistent horizontal slide across all eleven workbench tools, with direction following their navigation order and shorter travel on narrow screens.
  • Respond immediately to rapid selections without queuing transitions. Inactive pages leave keyboard and screen-reader interaction immediately.
  • Keep inputs, results, saved viewing positions and running work intact. Short pages retain enough space to avoid a navigation scroll jump.
  • Switch immediately with reduced motion or unavailable animation support. Help previews stay still, and account changes cancel any outgoing motion.
  • Updated Help to explain navigation, reduced motion and the existing explicit-save boundary.

Why it matters

Researchers can move between tools with clearer continuity, readable content and fewer visual interruptions while keeping their working context.

Checks & evidence

Software verification

  • 361 JavaScript checks and 34 Python checks passed; one existing optional JavaScript fixture was skipped. Tests covered navigation, Help, account boundaries and scientific regressions.
  • Desktop and phone-width recordings, intermediate frames, rapid switching, keyboard focus, input and scroll retention, delayed jobs, loading failure and saved-run reopening were checked. Seven displayed sequence values matched the previous release.
  • The deployed files matched the tested release. All eleven tools, Back and Forward, and guided Help were verified on the live site.

Scientific evidence

Presentation and software verification only. Scientific calculations, model defaults, research provenance and evidence limits are unchanged. This release adds no predictive-accuracy claim or biological or laboratory validation.

Usability

A clearer scientific workbench

Grouped navigation, clearer research context and more readable results make the existing workbench easier to use.

What changed

  • Organized the eleven workbench tools into three navigation groups, with a compact menu on narrow screens and consistent input, result and table styling.
  • Preserve each page's viewing position and usable focus when switching tools or using Back and Forward. Help returns to the starting view and inputs.
  • Keep newer Library draft edits made while a save is pending, and preserve expanded saved-protein details during refresh.
  • Clarify sequence heuristics and show saved-run assay units, recorded means and observation counts in experiment rankings. Predictions remain distinct from measurements.
  • Updated the guided Help instructions and kept ordinary route links available when scripts cannot initialize.

Why it matters

Researchers can keep their place between inputs and results, recover from interrupted loading, and inspect the meaning and provenance of existing outputs more easily.

Checks & evidence

Software verification

  • 352 JavaScript tests and 302 Python tests passed; one existing optional JavaScript fixture was skipped. Checks covered scientific calculations, navigation, account isolation, frozen history, Help and recovery.
  • A synthetic save, ranking, shortlist export and project-restore workflow passed locally, including delayed-save edits and a simulated connection failure. Seven displayed sequence values matched the previous release.
  • The deployed assets matched the tested release. Desktop and phone-width navigation, exact scroll return, Help keyboard exit, public reference preview and sign-in prerequisites were checked live.

Scientific evidence

Software and interface verification only. No model change, new predictive-accuracy evidence, retrospective biological evaluation or prospective laboratory validation. Existing scientific limitations remain; visual polish does not establish binding, function or therapeutic benefit.

Reliability

Stronger scientific source checks

New checks catch specific source mismatches and numerical errors before they become research results.

What changed

  • Require new structure responses to match the submitted sequence and contain unambiguous finite backbone coordinates. Target-evidence responses must match the selected target identifier.
  • Reject imported nonzero numeric text that would become zero, and correct very-small-error calculations and the fixed heuristic comparison on minimizing assays.
  • Hold new campaign preparation and current GammaFold review when preserved raw values or frozen predictions conflict with their source records. Historical records remain available for inspection and export.
  • Version new calculations and preserve historical checker behavior. Updated Help and Evidence Scope explain what these checks establish.

Why it matters

Researchers can identify these inconsistent inputs and distinguish recorded observations, predictions, heuristics and evidence checks while retaining their prior work.

Checks & evidence

Software verification

  • 127 integrated Python checks and 217 distinct JavaScript checks passed, covering numerical calculations, source identity, privacy, historical recovery and Help.
  • Analytical error calculations agreed with an 80-digit reference within the declared tolerance across 213 magnitude cases. Synthetic invalid-provider tests confirmed no result was saved.
  • Live assets and checker versions matched the tested release. Public Evidence Scope and signed-out Help were checked on desktop and phone-width layouts; synthetic signed-in failure and draft-preservation checks ran locally.

Scientific evidence

Analytical and synthetic software verification only. No new retrospective biological evaluation or prospective laboratory validation. The model default and existing mixed GB1 findings remain unchanged; these checks do not establish binding, function or therapeutic benefit.

Appearance

Brighter reference-protein transitions

The welcome proteins now change through a bright digital handoff and resolve into a crisp, fully lit structure.

What changed

  • Replaced the fading particle handoff with a staggered digital reconstruction and restrained cyan accents.
  • Improved protein lighting while preserving all seven exact reference structures, their order and factual captions. Each finished structure remains visible for 7.5 seconds.
  • Kept the lower decorative molecular figure, opening entrance, background, layout and branding unchanged.
  • Updated Welcome Help to describe the new effect and the existing animation controls.

Why it matters

Reference structures stay readable during handoffs, then settle into a clear view without changing the scientific content.

Checks & evidence

Software verification

  • 156 relevant JavaScript checks and 24 Python reference, Help and public-route checks passed.
  • Matched-pose renders checked all seven structures. Sixteen sampled renders of the protected molecular figure were pixel-identical to the baseline.
  • Live desktop and phone-width recordings covered all seven references and the final-to-first handoff. Deployed assets matched the tested source, and Welcome Help and Pause/Play were checked.

Scientific evidence

No new scientific or laboratory validation. The digital effect is decorative and does not represent a biological transformation.

Appearance

Welcome branding restored

The welcome page returns to Pure Performance x DeltaFold and Diagnose, Treat, Cure.

What changed

  • Restored the Pure Performance x DeltaFold headline and Diagnose, Treat, Cure slogan.
  • Removed the introductory audience label while retaining the research overview and workflow link.
  • Updated Welcome Help to match the restored wording.

Why it matters

The opening page uses the original DeltaFold branding while keeping the research tools easy to find.

Checks & evidence

Software verification

  • Existing slogan, Help and public-route checks passed.
  • Live desktop and phone-width checks confirmed the headline, slogan, matching Help and absence of horizontal overflow. Escape closed Help and returned focus.
  • Deployed homepage and Help assets matched the tested source; animation and molecular reference files were unchanged.

Scientific evidence

No new scientific or laboratory validation. This is a wording-only branding change.

Usability

A clearer path through protein research

The welcome page explains how candidates, measurements and next-batch decisions connect, with clearer instructions throughout the workbench.

What changed

  • The opening screen names protein-engineering researchers and explains the research workflow, with a direct link to its overview.
  • Explore, Measure and Decide headings replace the previous medical-outcome slogan. Clearer section summaries make the available tools easier to scan.
  • Example instructions distinguish filling inputs from running a tool, and saved-work descriptions use the current DeltaFold workspace.
  • An unchanged account-session check keeps keyboard focus on the account control. Actual account changes still replace the controls.
  • Welcome Help matches the new headings and navigation; the existing entrance and 7.5-second settled protein viewing period are preserved.
  • Help and synthetic-example messages consistently use research language while preserving the explicit actions and evidence warnings.

Why it matters

Researchers can understand the intended task and next action sooner, while keyboard users keep their place during routine session checks.

Checks & evidence

Software verification

  • 195 frontend and 43 backend checks passed, including new focus regressions, account isolation, routes, Help and animation behavior.
  • Desktop and phone-width browser checks covered the deployed welcome, Help and account focus. Isolated synthetic analysis, save/reopen and input preservation were checked before deployment.
  • Controlled reduced-motion initialization stayed static. Molecular reference assets and animation renderers are unchanged.
  • A final terminology follow-up passed 121 focused example, state-preservation and Help checks; the deployed assets and Sequence Help topic were read back afterward.

Scientific evidence

No new scientific or laboratory validation. These changes improve wording, presentation and keyboard behavior; existing scientific evidence and limits remain unchanged.

Appearance

7.5 seconds to view each protein

Each welcome-page reference stays fully formed for 7.5 seconds before its next particle transition.

What changed

  • The first viewing period starts after the entrance finishes, and subsequent proteins receive the same 7.5-second settled viewing period.
  • Smooth rotation, the existing particle handoff and motion controls are preserved.
  • Help now explains the updated protein timing.

Why it matters

The reference display advances a little sooner while retaining a clear, uninterrupted view of each structure.

Checks & evidence

Software verification

  • 115 affected animation and Help checks passed, including first-view timing, subsequent transitions and pause/resume boundaries.
  • Live asset, first-view timing, Help and motion-control readbacks confirmed the deployed change. A controlled reduced-motion preference retained a static display.

Scientific evidence

No new scientific or laboratory validation. This change affects presentation timing only; molecular coordinates and reference metadata are unchanged.

Appearance

Brighter, sharper welcome particles

The welcome entrance preserves luminous cyan particles and crisp structure detail as the protein and decorative network appear.

What changed

  • The structures reveal independently of the particles, avoiding an extra fade that dimmed the entire effect.
  • Particle transitions retain the original canvas resolution and fine detail on high-density displays.
  • The decorative network can join the opening effect when its optional animation code arrives late, without restarting the entrance or shortening the viewing period.

Why it matters

The new entrance retains the earlier rendition's fully formed colors, highlights and structural detail while keeping the existing rotation, framing and motion controls.

Checks & evidence

Software verification

  • 160 frontend and 36 backend, reference and route checks passed. Matched-pose canvas comparisons found identical fully formed pixels to the earlier rendering at two component sizes.
  • Desktop entrance, subsequent handoffs, no replay and a controlled reduced-motion preference were checked before release. The deployed welcome assets and Help instructions were checked separately.

Scientific evidence

No new scientific or laboratory validation. Molecular geometry and source metadata are unchanged; particles and the desktop network are decorative.

Appearance

A coordinated welcome-page entrance

The welcome page opens with a brief blue-and-cyan reveal that flows into its existing rotation and particle transitions.

What changed

  • The background, reference protein and decorative desktop network share a restrained entrance, followed by a subtle stagger for the headline, slogan and Open Workbench button.
  • The entrance runs once per page visit. Each first protein receives its full eight-second settled viewing period even when playback controls become ready later.
  • Interaction can settle the entrance immediately. Reduced-motion and data-saving preferences retain a stable reveal, and Help explains the behavior.

Why it matters

The opening feels cohesive while keeping text readable, controls available and reference structures intact.

Checks & evidence

Software verification

  • 155 frontend and 36 backend, reference and route checks passed. Regression tests reproduced and fixed premature first-cycle timing with delayed playback readiness.
  • Fresh desktop and phone-width visits, subsequent transitions and Help were verified on the live site. Controlled browser tests checked reduced-motion preferences, no replay and early-focus cleanup.

Scientific evidence

No new scientific or laboratory validation. Molecular coordinates and source metadata are unchanged; particles and the desktop network remain decorative.

Appearance

A more cohesive welcome animation

The rotating reference proteins and decorative desktop network share a fine blue-and-cyan particle effect and more dimensional lighting.

What changed

  • Fine luminous particles with subtle local drift replace the coarse colored blocks during protein transitions.
  • The decorative desktop network now uses the same dissolve-and-resolve effect, with eight seconds of clear viewing between transitions.
  • Restrained highlights and rim lighting bring out the reference proteins' shape while preserving their coordinates, framing and source details.
  • Help explains the shared effect and retains the existing Pause Welcome Animation and Play Welcome Animation controls.

Why it matters

A consistent visual style connects the welcome illustrations while keeping the structures and their descriptions easy to inspect.

Checks & evidence

Software verification

  • 121 frontend and 36 backend, reference and route checks passed, including reduced motion, interruption cleanup, rendering failure and Help behavior.
  • Three complete reference transitions were checked on the live site at desktop and phone viewport sizes. Desktop network cycles, Help pause/resume, focus restoration and layout checks also passed.

Scientific evidence

No new scientific or laboratory validation. Reference geometry and metadata are unchanged; the network and particles are decorative and do not represent proteins physically transforming.

Research workflow

Clearer evidence findings and explicit corrections

GammaFold explains evidence problems, links their sources and previews corrections before you accept them.

What changed

  • Evidence checks now detect duplicate observation identities, conflicting assay and frozen sample metadata, raw-value inconsistencies and changed censoring inequalities.
  • Review summaries separate failed checks from unsupported claims. Each finding explains its source, why it matters and a concrete next action.
  • Review a source fact stages an exact assertion. Correction previews show both values; explicit acceptance changes only the review draft and requires Check Evidence again.
  • Source links offer a return to the finding. Help walkthrough previews preserve the review draft, and correction values stack clearly on narrow screens.
  • Earlier checker reviews remain immutable, exportable and recoverable as history. Current approval requires a fresh check under the current checker.

Why it matters

Understand what the saved evidence supports, correct a review claim without rewriting research, and preserve the original decision history.

Checks & evidence

Software verification

  • The same targeted synthetic benchmark improved from 5 of 17 to 17 of 17 expected outcomes, detecting eleven previously missed contradictions and correcting one valid-count rejection. This is a regression corpus, not a general accuracy estimate.
  • Backend, recovery, account-boundary, frontend and Help checks passed. Live isolated Experiments and Campaigns completed incorrect-claim detection, explicit correction, rechecking, saving and export/import across desktop, tablet and narrow-phone layouts.

Scientific evidence

No new biological or clinical validation. These checks establish consistency with supplied records and declared rules; narrative screening remains heuristic. Live AI review is unavailable and no evidence is sent to an AI provider.

Research workflow

GammaFold evidence review

Check exact source records and review decisions alongside saved Experiments and Campaigns.

What changed

  • GammaFold Evidence Review checks structured facts, citations, observation statuses, provenance and configured allocation rules using authoritative saved project data.
  • Measured zero, missing, failed and censored outcomes stay distinct. Selected frozen rounds support exact comparisons between prior predictions and later individual observations.
  • Check Evidence, Save Review and Mark Reviewed are separate actions. Reviews preserve immutable source snapshots, become stale after evidence changes and never execute proposed work.
  • Private review exports and imports recheck integrity and context. Project and account recovery preserve historical reviews while resetting their approvals.
  • A separate recorded example revalidates the original synthetic GammaFold packet without an AI call. Help explains the controls, source inspection and scientific boundaries.

Why it matters

Inspect the evidence behind a result or proposed next batch, preserve the decision context and revisit it without silently rewriting history.

Checks & evidence

Software verification

  • Original GammaFold regressions and generalized backend/frontend checks passed, covering invalid evidence, account isolation, stale requests, tampering, numeric round trips and recovery.
  • On the deployed site, isolated synthetic Experiments and Campaigns completed calculations, evidence checking, saving, explicit review and export/import. Desktop, tablet and narrow-phone layouts were checked.

Scientific evidence

No new biological or clinical validation. These checks establish consistency with supplied evidence and declared rules; hashes do not authenticate measurements, and narrative screening is heuristic rather than a formal proof.

Usability

Pitch examples across the Workbench

Stage practical example inputs across the Workbench, then use the normal explicit actions to demonstrate each workflow.

What changed

  • Sequence, Design, Evolution and Model share a small synthetic sequence family. Model also offers a downloadable example FASTA; Structure can stage the same working sequence without replacing the displayed structure.
  • Molecular Design can fill plan settings after a target is selected. Library can stage a public reference search, and Disease Research can stage a clearly hypothetical target query without inventing a search result.
  • Fill Example preserves normal editing. Clear Example restores the earlier draft, with confirmation before replacing later edits. Loading inputs does not automatically save, train, calculate, search or contact a worker.
  • Exact synthetic-example Design, Evolution and Model runs are labeled Synthetic demo when explicitly executed. Runs continues to show real saved history, and the existing Experiments and Campaigns examples remain available.
  • Help explains the examples, their prerequisites, explicit next actions and how to clear them.

Why it matters

Demonstrate connected research workflows with less typing while keeping example inputs distinguishable from real research.

Checks & evidence

Software verification

  • Frontend and backend checks covered reversible filling, replacement confirmation, state preservation, account isolation and truthful saved-run labels. Local Design, Evolution, Model and Molecular Design preparation were exercised with isolated example inputs.
  • Desktop, tablet and phone viewport layouts were checked. On the deployed Workbench, Model filling staged all 64 synthetic variants, clearing restored the prior inputs, and neither action made an HTTP request.

Scientific evidence

No new scientific validation. Synthetic sequences and hypothetical scenarios demonstrate software behavior; public reference data remains distinct from predictions and experimental measurements.

Appearance

A calmer Workbench with smoother motion

A softer blue-gray Workbench, brief tab transitions, and larger protein illustrations in Design and Evolution.

What changed

  • Design and Evolution illustrations fill their existing space more naturally, with full-rotation clearance and unchanged geometry.
  • Rotation follows elapsed time at the display refresh rate and pauses when outside view, hidden or loading. Reduced-motion preferences keep the illustrations still.
  • The individual Pause Animation controls have been removed from Design and Evolution. Other visualization controls remain available.
  • Workbench backgrounds use softer blue-gray surfaces. Switching tabs uses a brief fade and slide while preserving inputs and results; reduced-motion navigation is immediate.
  • Help explains the current illustration and tab behavior.

Why it matters

Move between research tasks with less visual glare and smoother transitions, while keeping the surrounding controls and results clear.

Checks & evidence

Software verification

  • Application and Help checks covered frame-rate independence, full-rotation fitting, automatic suspension, reduced motion, rapid tab switching and preserved inputs.
  • Desktop, tablet and phone viewport layouts were inspected, and both local workflows completed using isolated test inputs. Deployed pages and assets were checked against the verified change.

Scientific evidence

No new scientific validation. These decorative illustrations are not predictions of the working sequence or experimental results.

Usability

Pitch examples for Experiments and Campaigns

Load consistent synthetic examples into each workflow, review the inputs and use the normal save, ranking and campaign actions.

What changed

  • Fill Example and Clear Example controls stage editable inputs and restore earlier drafts. Replacement asks for confirmation; saved research is not deleted or overwritten.
  • Experiments and Campaigns share a clearly labeled synthetic parent, 64 variants and 24 invented assay values. Compatible FASTA and CSV examples are available to download.
  • Demo batch settings use available candidates and the sample budget. Example outcomes use the actual frozen round's sample IDs, including explicit measured, censored, failed and missing outcomes.
  • Help explains the demonstration, explicit actions, synthetic-project restrictions and how to repeat it. Existing project permissions and the public GB1 teaching example remain unchanged.
  • The upper-left DeltaFold logo returns from the Workbench to the welcome page. The separate Home link has been removed.

Why it matters

Demonstrate the workflow from assay inputs to ranking, a frozen selection and outcome review without manually entering every example field.

Checks & evidence

Software verification

  • Frontend and backend checks covered reversible draft filling, file restoration, account isolation, duplicate protection, actual ranking, frozen records and outcome review.
  • Desktop, laptop, tablet and phone viewport layouts and keyboard controls were checked. Both deployed workflows completed with isolated synthetic records; published assets matched the verified commit.

Scientific evidence

No new scientific validation. The example sequences and assay values are synthetic and demonstrate software behavior, not laboratory performance or predictive accuracy.