Every SciKeep tool carries a verification grade, and the grade reflects what has been tested against something external — not how finished the tool looks. We publish the whole table, including the tools that score badly, because a product claiming uniform excellence is not one you should trust with a number that ends up in a paper.
Corrections log
Errors we found in SciKeep, newest first, including ones that could have changed a result. Nothing is removed from this list.
2026-09-25 — Site: Raw LaTeX delimiters showed as literal text in three method notes
Usability, no effect on results
What was wrong: Four sentences across the MSA, master-mix and drug-synergy method notes were written with LaTeX math delimiters ($...$, $$...$$, \text{}, \times) that the article renderer does not interpret, so a reader saw the literal markup — "$N$", "$$E_{\text{expected}} = ...$$" — instead of the intended notation.
Who was affected: Anyone reading those three posts; no calculation or number was affected, only the prose.
What changed: Rewritten as the plain-text math notation the rest of the site uses (e.g. "Penalty = g_o + (k − 1) × g_e").
2026-09-24 — Site, Electronic Lab Notebook, Grant Justification Kit: Compliance and validation language overclaimed in five places, on top of the "immutable" fix earlier the same day
A statement on the site was wrong
What was wrong: A sweep for the same pattern as the "immutable" audit trail found five more: (1) the section introducing all 42 tools said every one was "validated against wet-lab benchmark data" — false for most of them; freezer, antibody and reagent trackers are record-keeping with nothing to benchmark, several calculators have no automated test at all, and the tool-by-tool validation page itself lists which are checked against a formula, hand arithmetic, or real image data instead. (2) The same overclaim appeared as the shorter tagline "Published methods, benchmarked and cited" in the hero and the pricing modal. (3) The Western blot tile's badge read "ImageJ Standard", naming a specific competing product as a match, while the validation page says lane densities are "not benchmarked against ImageJ on real film or ChemiDoc images." (4) Signing a notebook entry — a typed name and a checkbox, with no identity verification — stamped the record with the declaration "Certified accurate, complete, and reproducible experimental record under 21 CFR Part 11 standards", and the modal was titled "21 CFR Part 11 Electronic Signature", though the validation page states this signature "does NOT meet 21 CFR 11.200." (5) The Grant Justification Kit generated boilerplate meant to be pasted into a real NIH, NSF or Horizon Europe grant application asserting SciKeep's local execution was "ensuring compliance with institutional IRB, HIPAA, and IP governance protocols", would "guarantee... FAIR... compliance", was "fully aligning with NSF... mandates", and — in the modal's own subtitle — was flatly "Compliant with NIH NOT-OD-21-013 DMS Policy requirements". A researcher who pasted that text into a federal submission was making a compliance representation on SciKeep's behalf that the product cannot back up.
Who was affected: Anyone who read the 42-tool intro or either tagline as a validation claim, anyone who took a signed notebook entry's declaration as a Part 11 certification, and anyone who copied the grant boilerplate into an actual NIH, NSF or Horizon Europe application.
What changed: The tool-intro copy and both taglines now say tools are graded openly on the validation page rather than uniformly "benchmarked". The blot badge names its method ("Rectangular Lane Profile") instead of a competing product. The signature modal is titled "Electronic Signature" with a line stating it does not meet 21 CFR 11.200, and the stored declaration is the signer's own attestation rather than a certification claim. The grant kit's three templates now say a claim is "supported by" or "supports" the relevant requirement rather than "ensures compliance" or "guarantees" it, its subtitles no longer assert compliance status, and the modal carries a standing notice that it is a draft to edit, not a compliance certification.
2026-09-24 — Site, Electronic Lab Notebook: The audit ledger was still called "immutable" in nine places
A statement on the site was wrong
What was wrong: The 31 August correction removed "legally immutable" from the notebook itself, and the tool, the features page and the validation ledger have since said plainly that a hash chain held in your own browser is tamper-EVIDENT and not tamper-proof — someone with the developer console can rewrite the whole chain, and clearing site data removes it. The word survived everywhere else: the landing page (three places), the pricing modal, the comparison table, the FAQ, the Part 11 method note and its search description, and the notebook's own empty audit-trail message. A reader who saw only those surfaces would take the ledger to be something the product has never claimed to be in the places that describe how it works.
Who was affected: Anyone who judged SciKeep's Part 11 posture from the marketing pages, the pricing modal, the comparison table or the FAQ rather than from the tool and the validation page.
What changed: Every remaining "immutable" is now "tamper-evident", and the FAQ answer states outright that the chain gives detection rather than prevention, that the console can rewrite it and that clearing site data removes it. The comparison table row and the landing-page feature name were renamed to match. The word is left untouched in this log, where it records what was said.
2026-09-24 — Site: The isoelectric point article cited a paper about extinction coefficients as its source
A statement on the site was wrong
What was wrong: The method note "Why Your Protein's pI Matters" carried DOI 10.1002/pro.5560041120 — Pace et al. 1995, "How to measure and predict the molar absorption coefficient of a protein". That paper is the correct source for the article's ε280 formula and nothing else in it; it says nothing about isoelectric points. The article also stated the pKa set and the algorithm without saying that different pI calculators use different pKa tables, that ExPASy uses the Bjellqvist values rather than the textbook ones SciKeep uses, or that a folded protein's measured pI can sit half a pH unit or more from any sequence-derived number. Its markdown also rendered raw: bullet lists collapsed into run-on paragraphs and italics printed with their asterisks showing, across nine method notes.
Who was affected: Anyone who followed the citation expecting the source of the pI calculation, or who read the article as saying a calculated pI should agree with a published or measured one.
What changed: The citation is now Bjellqvist et al. 1993, Electrophoresis 14(10):1023–31, doi:10.1002/elps.11501401163 — the primary reference for predicting focusing positions from sequence. Pace 1995 is still cited, in the section it actually supports. The article was rewritten to state which pKa table SciKeep uses, why tools disagree, and why the folded protein disagrees with all of them; the validation page already recorded that the pI is not benchmarked against ExPASy. The article renderer now handles lists, line breaks and italics.
2026-09-24 — Sanger Trace Viewer, Reagent Tracker, Buffer Calculator: A schematic peak diagram was described as fluorescence channels, and three smaller mislabels
A statement on the site was wrong
What was wrong: The peak diagram is drawn from the called bases, one peak per base, but the page described it as "four fluorescence channels" and the legend named BigDye terminator dyes — a chemistry this view never reads. It cannot show a secondary peak the basecaller did not call, which is the main thing someone looks at a trace for. The reagent tracker opened on six invented orders with made-up people and grant codes, with no example notice, driving a red low-stock alert and a spend-by-grant panel that looked like the user's own. Its CSV export joined fields with a bare comma, so a reagent name containing a comma shifted every later column. The buffer list also offered "NaCl Stock (5M)" while the default produced 1 M, and showed a concentration control on the 0.5 M EDTA recipe that changed nothing.
Who was affected: Anyone who read the peak diagram as a real trace, took the reagent low-stock or spend figures as their own, or exported a reagent CSV containing a comma in a name.
What changed: The diagram is labelled a schematic and the dye names are gone; the reagent tracker shows an example notice and both CSV exports use the shared escaping exporter; the NaCl label no longer claims 5 M and the inert EDTA control is hidden.
What to do: Re-open any reagent CSV exported before today if a reagent name contained a comma.
2026-09-24 — Sequence Toolkit, Cloning Planner: Half of all sites for the Golden Gate enzymes were invisible, and Dcm warnings fired on unrelated motifs
A result could have been wrong
What was wrong: The restriction digest scanned only the forward strand. For palindromic enzymes that is the same answer either way, but for BsaI, BsmBI and BbsI — the enzymes people check before Golden Gate — roughly half of real sites sit on the reverse strand and were simply not found. The Dcm methylation check also flagged a site whenever a CCWGG motif appeared anywhere within ten bases, without requiring it to overlap the recognition sequence, so sites that would cut perfectly well were reported as blocked. The Gibson planner additionally demanded an overlap melting temperature between 48 and 55 °C; the manufacturer's guidance is a floor of about 48 °C with no upper limit, so ordinary overlaps were marked "not optimal". Toolkit descriptions also claimed six-frame translation, a sliding GC window, a mouse codon table and NCBI gene retrieval, none of which this tool has.
Who was affected: Any digest or Golden Gate check for a non-palindromic enzyme, any site reported as Dcm-blocked, and any Gibson overlap rejected for being too warm.
What changed: Both strands are scanned, Dcm requires a real overlap, the Gibson check enforces only the floor, and the descriptions list what the tool does.
What to do: Re-check any construct you cleared for internal BsaI, BsmBI or BbsI sites.
2026-09-24 — Electronic Lab Notebook: The audit trail recorded only signatures, and "hash chain verified" was never checked
A result could have been wrong
What was wrong: The ledger was written only when an entry was signed. Creating, editing, deleting and exporting recorded nothing, so a notebook of unsigned entries had an empty ledger and an entry could be rewritten with no trace — while the page said "Every create, edit, signature and deletion is recorded in an append-only ledger". The chained value also left the entry text out, so even a logged action did not commit to what the entry said. The green "✓ Cryptographically sealed — hash chain verified" line appeared whenever a signature existed; no verification code existed anywhere in the product. Editing a signed entry kept the signature, so a PI's name stayed on text they never saw. The "This month" count compared the month but not the year. Site copy also called the ledger "legally immutable", offered a "GLP compliance certificate (PDF)" that is a JSON file, and said SciKeep "complies with FDA requirements".
Who was affected: Anyone relying on the ledger as a record of what happened to their entries, or on the signature seal as evidence that an entry was unchanged.
What changed: Create, edit and delete are recorded, the chain covers a hash of the entry text, and the chain is recomputed in the browser with the real result shown. Editing a signed entry now clears the signature and says why. The compliance and certificate claims were corrected, and the tool now states plainly that signing needs only a typed name, which does not meet 21 CFR 11.200, and that a ledger in your own browser is tamper-evident rather than tamper-proof.
What to do: Treat any entry edited before today as unverified, and re-sign entries whose text changed after signature.
2026-09-24 — Western Blot Analyzer: Per-lane significance stars were uncorrected, and two stated limits were not what the tool does
A result could have been wrong
What was wrong: The per-lane table ran a separate Welch test of every treated lane against the control and printed stars from the raw p-value, with no correction for the number of comparisons — five treated lanes carries about a 23% chance of at least one false star — while the public page said a Holm correction was applied by default. That was true only of the group chart. The page also said a band under about 10% above background is "flagged as under-measured", and the confidence note put the shortfall at "roughly 20%": no code implemented either figure. Measured on bands of known density, a band below about 10% above background is not detected at all, and a detected band recovers about 9-10% low whatever its intensity. The offline record for vinculin also gave 116.6 kDa, which is an isoform mass, against an accession whose canonical protein is 123.8 kDa — while the tool's own dropdown said 124.
Who was affected: Significance stars in the per-lane table, and anyone who relied on the stated faint-band behaviour or the offline vinculin weight.
What changed: Lane p-values are Holm-adjusted across the lanes tested and the stars follow the adjusted value. The faint-band and pairwise-comparison descriptions now say what the code does, backed by tests that measure it. Vinculin reads 123.8 kDa.
What to do: Re-check any significance star taken from the per-lane table where more than one treated lane was compared.
2026-09-24 — RNA-Seq Volcano Plot: A "Real Data" badge sat above invented example tables, and raw p-values were plotted as adjusted
Showed fixed or invented output as analysis
What was wrong: The tool carried a green "✓ Real Data" badge, and its three example datasets were described as published DESeq2 output from named cell lines. They are illustrative tables written by hand; no source, accession or citation exists for any of them. Separately, if an uploaded table had no adjusted p-value column the raw p-value column was plotted and thresholded silently, while the axis, the slider and the exported column all still read "p-adj" — the tool's own FAQ warns that this overstates confidence. Rows with a blank or zero p-adj were dropped with no count, and a blank p-adj is ordinary DESeq2 output for independently filtered genes, often a third of a real table. Site copy also described the tool as applying "DESeq2 thresholds" and "DESeq2 & EdgeR statistical standards"; it runs no model or test of any kind.
Who was affected: Anyone who read the example data as measured results, or who uploaded a table without an adjusted p-value column, or who read the up/down/not-significant totals as covering their whole table.
What changed: The badge is gone and the examples are labelled illustrative. A table without an adjusted p-value column now says so plainly, and dropped rows are counted and explained. The copy says what the tool does: it plots and thresholds a table you already have.
What to do: If you exported significant genes from a table with only a raw p-value column, re-run with adjusted values.
2026-09-24 — Figure Panel Composer: Scale bars were drawn from an assumed calibration, and Science figures exported 9 mm too narrow
A result could have been wrong
What was wrong: Every scale bar was computed as if the source image were 512 pixels wide and every pixel 0.16 µm, and there was no control anywhere in the interface to enter your own pixel size. A 2048-pixel confocal frame therefore got a bar four times too long. The drawn bar was also capped at 40% of the panel while its micron label was printed unchanged, so once capped the bar and its label no longer agreed. Scale bars were on by default. Separately, the Science preset used 55, 120 and 175 mm where Science publishes 5.7, 12.1 and 18.4 cm, so a full-width figure exported 9 mm narrow and the journal would scale it up, dropping it below the 300 DPI the tool tells you to hit. PNAS 1.5-column was 115 mm rather than 114.
Who was affected: Any exported figure with a scale bar, and any figure sized to the Science or PNAS presets.
What changed: Scale bars need a pixel size you enter and the image's real width, both now used in the calculation; a bar that cannot be drawn to scale is not drawn, and the panel editor says why. All four journals' widths are now asserted against their published guidelines — the previous test read the width from the preset it was checking, so it would have passed on any number.
What to do: Re-measure the scale bar on any figure exported from this tool, and re-export anything sized for Science.
2026-09-24 — Cell Culture Tracker: Five published doubling times were not in the Cellosaurus record they were badged with
A result could have been wrong
What was wrong: Each line showed a doubling time beside a "Cellosaurus ✓" badge, but five of the ten values do not appear in the cited record: HeLa was 24 h where Cellosaurus reports 30-48 h, HEK293 20 h and HEK293T 18 h where it reports about 24-30 h, CHO-K1 16 h where it reports about 24 h, and NIH-3T3 18 h where it reports 20-21 h. "Jurkat" also carried the accession of the Jurkat E6.1 clone rather than the parent line. Separately, "Mark as Split" wrote a post-split confluency of 15% no matter what the split ratio was, and that figure is the starting point for the observed doubling time shown on the card — so a measurement presented as the dish's own was computed from a number the user never entered.
Who was affected: Anyone who planned a split or an experiment around the published doubling time for HeLa, HEK293, HEK293T, CHO-K1 or NIH-3T3, and every observed doubling time after a split.
What changed: All ten values were transcribed from their own Cellosaurus records and are checked against saved copies of those records by a test. The split action now asks for the confluency you actually see.
What to do: Re-check any timing planned around these figures — HeLa in particular was stated at half its published doubling time.
2026-09-24 — Buffer Calculator: The RIPA lysis buffer recipe was ten times too dilute
A result could have been wrong
What was wrong: RIPA's five amounts were written as grams per 100 mL but multiplied by the volume in litres. At the tool's default 100 mL batch it printed 0.08 g Tris-HCl, 0.09 g NaCl, 0.10 mL NP-40, 0.05 g deoxycholate and 0.01 g SDS — a tenth of every concentration, while the labels still read 50 mM, 150 mM, 1%, 0.5% and 0.1%. The other recipes in the same list use per-litre amounts and were correct.
Who was affected: Anyone who made RIPA from this recipe: the buffer had a tenth of the detergent and will not solubilise membrane or nuclear proteins.
What changed: RIPA now uses the same tested molarity x molecular weight arithmetic as PBS and TBS, with the detergents as real percentages of the batch volume.
What to do: Remake any RIPA prepared from this tool, and treat a western blot that showed no band for a membrane or nuclear protein as untrustworthy.
2026-09-24 — Cloning Planner: Golden Gate ignored the chosen enzyme, and gave BsmBI the wrong reaction temperature
A result could have been wrong
What was wrong: The type IIS enzyme you selected was shown in the output but used in no calculation, so parts were never screened for the enzyme's own recognition site — the most common reason a Golden Gate reaction fails. The planner's own default CMV promoter part contains a reverse-strand BsaI site and was reported as a valid, high-fidelity assembly with no warnings. The printed protocol also cycled BsmBI at 55 °C, which is its plain-digest temperature; NEB's Golden Gate protocol for BsmBI-v2 cycles at 42 °C, and T4 ligase is inactive at 55 °C, so nothing would have been ligated.
Who was affected: Any Golden Gate assembly planned here, and anyone who ran the printed BsmBI protocol.
What changed: Every part is scanned on both strands for the selected enzyme's site; a hit invalidates the assembly and names the position. Warnings are now shown at all — the result carried them but nothing rendered them. BsmBI cycling reads 42 °C.
What to do: Re-check any Golden Gate design from this tool for internal sites, and re-run a BsmBI reaction that used the 55 °C protocol.
2026-09-24 — CRISPR Indel Analysis: Reads trimmed differently from the reference were called as edits
A result could have been wrong
What was wrong: The alignment charged a full gap penalty at the ends, so the cheapest way to reconcile two reads that started or finished at different bases was to call the overhang an indel. A read identical to the reference but trimmed 12 bases was reported as a "12 bp deletion, in-frame", and a genuine 3 bp in-frame deletion in a read trimmed at both ends was reported as a 23 bp deletion causing a frameshift. Two Sanger reads almost never start and end at the same base, so this affected ordinary use.
Who was affected: Any indel call where the edited read and the reference were trimmed differently — including negative controls, which could read as deletions.
What changed: Leading and trailing gaps are free, so only indels inside the overlapping region are counted. Trimmed identical reads now report "no edit", and a 3 bp deletion reads as 3 bp and in-frame whether or not the reads are trimmed.
What to do: Re-run any indel call made here, especially one that reported a deletion the size of your trimming.
2026-09-24 — Sanger Trace Viewer, CRISPR Indel Analysis: Trace files were read from the wrong byte offsets, and the quality histogram showed fixed numbers
A result could have been wrong
What was wrong: Both .ab1 readers used directory-entry offsets one field too far. In the indel tool the data offset came from a field that is zero in real files, so every genuine .ab1 was rejected with a message telling the user to re-export it from their sequencing provider. In the trace viewer the element count evaluated to a constant 65,536, so the reader either pulled tens of kilobytes of trace binary in as extra "bases", or skipped the basecall record and fell back to matching the longest run of A/T/G/C in the raw binary and presenting that as the read. The per-base quality histogram was drawn from fixed values (18, 24 or 45) and never used the quality record at all, and the accuracy caption under the mean quality was hardcoded to the Q30 figure, so a mean of Q12 was labelled ">99.9% accuracy" when it is 94%.
Who was affected: Anyone who uploaded a real .ab1: the indel tool refused it, and the trace viewer could show a sequence and a quality profile that were not from the file.
What changed: One shared reader now uses the offsets the ABIF specification defines, cross-checked against the file header's own root entry. It returns the real per-base quality, the histogram and the accuracy figure are computed from it, and the binary-mining fallback is gone. The test fixture was built to the specification rather than to the parser — the previous fixture reproduced the parser's own wrong offsets, which is why this passed unnoticed.
What to do: Re-upload any .ab1 you read here, and re-check any sequence or quality figure taken from one.
2026-09-24 — siRNA Knockdown Designer: The off-target seed screen was reading the wrong strand, and the antisense overhang was on the wrong end
A result could have been wrong
What was wrong: The microRNA seed screen and the seed GC% were computed from the sense (passenger) strand. RISC loads the guide (antisense) strand, and that is the strand whose positions 2-8 cause microRNA-like off-target silencing, so a guide carrying a real oncomir seed was reported as "Low" risk. Update 75 corrected the seed sequences; the check was still pointed at the wrong molecule. Separately, the antisense oligo was written with TT at its 5' end instead of a 3' dTdT, so an oligo ordered from that line was wrong at both ends of the guide.
Who was affected: Every siRNA ranked or ordered from this tool: off-target risk, seed GC% and the antisense sequence.
What changed: Both the seed screen and the seed GC% run on the guide strand, and both strands carry a 3' dTdT. A test now builds the target from the guide, so it would fail on the old behaviour.
What to do: Re-run any siRNA you designed here, and re-check the antisense sequence of anything already ordered.
2026-09-24 — Standard Curve Calculator: On a 4PL curve the top standard interpolated to zero, and mid-curve values were up to 23% out
A result could have been wrong
What was wrong: The curve's two asymptotes were fixed to the lowest and highest observed readings rather than fitted. The top standard's reading therefore equalled the upper asymptote exactly, the inverse divided by zero, and any sample at the top of the curve was reported as 0 — a saturated well read as "analyte absent". Pinning the asymptotes also bent the curve: on noiseless data from a known 4PL, points across the middle came back 11-23% low. Readings above the curve were reported as 1.5x the top standard, a number no measurement produced, which was written into the exported CSV. The below-range flag compared against the blank instead of the lowest calibrator, so readings under the first standard were reported as quantifiable.
Who was affected: Every 4PL interpolation (the default for ELISA): concentrations across the curve, and anything at or above the top standard.
What changed: All four parameters are fitted, so the same known curve now reads every standard back within 5% and the top standard to itself. A reading beyond an asymptote reports "outside the curve" instead of a number, and range flags use the calibrated range.
What to do: Re-interpolate any ELISA or BCA result taken from a 4PL curve here, especially samples near the top standard.
2026-09-24 — Share Tools: Every "Copy Link" button copied a link that 404s
Usability, no effect on results
What was wrong: The eight tool cards carried hand-written paths (/tools/qpcr, /tools/crispr, /tools/ic50 and so on). None matched the tool's real address, so every copied link opened a not-found page for whoever received it, while the toast said "Copied to clipboard". The links also used the non-www domain, which redirects. The "Share Protocol" checkbox copied a /protocol/<id> link to a page that does not exist.
Who was affected: Anyone who sent a tool link or a protocol link from this panel.
What changed: Links are built from each tool's real route, so they are correct by construction, and a test fails if one ever drifts again. The protocol checkbox copies the protocol text instead of a dead link.
What to do: Re-send any SciKeep tool link you shared before today.
2026-09-24 — Team Collaboration: "Invitation Sent!" sent nothing, and "Team access revoked" revoked nothing
A statement on the site was wrong
What was wrong: Adding someone to the lab roster showed "Invitation Sent!" but no email was sent and no account access was granted; removing them said "Team access revoked" while changing only the local list. A "Copy Lab Join Link" button produced a link containing a token that nothing in the app reads.
Who was affected: Anyone who believed colleagues had been invited or that removing a row had revoked their access.
What changed: The roster is described as a list the lab keeps, with an Email invite button that opens your mail client, and the dead join link was removed. Account access is granted through lab membership, which is separate.
What to do: Confirm directly with anyone you believed was invited or removed.
2026-09-24 — 3D Structure Viewer: Alternate conformers were drawn twice, and AlphaFold files lost their confidence colouring
A result could have been wrong
What was wrong: Atoms modelled in more than one position were all kept, so a high-resolution structure showed extra atoms and its backbone trace zig-zagged between conformers: lysozyme 2VB1 was drawn with 2,900 atoms and 70 residues carrying two alpha-carbons. Residues numbered 52, 52A and 52B (antibody numbering) were counted as one residue. An AlphaFold model downloaded and then uploaded as a file was not recognised as a prediction, so pLDDT colouring showed flat grey. The page also said ribbons were derived from backbone geometry; helix and sheet are read from the file's own records, so any file without them — including every AlphaFold model — is drawn entirely as coil with nothing said.
Who was affected: Atom and residue counts, and backbone traces, for structures containing alternate conformers or insertion codes; pLDDT views of uploaded AlphaFold files.
What changed: Only the first conformer is kept (2VB1 now parses to 2,200 atoms and 129 alpha-carbons for its 129 residues), insertion codes are distinguished, AlphaFold files are recognised from their title and open in pLDDT colouring, and the viewer says when a file carries no secondary-structure records.
What to do: Re-check any atom or residue count taken from a structure with alternate conformers.
2026-09-24 — CRISPR gRNA Designer: A fetched or pasted gene froze the tab for seconds to minutes
Usability, no effect on results
What was wrong: The off-target search ran for every candidate guide during the scan, and each run walked both strands of the whole sequence, so the work grew with the square of the sequence length. Measured on a 24 kb sequence, the page was unresponsive for 24 seconds.
Who was affected: Anyone who pasted or fetched more than a few kilobases; no result was wrong, but the tool appeared to hang.
What changed: The off-target search now runs only for the ten guides that are shown, which does not change any reported number. The same 24 kb sequence now blocks the page for 12 ms.
2026-09-24 — CRISPR gRNA Designer: Cloning oligos were labelled by enzyme when the difference is the vector promoter
A statement on the site was wrong
What was wrong: Two oligo pairs were offered, "BsmBI" without an added G and "BsaI" with one. The extra G is what the U6 promoter needs to start transcription when the guide does not begin with G; it has nothing to do with which type IIS enzyme cuts the vector. A guide starting with G was given an extra G anyway, and oligos were shown for SaCas9 and Cas12a guides, which use different vectors.
Who was affected: Anyone who ordered oligos from the row whose label matched their enzyme rather than their vector.
What changed: One pair is shown for SpCas9 20-mers, with the G added only when the guide needs it, and the note names the vector (lentiCRISPRv2 with BsmBI, pX330 with BbsI). No oligos are shown for nucleases that need another vector.
What to do: Check any ordered oligos: the top should read CACC + G only if your guide does not already start with G.
2026-09-24 — Freezer Inventory: Every added sample was saved as a plasmid at 100 ng/µL, 50 µL, added by "Current Researcher"
Showed fixed or invented output as analysis
What was wrong: The Add form had no fields for type, concentration, volume or owner, so every item was stored as type Plasmid with concentration 100 ng/µL, volume 50 µL and owner "Current Researcher", and these values synced to your account. Items could not be deleted. The page described rack and shelf maps and 81- or 100-place boxes that the tool does not have.
Who was affected: Every item added through the form: its type shows as Plasmid regardless of what it is, and exported or synced records carried the invented values.
What changed: The form has type, concentration, volume and owner fields, items can be removed, the invented values are cleared from items saved earlier, and the description matches the tool.
What to do: Check the type of each item you added before today; it was recorded as Plasmid.
2026-09-24 — Oligo & Primer Resuspension Calculator: The G extinction coefficient came from a different reference set, so G-rich oligos got too much water
A result could have been wrong
What was wrong: ε260 used dA 15,200, dC 7,050 and dT 8,400 from one published set but dG 11,500 from another; the matching value is 12,010. ε was therefore too low for G-containing oligos (about 2.5% for a G-rich 25-mer), so the calculated amount and the resuspension volume were too high and stocks came out slightly below their stated concentration. The page also described ε as nearest-neighbour; it is a nucleotide sum.
Who was affected: Resuspension volumes and nmol amounts for G-containing oligos.
What changed: dG is 12,010, the calculator uses the shared tested functions, and the description says what the method is.
What to do: For G-rich primers where a few percent matters (for example qPCR standards), recompute the stock concentration.
2026-09-24 — Dilution Calculator: Impossible dilutions were printed as bench instructions
A result could have been wrong
What was wrong: A target more concentrated than the stock gave a negative diluent volume (for example "add 50000 µL stock to -40 mL buffer"), and a zero stock gave "Infinity µL". The page also described unit conversion and serial dilution, which the tool does not do.
Who was affected: No correct calculation was affected; impossible inputs produced impossible instructions.
What changed: Impossible inputs are refused with the reason, and the description matches the fixed units (stock mM, target µM, volume mL).
What to do: If you made a solution from a negative or infinite volume shown here, remake it from a stock more concentrated than the target.
2026-09-24 — Pre-Submission Manuscript Audit: The audit passed items it had not found and wrote claims into the suggested reviewer response
A result could have been wrong
What was wrong: The MIQE amplification-efficiency item passed for any text containing "R2", so a qPCR paper about HER2 or NR2B passed with no efficiency data, while a real "slope of -3.35 (R2 = 0.99)" did not match. The suggested reviewer response said "Statistical power analysis was performed" whenever an n was given and called any RRID "authenticated"; a found RRID was reported as "compliant with ARRIVE antibody guidelines", though ARRIVE has no antibody item. The page claimed 25 checks against ARRIVE, CONSORT and MIQE (there is no CONSORT check, and several checks apply only to some manuscripts), numbered SAMPL items S1-S5 (SAMPL does not number its items), and described the AI reviewer as a Gemini model trained on journal review rubrics.
Who was affected: MIQE efficiency results for manuscripts containing gene names with "R2", and any reviewer-response text copied from the audit.
What changed: The efficiency pattern requires a slope with its R², an efficiency or a standard curve. The response paragraph states only what the text reports. The counts, guideline names, SAMPL labels and AI description now match what runs.
What to do: Re-read any reviewer response copied from the audit and remove claims about power analysis or antibody authentication your manuscript does not make.
2026-09-24 — Transfection Calculator: Lipofectamine 3000 amounts were wrong for large vessels and left out P3000 Reagent
A result could have been wrong
What was wrong: Amounts came from fixed multipliers that were not the vessels' growth areas: a 10 cm dish got 10 µg DNA in 8 mL, where the manufacturer's table gives 14–28 µg in 10 mL, and a T-75 was under-scaled similarly, while a 96-well well got 0.13 µg instead of 0.1 µg. The steps for Lipofectamine 3000, the default reagent, never mentioned the required P3000 Reagent, and used 2 µL of lipid per µg where the protocol tests 1.5 and 3 µL. Every other reagent was given the same DNA amounts and Opti-MEM steps, and an "expected efficiency" was shown with no source.
Who was affected: Transfections planned with this tool, especially in 10 cm dishes, T-75 flasks or 96-well plates, or with Lipofectamine 3000 without P3000.
What changed: Lipofectamine 3000 amounts now come from Thermo Fisher's protocol and scaling table, including P3000 and both lipid doses. For other reagents, you enter a condition that works and it is scaled by growth area. The efficiency estimate was removed.
What to do: If a Lipofectamine 3000 transfection planned here underperformed, check that P3000 Reagent was included and re-scale from the manufacturer table.
2026-09-24 — Antibody Tracker: RRID lookup filled in the wrong antibody for most RRIDs
Showed fixed or invented output as analysis
What was wrong: The lookup read the registry's reply in a format it never uses, so it never used the live SciCrunch record. It fell back to five stored records, three of which were attached to the wrong RRID (AB_2687626 is an HRP secondary, not Ki-67; AB_10694088 is a biotinylated cleaved caspase-3 antibody, not GAPDH; AB_2178887 is an NF-κB p65 antibody, not cleaved caspase-3), and any other RRID was answered with the Ki-67 record under your RRID. Missing fields were filled with "Rabbit", "Cell Signaling Technology" and WB/IHC/IF.
Who was affected: Any antibody added with the RRID Lookup button: its name, target, host, vendor and catalogue number may belong to a different antibody.
What changed: The live registry record is read correctly; offline, only five records copied from the registry are used, and any other RRID returns "not found". Nothing is filled in that the registry does not state. The SEO page's example RRID was also wrong and has been replaced.
What to do: Check each antibody you added with RRID Lookup against scicrunch.org/resolver before citing it.
2026-09-24 — Methods Section Writer: Template Methods paragraphs stated kits, catalogue numbers and guideline compliance you never entered
Showed fixed or invented output as analysis
What was wrong: Without AI, the writer filled in specific kits, catalogue numbers, software versions, dilutions and cell densities that were not among your inputs, and some templates ended with a compliance statement (ARRIVE, MIQE, "Nature Reporting") that nothing had checked. The example flow-cytometry RRID (AB_314154) is an anti-CD11b antibody, not the CD4/CD8 antibodies named; two other example RRIDs did not match or did not exist.
Who was affected: Any Methods paragraph copied from the template (non-AI) path.
What changed: Anything you have not entered is now a [bracketed placeholder]; no template claims compliance. The page shows an example notice until you change the pre-filled fields, and the wrong example RRIDs were replaced with placeholders.
What to do: Re-read any template-generated Methods text against your records; remove any kit, catalogue number, version or compliance sentence you did not supply.
2026-09-24 — Protocol & SOP Writer: The protocol library was described as generated from your inputs and "Bench Validated"
A statement on the site was wrong
What was wrong: The page and its confidence note said protocols were generated from your inputs with values echoed from what you entered, and a badge said "Bench Validated". The tool is a fixed library of twelve reference protocols that takes no input and was not validated. Its CRISPR protocol also gave Nucleofector program CA-137 for HEK293T; Lonza's own HEK293 protocol uses CM-130.
Who was affected: Anyone who followed a library protocol believing its values were tailored to them or validated.
What changed: The page, badge, description and note now call it a reference library, the copied and downloaded text carries a notice to check each value, and the program code points to Lonza's optimized-protocol list.
What to do: Check any Nucleofector program you took from the library against Lonza's protocol for your cell line.
2026-09-24 — Drug Synergy Matrix: A "published NCI ALMANAC reference" panel showed synergy scores that no published source contains
Showed fixed or invented output as analysis
What was wrong: The tool offered five "verified" drug pairs with Bliss and Loewe scores and record IDs attributed to NCI ALMANAC, DrugCombDB and SynergyFinder, and showed the matching score next to your result as the "expected" value. NCI ALMANAC screened only FDA-approved drugs on the NCI-60 cell panel and reports ComboScores, not Bliss or Loewe scores; three of the five pairs used cell lines outside that panel, one used a drug that is not FDA-approved, and none of the record IDs could be traced. Site pages also claimed a "5,000+" or "2M" pair benchmark database and Loewe, ZIP and HSA models. Only Bliss has ever been computed, and only five pairs were ever stored.
Who was affected: Anyone who compared their result to, or cited, one of the reference scores, or who described the tool as benchmarked against NCI ALMANAC.
What changed: The reference table, the pair selector and the reference panel were removed, and every page now describes the tool as a Bliss independence calculator. The Methods paragraph no longer claims cross-referencing against NCI ALMANAC. Your own Bliss scores were never affected.
What to do: Remove any reference score or NCI ALMANAC comparison taken from this tool from your notes or manuscript.
2026-09-24 — qPCR ΔΔCt Expression: Pfaffl mode ignored the reference gene's own efficiency
A result could have been wrong
What was wrong: In Pfaffl mode, which is the default, the whole ΔΔCt was raised to the target gene's efficiency, so the reference gene's measured efficiency was never used. With a target at 90% and a reference at 110%, a 2-cycle target shift and a 1-cycle reference shift gave 1.90-fold instead of Pfaffl's 1.72-fold. The error grows with the gap between the two efficiencies and with the size of the shift.
Who was affected: Pfaffl fold changes, intervals and p-values where the target and reference efficiencies were not equal. Livak results, and Pfaffl results with equal efficiencies, were correct.
What changed: Each gene's Ct is now weighted by its own efficiency, exactly as in Pfaffl (2001), and a hand-worked test checks the 1.72 case.
What to do: Re-run any Pfaffl analysis where the target and reference efficiencies differed.
2026-09-24 — siRNA Knockdown: Four of the five microRNA seeds in the off-target screen were wrong
A result could have been wrong
What was wrong: The screen that warns when a guide carries a known microRNA seed held the wrong 7-nucleotide seed for miR-21, let-7a, miR-34a and miR-126 (only miR-17 was right), so guides carrying those real seeds were not flagged, and some unrelated guides were.
Who was affected: siRNA designs ranked before this date: a guide could have been shown as low-risk while carrying one of these seeds.
What changed: All five seeds are now positions 2-8 of the miRBase mature sequence (accession cited), checked by a test that derives each seed from the published sequence.
What to do: Re-run the designer on any guide you chose, and check its seed region against miRBase.
2026-09-23 — Reagent Orders & Stock: The order-count and spend tiles showed fixed numbers
Showed fixed or invented output as analysis
What was wrong: The "orders this month" and "spent this month" tiles always showed 12 and $3,450, whatever you had entered, and the add, edit and delete buttons did nothing.
Who was affected: Anyone who read those tiles as a summary of their own orders.
What changed: Both tiles and the spend-by-grant breakdown are computed from your entries, and adding, editing and deleting reagents works.
2026-09-21 — Cell Imaging & Quantification: Public pages quoted out-of-date accuracy figures and presented our deep-learning model as ahead of Cellpose
A statement on the site was wrong
What was wrong: The landing page, the user guide, the deep-learning control and the note at the top of the imaging tool quoted the built-in engine at F1 12.0% on H&E tissue and 60-64% on fluorescence, and presented the tool's own deep-learning model as ahead of Cellpose (67.2% against 63.2%). The engine figures were measured before the 19 August watershed fix. The comparison with Cellpose is an August measurement that has not been re-run.
Who was affected: Visitors and users deciding whether to use the built-in engine, the deep-learning option or Cellpose on H&E tissue, or how far to trust the engine on fluorescence.
What changed: The figures were replaced with the 21 September re-measurement: on two samples of 62 NuInsSeg H&E patches Cellpose scored F1 65% and 69%, StarDist 45% and 52%, and the built-in engine 23% and 26%; on fluorescence the built-in engine scores 87% (BBBC039), 80% (BBBC020) and 71% (BBBC038). The tool now recommends Cellpose for H&E tissue you plan to publish: at the top of the imaging tool, in the user guide, on the landing page, in the deep-learning control and in the upload summary when H&E is recognised. The claim that our model is ahead of Cellpose was removed, because it rests on the August measurement.
What to do: If you used the built-in engine on H&E tissue because of the published figures, re-check those images with Cellpose, and check the overlay before quoting counts.
2026-09-21 — Cell Imaging & Quantification: The default segmentation setting was too permissive on most image types; it has been changed
A result could have been wrong
What was wrong: The classical (non-deep-learning) segmentation grows each detected nucleus outward to a low threshold placed halfway (0.50) between the background and the Otsu level. That value was chosen in August to correct nuclei that came out about 38% too small, which was largely a side effect of the watershed defect fixed on 19 August. On the current code it scored below a plain single threshold on four of five benchmark sets, mostly by merging nuclei and by over-detecting on H&E tissue.
Who was affected: Object outlines, counts and areas from the classical pipeline before this date, on macrophage, H&E and mixed images. On evenly stained fluorescence nuclei (BBBC039) the change makes no difference. On the benchmarks the object count moved by 1 to 4% on most sets and by about 11% on TNBC (H&E: 9,666 objects before, 8,556 after, against 4,028 true).
What changed: The low threshold was raised from 0.50 to 0.85. It was chosen on BBBC039 and BBBC038 alone by a protocol fixed before the run, then confirmed on four sets that took no part in choosing it, including a second NuInsSeg sample of different patches fetched afterwards: F1 rose on every one (BBBC020 78.3% to 79.9%, TNBC 29.5% to 35.1%, NuInsSeg 21.2% to 22.8% and 22.0% to 25.5%; mean +3.1 points). A plain single threshold still scores about as well overall on the held-out sets, so this corrects a value that was too low; it does not show that the two-level mask is better than a single threshold. The deep-learning path and the growth cap are unchanged.
What to do: If you counted or measured nuclei on H&E, macrophage or mixed images with the default settings before this date, re-run them and check the overlay. H&E counts remain a screen, not a measurement.
2026-09-21 — Cell Imaging & Quantification: The H&E accuracy figure across organs was out of date, and the comparison with deep-learning tools overstated the gap
A statement on the site was wrong
What was wrong: The tool's note gave F1 12.0% for H&E tissue across 31 organs (NuInsSeg) and described a 53-point gap to Cellpose. That was measured on the code before the 19 August watershed fix. The note also did not say that the tool's automatic stain detection recognised only a small share of those patches as histology.
Who was affected: Anyone who read the note to judge the tool on H&E tissue, or to decide between it and Cellpose or StarDist.
What changed: Re-run on the same 62 patches (3,248 nuclei, 31 organs): F1 21.2% at the shipped settings (precision 15.6%, recall 32.9%, count 2.1 times too high), 22.7% with the plain single threshold; by organ from 64.4% (cerebellum) to 0.0%. Automatic detection recognised 9 of the 62 as histology. The scoring was cross-checked with an independent scorer, and reverting the single watershed line reproduces the old 12.0% exactly. The gap to Cellpose (64.9%, not re-run) is now about 44 points. Nothing in the tool's settings was changed.
What to do: On H&E tissue treat counts as a rough screen, and use a dedicated histology model (Cellpose, StarDist, HoVer-Net) for anything you intend to publish.
2026-09-21 — Cell Imaging & Quantification: The macrophage and H&E accuracy figures were out of date, and the two-level mask does not help on three of four benchmarks
A statement on the site was wrong
What was wrong: The tool's note gave F1 62.9% on BBBC020 (mouse macrophage nuclei) and 24.6% / 20.5% on TNBC (H&E breast cancer), measured on the code before the 19 August watershed fix. The note also presented the two-level mask as an improvement and the 3-pixel growth cap as best on every set. The two TNBC figures were the same settings on different subsets (the first 25 patches and all 50), which the note did not say.
Who was affected: Anyone who read those figures, or the claim that the mask improves results, to judge the tool on macrophage or histology images.
What changed: All four benchmark sets have now been re-run on the current code. Shipped settings: BBBC039 87.0%, BBBC020 78.3%, BBBC038 67.8%, TNBC 29.5% (all 50 patches; 36.0% on the first 25). The plain single threshold scores higher than the shipped two-level mask on BBBC020 (81.8%), BBBC038 (70.2%) and TNBC (36.6%), and lower only on BBBC039 (85.7%); the problem the mask was added to fix (nuclei about 38% too small) no longer occurs without it. On H&E the tool finds 2.4 times too many objects, so counts there are a screen, not a measurement. Scores were cross-checked with an independent scorer, and reverting the single watershed line reproduces every old figure. Nothing in the tool's settings was changed; the note and the benchmarks now say what was measured.
What to do: On H&E tissue treat counts as a rough screen only. On any images, check the overlay before quoting counts or per-cell measurements.
2026-09-21 — Cell Imaging & Quantification: The "will it work on my images?" figure was out of date, and the note said the two-level mask helped
A statement on the site was wrong
What was wrong: The tool's note gave F1 60.0% on the BBBC038 collection (30-plus experiments, mixed stains and magnifications) as the generalisation figure, and said the two-level mask improved it from 58.3%. Both were measured on the code before the 19 August watershed fix. Re-measured on the current code, the shipped settings score higher than 60.0% but the two-level mask no longer helps on this set.
Who was affected: Anyone who read that figure to judge how the tool would do on images unlike a single tidy fluorescence set.
What changed: Re-run across the whole collection (670 fields, 29,461 nuclei): F1 67.8% at the shipped settings (precision 77.0%, recall 60.6%; 3.1% missed, 0.9% split, 8.2% merged; the count is 21% low). The plain single threshold scores 70.2% there, so on this heterogeneous set the two-level mask lowers F1 by about 2 points, while on the single-preparation BBBC039 it raises it by 1.3. Scores were cross-checked with an independent scorer. The note states all of this.
What to do: On images that are not evenly stained fluorescence nuclei, expect merged nuclei and a count that can run roughly a fifth low across mixed image types. Check the overlay before quoting counts.
2026-09-21 — Cell Imaging & Quantification: The published accuracy figure was out of date and understated
A statement on the site was wrong
What was wrong: The tool's note quoted F1 63.8% on the BBBC039 nuclei benchmark. That was measured on the code as of 17 August. A watershed defect fixed on 19 August (the flood stalled wherever the distance map rose, dropping about two-thirds of each rotated or elongated nucleus) had been lowering it. The note was not updated, and the benchmark scripts that produced it were not in the project.
Who was affected: Anyone who read the note to judge how well the tool finds nuclei. The figure understated it; the defect itself was fixed on 19 August.
What changed: The benchmark scripts are now in the project. Re-run on 21 September across all 200 BBBC039 fields (19,389 nuclei), the segmentation stage scores F1 87.0% at IoU 0.5 (89.9% on the 20-field subset behind the old figure), cross-checked with an independent scorer. The other datasets quoted in the note (BBBC020, BBBC038, TNBC, NuInsSeg) predate the fix and are marked as out of date until they are re-run.
What to do: Check the overlay on your own images before quoting counts or per-cell measurements; the benchmark measures one fluorescence dataset, not your preparation.
2026-09-21 — Kaplan-Meier Survival Curves: Log-rank p-values were wrong
A result could have been wrong
What was wrong: The tool converted the log-rank statistic to a p-value with a normal approximation that is poor at one degree of freedom, and then doubled the tail probability. A chi-squared upper tail already covers both directions, so the doubling was an error on top of the approximation.
Who was affected: Every log-rank p-value the tool displayed or exported before this date. Weak and moderate effects were reported as less significant than they are (a statistic of 3.84 gave p = 0.077 instead of 0.05; a small worked example gave 0.913 instead of 0.433). Strong effects were reported as more significant (a statistic of 10.83 gave about 0.0003 instead of 0.001). Anything below 0.0001 was floored at 0.0001.
What changed: The p-value is now the exact chi-squared tail for one degree of freedom, using the same function that the statistics tool checks against published critical values. Two hand-worked examples are now pinned by tests. The Kaplan-Meier curve itself, the risk table and median survival were not affected.
What to do: Re-run any survival comparison whose p-value you reported or whose conclusion depended on it. If a result crossed 0.05 in either direction, it may change.
2026-09-21 — Site: Verification notes cited tests that were not in the project
A statement on the site was wrong
What was wrong: Notes on several tools said their outputs were "asserted in npm test against published reference values". A review found that many of those tests were not in the repository: the original suite did not survive a platform migration, and the notes kept describing it. Several tools were graded 9 or 10 with no automated check on their own code, and the built-in flow-cytometry spillover values were described as "published" without a source.
Who was affected: Anyone who read a tool's grade or note as evidence that a specific reference check existed.
What changed: The missing checks that could be rebuilt were rebuilt, this time from published values or hand calculation written before the code was run (the "verified against" table below lists every check and the test behind it). Eight tools moved from "verified" to "standard method", and one from 10 to 9, where no check on the tool's own code exists; plain-arithmetic calculators without a test are shown as such. Notes that described tests which do not exist now say so. A new test fails if a tool is graded "verified", or a note says "asserted", without a real test behind it.
2026-09-21 — Site: Pricing and billing statements did not match what exists
A statement on the site was wrong
What was wrong: The FAQ quoted annual prices for plans that no longer exist, the landing page comparison showed a "$79/mo" plan, and the comparison-page cost calculator used $499 a year. The site also claimed a "14-domain automated verification suite", SSO and a dedicated account manager, and wire and ACH billing. None of those exist.
Who was affected: Visitors comparing prices, and institutions asking about procurement.
What changed: The one paid plan (Lab, $49/month or $468/year, per lab) is stated everywhere, and the unsupported claims were removed. Card payment through Stripe is the only billing offered today; purchase orders are handled case by case.
2026-09-21 — Site: Four references pointed to the wrong or a non-existent DOI; the MSA tool was labelled "Clustal"
A statement on the site was wrong
What was wrong: The DOIs for Thompson 1994 (CLUSTAL W), Vincent & Soille 1991 (watersheds) and Sebaugh 2011 (IC50 guidelines) did not resolve, and a citation of 21 CFR Part 11 was shown as if it had a DOI. Separately, the multiple sequence alignment tool was labelled "Clustal-W" or "Clustal Omega" although it runs Needleman-Wunsch alignment with a UPGMA guide tree.
Who was affected: Anyone who followed those reference links, or who described the alignment as Clustal in a methods section.
What changed: The DOIs were corrected and checked against doi.org; the tool now describes what it actually runs. Every DOI in the source is now tested for form, and DOIs are links. The tool still uses a "Clustal-style" conservation notation (* : .) and says so.
What to do: If you cited the MSA tool as Clustal, describe it as progressive Needleman-Wunsch alignment with a UPGMA guide tree, or use Clustal Omega or MAFFT for deep alignments.
2026-09-21 — All tools: Text could not be selected or copied
Usability, no effect on results
What was wrong: A page-wide setting that prevents text selection was applied to the whole app, so no result, table cell or computed value could be selected or copied.
Who was affected: Everyone. No numbers were changed.
What changed: Selection is enabled everywhere except the toolbars and image viewers, where dragging is the interaction. A test prevents it coming back.
2026-09-16 — Sample Size & Power Calculator: Proportion and survival designs used the t-test formula
A result could have been wrong
What was wrong: The Test Type picker offered Proportions (chi-square) and Survival (log-rank), but both ran the same two-sample t-test formula for a standardised mean difference. A proportion or a hazard ratio has no such relationship.
Who was affected: Any sample size or power figure produced for a proportion or survival design.
What changed: Each design now uses its own formula (Fleiss two-proportion z-test; Schoenfeld log-rank events). The t-test sample sizes were also checked against G*Power (d = 0.2 / 0.5 / 0.8 / 1.0 need 394 / 64 / 26 / 17 per group).
What to do: Recompute the sample size for any proportion or survival study you planned with this tool.
2026-09-16 — Standard Curve Engine: qPCR standards were blank-subtracted by default
A result could have been wrong
What was wrong: "Subtract zero-standard blank" was on by default and was not reset when the qPCR preset was chosen. qPCR standards are Ct values, which fall as concentration rises, so the lowest-concentration standard was subtracted from every point, driving later values negative and floored at zero.
Who was affected: A standard curve built from the qPCR preset with the box left checked reported R² = 0, slope 0 and an undefined efficiency.
What changed: Blank subtraction is off and disabled for qPCR standards, and the blank is taken from the standard whose concentration is actually zero.
What to do: If you built a qPCR standard curve with the qPCR preset and the blank box checked, rebuild it and check that R² and the slope are sensible before using the efficiency.
2026-09-16 — PCR Primer Designer: Pasting a new template did not recompute primers
A result could have been wrong
What was wrong: Primer candidates were computed once for the bundled example and not again when a real template was pasted over it.
Who was affected: Primers shown after pasting a template could belong to the example sequence rather than yours.
What changed: Candidates recompute whenever the template changes. The example template is now labelled as an example.
What to do: If you ordered primers from this tool by pasting a template over the default, check the primer sequences against your template.
2026-09-16 — Several exports: CSV files were cut off at the first "#"
A result could have been wrong
What was wrong: Eight CSV exports built the file as a web address, which silently truncates the file at the first # character: the primer order sheet, both CRISPR exports (guide order sheet and indel call), the plate-reader dose-response and 96-well plate exports, the master-mix export, and two administrator-only exports. (The blot export had the same fault and is listed under Western Blot.)
Who was affected: Any of those files that contained a "#" — in a compound name, a comment row or a note. Everything after it was missing, with no warning.
What changed: All exports now use the shared downloader, which encodes the file correctly and also escapes commas and quotes.
What to do: If you exported a CSV from these tools and a name or row looked short, export it again.
2026-09-16 — Wound Healing: Opened with a fixed demo and a typed-in p-value
Showed fixed or invented output as analysis
What was wrong: The tab showed a fixed demo dataset that never changed, with a p-value of 0.0004 typed into the sample data and displayed as if calculated.
Who was affected: Anyone who read the on-screen result as an analysis of their own image.
What changed: Rebuilt on real segmentation of the uploaded scratch image. The example is labelled as an example, and no p-value is shown that was not computed from measured replicates.
2026-09-16 — 3D Spheroid Morphometry: Opened with a fixed demo and a typed-in p-value
Showed fixed or invented output as analysis
What was wrong: Three hand-typed conditions with a typed-in p-value of 0.0002, unaffected by anything uploaded. The volume and sphericity formulas were sound but only ever received hand-typed axis lengths.
Who was affected: Anyone who read the on-screen result as an analysis of their own images.
What changed: The spheroid is now detected in the uploaded image and measured; the example is labelled.
2026-09-16 — Subcellular Puncta & Foci Counter: Foci counts were random numbers
Showed fixed or invented output as analysis
What was wrong: The tab opened with three conditions of Math.random()-generated foci counts and did not respond to uploaded images.
Who was affected: Anyone who read the counts as measured.
What changed: Nuclei and spots are now detected in the uploaded channels; the example is labelled.
2026-09-16 — Neurite Outgrowth & Angiogenesis: Skeleton metrics were hand-typed
Showed fixed or invented output as analysis
What was wrong: Total length, junctions, endpoints and a Sholl curve shaped like a Gaussian were typed in and never changed with the uploaded image.
Who was affected: Anyone who read the metrics as measured.
What changed: The image is skeletonised (Zhang-Suen thinning) and measured; the example is labelled.
2026-09-16 — 2D Scatter Gating: The gated population was regenerated at random on every drag
Showed fixed or invented output as analysis
What was wrong: The tab generated 240 entirely synthetic cells every time it was drawn, including on every gate-slider movement, so moving a gate created a new population instead of re-gating one.
Who was affected: Anyone who read gate percentages as describing their own cells.
What changed: Gates now read the real per-cell measurements already made in the imaging tool.
2026-09-16 — SuperPlots: Replicates were invented from cell ID parity
A result could have been wrong
What was wrong: When no condition label said "control", the tool split one image's cells into fake Control and Treated groups by even and odd cell ID, cut those into thirds and called the thirds biological replicates. The p-value came from a lookup table, not a test.
Who was affected: Any SuperPlot built from a single image, and its p-value.
What changed: Groups come only from real condition labels on separately imaged samples, and a p-value is shown only when every shown condition has at least two real replicates. Otherwise it says why it cannot.
What to do: Do not use a SuperPlot or p-value from before this date as evidence of a biological effect unless it was built from separate samples.
2026-09-16 — Site: Unknown tool addresses returned "found"; two tools did not label their example data
Usability, no effect on results
What was wrong: A mistyped or removed /tools/ address returned a normal page (and opened whichever tool loaded first) instead of a "not found". Separately, the Kaplan-Meier tool and the Primer Designer opened on example data and computed results from it without saying so.
Who was affected: Visitors following old links; search engines indexing non-existent pages.
What changed: Unknown addresses now return not-found, and both tools show the example-data notice until you enter your own.
2026-09-15 — Cell Imaging & Quantification: The multi-image comparison had a silent control and a fabricated ANOVA p-value
A result could have been wrong
What was wrong: Whichever image was loaded first was silently the control, with no way to change it. The one-way ANOVA p-value was read from a hand-picked table of F thresholds that ignored degrees of freedom, and starred results on that number. The help text promised Dunnett testing that was never implemented.
Who was affected: Significance stars and p-values in the "Impact vs Ctrl" comparison before this date.
What changed: You choose the control; the ANOVA p-value is the real F-distribution value; pairwise testing is Tukey-Kramer, as the help now says.
What to do: Re-check any multi-image comparison you reported, especially where the first image was not the control.
2026-09-15 — Western Blot: Six defects in blot quantification
A result could have been wrong
What was wrong: Significance stars were placed on the control lane (a missing p-value compared as less than 0.05). A lane with no loading-control band produced a fold change around 4.5 trillion. Lane IDs could duplicate after removing and adding lanes. Excluded lanes still fed later steps. Molecular weight was measured in the wrong coordinate frame when the band window changed. The exported CSV was cut at the first "#".
Who was affected: Fold changes, stars, molecular-weight estimates and exports from the blot tool before this date.
What changed: A lane without a control band shows "n/a", stars appear only with a real p-value, excluded lanes stay excluded, lane IDs are unique, molecular weight is measured against the ladder, and the export uses the shared downloader. You can also choose the control lane and band order, and error bars show SD, SEM or 95% CI with Holm-adjusted comparisons.
What to do: Re-run any blot quantification you reported, especially where a lane lacked a loading-control band.
2026-09-07 — Methods Generator: The prompt allowed invented experimental values
A result could have been wrong
What was wrong: The AI prompt asked the model to add placeholders "only where critical lab-specific catalogue details are missing", which licensed it to invent centrifuge speeds, incubation times and dilutions for everything else.
Who was affected: Any Methods text generated before this date could contain a plausible-looking value you never supplied.
What changed: The prompt now says to use only the parameters you provided, to mark anything not supplied with a bracketed placeholder, never to substitute a "typical" value, and not to add citations or claim compliance with a reporting standard.
What to do: Re-read any generated Methods paragraph against your own records, sentence by sentence.
2026-09-05 — Drug Synergy Calculator: The heading claimed Loewe and Combination Index; only Bliss is computed
A statement on the site was wrong
What was wrong: The tool was titled "Drug Synergy & Combination Index Suite (Bliss & Loewe)" but computes the Bliss independence score only. Loewe values appeared only as reference numbers.
Who was affected: Anyone who described a Bliss score from this tool as Loewe or Chou-Talalay.
What changed: The heading and FAQ state that it computes Bliss.
2026-09-05 — qPCR ΔΔCt, Statistical Tests, Drug Synergy: Example data produced real-looking results with no label
Showed fixed or invented output as analysis
What was wrong: These tools open on a worked example and calculate fold changes, p-values and synergy scores from it, with nothing on screen saying the data were an example.
Who was affected: Anyone who screenshotted or exported before entering their own data.
What changed: An "example data" notice appears until the first real edit.
2026-09-03 — Cell Imaging & Quantification: Analysis ran with settings from before the image was profiled
A result could have been wrong
What was wrong: The analysis closed over a stale copy of the image profile, so it ran with pre-profiling channel and size settings and hid the zero-result explanation.
Who was affected: On the test image, the same clicks gave 0 objects before the fix and 1,328 after. Counts from before this date may be too low.
What changed: The analysis uses the current profile.
What to do: Re-run any counts you reported, and check the overlay.
2026-09-03 — Cell Imaging & Quantification: The toolbar Upload button silently failed on TIFF files
Usability, no effect on results
What was wrong: The cyan Upload button on the image toolbar read files in a way no browser can decode as TIFF, and had no error handler, so a .tif produced no image and no message.
Who was affected: Anyone who uploaded TIFFs with that button. The separate "Open File" button worked.
What changed: Both buttons use the same loader, and failures say what went wrong.
2026-09-02 — Cell Imaging & Quantification: Missing calibration was assumed; a channel with no signal was segmented
A result could have been wrong
What was wrong: When a file carried no real scale, the tool assumed 0.325 µm per pixel without reading the scale bar burned into the image; on one test file the true value was 0.180, so areas were off by 3.25 times. Separately, segmentation defaulted to the blue channel even on images with no nuclear stain, giving zero objects.
Who was affected: Areas and object counts on images without embedded calibration, or without a nuclear channel.
What changed: The burned-in scale bar is detected and used, with a warning where the scale is assumed; the channel is chosen from a profile of the image and the choice is stated.
What to do: Check the scale on any image whose file had no embedded calibration; re-measure areas if the tool had assumed one.
2026-09-02 — Site: The Terms listed Syria as comprehensively sanctioned
A statement on the site was wrong
What was wrong: The sanctions section of the Terms named Syria as a comprehensively sanctioned country. US sanctions on Syria under those orders were revoked in 2025.
Who was affected: Visitors reading the Terms.
What changed: Syria was removed from the Terms and from the access block list, with a test asserting it stays out.
2026-08-31 — Western Blot: Band detection framed the wrong region on photos with a dark surround
A result could have been wrong
What was wrong: The tool judged whether bands were dark-on-light from the median of the whole image, so a photo with a dark surround was read as bright-on-dark and the blown-out white block became the strongest "signal". Separately, the minimum band height was a fixed constant no caller could change, so faint blots reported no bands.
Who was affected: Lane detection and densities on blots photographed with a dark surround, and faint blots reported as empty.
What changed: Polarity is measured inside the region you draw, with a notice when it changes; the threshold scales with the image's own noise and is deliberately conservative so a blank lane does not produce invented bands.
What to do: Re-run any blot whose lanes looked wrong or whose bands were reported as missing.