quickTrace — user guide

What quickTrace is

quickTrace recovers the real numbers behind a chart you can no longer edit — a chart flattened into a picture inside a client deck, a screenshot, a PDF export. Instead of guessing values by eye and retyping them, you load the image, the tool learns the axis scale, measures the plotted shapes against it, and hands back an editable table, a CSV, a rebuilt SVG and an Excel file with a live chart.

It gives you one of two honest outcomes, and never blurs them:

• Recover data — numbers first. Exact when they're still inside a PowerPoint file, measured from pixels when they're not.
• Recreate appearance — artwork first. An editable vector reproduction of what's on screen, claiming no numbers at all.

The design principle behind everything: a plausible wrong number is more dangerous than no number. When the tool isn't confident, it refuses and says why, rather than handing you a value that merely looks trustworthy. A refusal is the tool working, not failing — see When the tool refuses.

New to the tool, or showing it to someone? Press ▸ Guided demo in the header. It walks the real interface in ten steps, and every figure it quotes is read live from the running tool — nothing staged.

Quick start: your first chart

Fastest way to see the whole loop — about a minute, on a sample whose true values are known so you can check the tool's honesty at the end:

1. In step 1 of the workflow rail, press Sample: bar chart.
2. The rail advances on its own. In step 3, press Read the axis automatically — it finds the gridlines, reads the numbers beside them, and reports the recovered scale with a trust score.
3. In step 5, press Read the values.

The right-hand panel now shows every bar's value in an editable table, each with a provenance dot, plus a Self-check against known truth comparing what was measured against the values the sample was drawn from. From here, Download CSV or Download Excel + live chart.

On your own charts the flow is the same, with two extra clicks when needed: drag the plot area (step 2) if the automatic guess is off, and pick series colours (step 4) if the suggestions aren't right. Details in Measure a chart image.

Exact data from PowerPoint — no measuring at all

If the chart is still a native chart object inside a .pptx, its real numbers never left the file. quickTrace finds them and lifts them out exactly — no calibration, no measurement, no error.

1. Step 1 → Open PNG / JPG / PPTX… and choose the deck.
2. Each native chart appears at the top of the rail: “Slide 4 · native chart 1 found with its embedded workbook. The real numbers are still inside this deck.”
3. Press Extract data — no measuring. The table fills with values marked exact.

If a chart's cached values were stripped, the tool automatically reads the workbook embedded alongside it instead. And every ordinary image in the deck is listed below the native charts at original resolution — click one to load it for measurement, exactly as if you'd opened the image file directly.

Measure a chart image

The workflow rail on the left is an accordion: one step open at a time, advancing as you finish each. You can click any step header to go back; the tool won't fight you.

Step 1 — Load image

Open PNG / JPG / PPTX… accepts PNG, JPG, GIF, BMP, WEBP and PPTX. Images over 6000 px on a side or 16 megapixels are refused up front rather than allowed to hang the tab. Two built-in samples with known true values are there whenever you want to sanity-check the tool itself.

Step 2 — Plot area

Drag a box round the plot — just the plotting region, excluding the title, legend and axis text. This tells detection where to look and what to ignore. If bars or gridlines are being missed, the first thing to re-check is this box.

Step 3 — Calibrate Y

Read the axis automatically finds horizontal gridlines, reads the numbers beside them, and fits a scale. The fit note reports the recovered range and a trust score — a composite of how many ticks were read, how confidently, and how well they agree. Every tick appears in a review list where you can correct a misread value, exclude a bad tick, and Refit selected ticks.

When automation can't get there — no gridlines, unusual labels — calibrate by hand: Click low gridline, type its value, Click high gridline, type its value. Two known points are all the scale needs. The magnifier loupe follows your cursor showing the live calibrated value, so you can verify the scale before measuring anything.

Step 4 — Series

The tool suggests the most common saturated colours in the plot as one-click series candidates. For anything else: Add a series… then click a bar or line of that series on the image. Each series has a colour tolerance slider — Preview selected colour mask paints exactly which pixels will be counted before you commit, which is the honest way to tune it.

Set Type to vertical bars or 2D line. For multi-series bars the layout is detected automatically, or force Grouped / Stacked if it guesses wrong.

Step 5 — Measure

Read the values measures every bar height or line position against the calibrated scale and fills the table. Category labels under the plot are read too — and if two reads in an ordered run are confident (say Jan and Mar), the whole month sequence is pinned from them. Values stay editable afterwards.

This is also where the tool may refuse instead of answering — that's deliberate, and it tells you exactly why. See When the tool refuses.

Step 6 — Recreate — is not part of measuring. It's the other outcome: reproducing the chart's appearance when measurement doesn't apply. It has its own section: Trace vs Reconstruct.

Read the results

Every value in the Recovered data table is editable — click a cell and type. Each cell carries a small dot naming where its value came from. Hover any cell and the tooltip spells it out.

Below the table, a note counts any flagged cells: “unusual bar width, a probable series crossing, or no pixels found — check these against the image before trusting them.” Take it literally; those are the cells the tool is least sure of.

Verify before you export

• Reconstruction overlay redraws the recovered numbers back onto the source image, so drift is visible instead of assumed. If a redrawn bar doesn't sit on its original, you can see exactly which value to fix.
• On the built-in samples, Self-check against known truth reports the worst and mean error against the values the sample was drawn from — the tool grading itself in the open.
• Fit / 50% / 100% / 200% zoom the canvas. Zoom in before clicking a gridline by hand — on a large screenshot a click that looks accurate at Fit can be several values out.
• Undo / Redo cover the whole reconstruction state, so no experiment is destructive.

When the tool refuses — and what to do

quickTrace refuses to report numbers it can't stand behind. Each refusal names its reason and its remedy. The three you'll meet:

“The automatic axis trust score is below the 78% floor”

The axis reading wasn't confident enough — misread ticks, too few ticks, or labels that don't agree. In order of preference:

1. Open step 3's tick review: correct or exclude the misread ticks and Refit selected ticks.
2. Calibrate by hand — two clicked gridlines with typed values beat a shaky automatic read every time.
3. Override and read anyway — only after you have personally verified the scale with the loupe. The override exists for charts you've checked, not for saving time.

“These shapes don't stand on the value-axis baseline”

Real bars stand on the axis baseline; candle bodies and box plots float. Measuring a floating shape would return its position, not its value — a number that looks right and means nothing. Press Trace the artwork instead to recover the chart as editable artwork (next section).

“No pixels matched”

The series colour and tolerance matched nothing inside the plot box. Widen the tolerance slider, re-pick the colour from the actual bar or line, and use Preview selected colour mask to see what will be counted before measuring again.

Recreate the look: Trace appearance vs Reconstruct chart

For the many charts measurement doesn't cover — pies, radars, sankeys, gauges, candlesticks, and anything else — step 6, Recreate, reproduces the chart as editable vector artwork. It claims no numbers; it recovers the picture, as real shapes and, where text is read confidently, real editable text.

The mode toggle (in step 6 and beside the exports — same control, mirrored) chooses what the output is allowed to claim:

Then press Trace the artwork (or Trace artwork — no values beside the exports). Tracing runs at the image's native resolution; the largest screenshots take a few seconds.

Start with Trace appearance when you just need faithful editable artwork. Use Reconstruct chart when you want cleaner structure — one true ring instead of arc fragments — and the tool's reading of what the chart is.

Traced outlines are real curves, not many-sided polygons following the pixel staircase, and a run that genuinely lies on a circle is emitted as a true arc. That matters the moment you open the file in Illustrator or Inkscape: a curve has a handful of editable control points, where a traced staircase would have hundreds and would look jagged at any zoom.

Rotated text is currently reproduced as drawn outlines, not editable text — it looks right but can't be retyped. Horizontal text the engine reads confidently comes back as real text.

The verdict panel, and what the provenance words mean

Running Reconstruct chart opens the Reconstruction verdict panel beside the preview. It reads the chart back to you in plain words and — this is the point — separates three kinds of claim, so you always know what was seen, what was guessed, and what you vouched for. That distinction is written into the exported file itself.

What the panel shows

• The family line — what kind of chart this is, with confidence: “Radar chart detected — confidence 0.84”. Combination charts get a layered reading (bar + line). When the evidence doesn't single out a family it says so honestly: “No single family claimed — the safe reading”, with the closest candidates named.
• The family selector — your override. If you know what the chart is and the reader doesn't, pick from Bar chart, Line chart, Scatter plot, Pie / doughnut, Radar chart, Box / candlestick — or No single family, which is how you take an override back off. Your choice is recorded as operator-selected — the panel then reads “operator-selected (reader said: …)” — and it draws nothing by itself.
• The structure inventory — every observed structure counted (“6 spokes found”), and every open proposal listed with its location and evidence: “Track at radius 312 — support 78%”. Support is how much of the predicted structure's path has real pixel evidence behind it.
• The verdict note — what the machine decided and why, in its own words: “Restored 4 observed structures automatically”, or the exact reason nothing was restored.

Confirming an inferred structure

1. Anything inferred is drawn dashed amber over the original image, so you judge the hypothesis against the source itself — not against a redraw.
2. Press Accept on what's right, Dismiss on what's not. Both are toggles; nothing is final yet, and dismissed proposals simply stay out of the file.
3. Press Apply confirmed structures. Accepted structures are drawn into the reconstruction, each labelled user-confirmed, and the note confirms: “Added 1 user-confirmed structure to the file.”

Your accept supplies the judgment that the structure belongs; the machine still enforces its own safety line. If drawing your accepted structure would damage the picture — cover real ink, lose detail the trace already had — the tool refuses and quotes the exact reason in the panel, for example: “rejected: fidelity fell 0.6% — user-confirmed furniture must not cost the picture.” A person decides what's worth adding; the machine guarantees the addition can't quietly cost you what you already had.

Applying again later rebuilds from the original reconstruction rather than stacking on top of the last apply — accepting, dismissing and re-applying is always safe to redo.

What lands in the exported file

In the downloaded reconstruction SVG, every restored structure carries its provenance as a machine-readable label: restored observed structures sit in a reconstructed-furniture group, your confirmed ones in a separate confirmed-furniture group marked data-provenance="user-confirmed". Delete both groups in any editor and what remains is exactly the plain appearance trace, byte for byte. Anyone downstream can see — and strip — precisely what was reconstructed versus what was traced.

Exports and saving

The rebuilt-SVG / reconstruction-SVG distinction is the tool's core promise in miniature: one is recovered data drawn honestly, the other is recovered appearance claiming nothing — and neither ever pretends to be the other.

What's supported — and what's refused by design

Measurement (numbers out)

• Vertical bar charts — single, grouped and stacked, including negative bars and truncated axes.
• 2D line charts — multiple series separated by colour, gaps preserved.
• Native PowerPoint charts — any type, because nothing is measured: the data is lifted exactly from the file.
• Linear and logarithmic value axes, detected automatically.

Refused for measurement, deliberately

These produce confident nonsense under colour-based measurement, so the tool declines rather than guessing:

• Horizontal bars, pie/doughnut, dual-axis, combo and 3D charts
• Candlesticks and box plots — caught by the baseline check even when their axis reads perfectly
• Series sharing one colour that cross each other
• Broken (discontinuous) value axes — surface as low calibration trust

Every refused chart still has a full answer in the appearance lane — trace or reconstruct it as editable artwork. And when a category label can't be read confidently, it's left as a placeholder to fill in, never guessed.

Privacy and confidentiality

Nothing you load ever leaves your machine. Every byte of processing — image analysis, text reading, measurement, export — happens inside your browser tab. There is no upload, no account, and no server that ever sees a chart.

This is enforced, not promised: the page is served with a security policy that makes it impossible for it to send data anywhere, and the deployment's test suite verifies on every release that no request leaves your machine once the page is loaded. Charts from confidential client decks are safe here — that constraint shaped the whole tool.

The one nuance worth knowing: the tool is served from the team's own box at trace.wlteamtools.com (including its text-reading engine — also self-hosted, never a third-party service), and then processes everything locally. Being online is how it arrives, not how it works.

Troubleshooting

Common questions

Why does it refuse instead of just giving me its best guess?

Because its best guess would look exactly like a solid answer. A wrong number you were warned about is a nuisance; a wrong number presented as trustworthy ends up in a client document. Every guard in the tool exists to keep the second kind at zero — and every guard has a deliberate override for the case where you've verified the chart yourself.

What does “support 78%” mean on a proposal?

The share of the predicted structure's path that has real pixel evidence under it. High support with an inferred chip means the structure is likely but the tool held itself to its evidence standard — which is exactly the question it hands to you.

I accepted a proposal — why was it still rejected?

Your accept answers “does this belong?”. The machine still separately checks “does drawing it damage the picture?” — covering real ink, losing detail, adding clutter. The refusal reason is quoted verbatim in the panel. This is a feature: nothing you confirm can silently cost you what the trace already had.

Can another tool tell what was reconstructed versus traced?

Yes. Every restored structure in the SVG carries a machine-readable provenance label, grouped so it can be stripped in one action — removing the restored and confirmed groups leaves the plain literal trace, exactly.

Is the measurement accurate?

On the built-in samples — where true values are known — worst-case bar error is well under 1%, and the tool shows you its own self-check every time. On real charts, accuracy depends on calibration; that's why the trust score, the tick review and the reconstruction overlay exist. Verify with the overlay before exporting anything that matters.

Who do I contact about the tool?

The Williams Lea team tools owner. This guide lives beside the tool itself and is updated with it.