TraceLab

Analysis software is good either side of the moment that matters

Tools for qualitative research collect well and present well. The part between — where a sentence somebody said becomes a claim a team acts on — is where the provenance is either kept or quietly lost.

TraceLab's slice is built entirely around that middle. Everything in it exists to keep a claim tied to the excerpt it came from, and to make the tie legible to somebody who did not do the coding. Three commitments were fixed before any screen was drawn, and every later decision in this file follows from one of them.

1. i

   A machine may propose. It may not assert. No confidence percentage appears anywhere in the product; the machine's output is named, provisional, and unusable as evidence until a person confirms it.

2. ii

   Contradiction is data. Two excerpts that disagree are both kept. Publishing an unresolved disagreement is a first-class outcome, not a failure to tidy up.

3. iii

   A quote is only evidence in the company of what surrounded it. Citations open inside the turn they were taken from, one at a time, with a route back to the source.

Scope was fixed before pixels too: six views, three edge states, no marketing surface. The public build exposes no states menu, no help overlay, no reduced-motion toggle, no debug panel, no inspector and no flow-step list — reduced motion is honoured through prefers-reduced-motion alone, because this concept has no genuine product need for a switch. That is asserted at runtime on every view, not only in the source.

The one study the prototype ships

Corveth Research, a four-person studio (fictional). Project Workshop Handover — how two-shift repair workshops pass an unfinished job across the change of shift. 14 sources S-01 … S-14, one session transcript each; six codes; six themes; three insights; two named reviewers. Every person, session, quote and timecode is invented, and no outcome metric is claimed anywhere in the product. The counts are the structure of the fiction, not results.

The machine may propose. It may not assert

Machine help appears exactly twice in six views, as two named, bounded operations: Suggest codes in the transcript, and Draft with evidence in the composer. There is no prompt box, no message list and no assistant turn anywhere in the product.

That is a decision about where the confidence lives. An assistant that talks invites the researcher to argue with it; an operation that returns a provisional object invites the researcher to inspect it. TraceLab does the second, and it holds four rules while doing it.

1 · No number stands in for judgement

The suggestion band is headed Needs your review · 4 suggested codes, and there is no percentage on the screen. The band's body says what was read and what the reader owes it: TraceLab read S-09 and proposed these codes. They are a starting point. Confirm each one before it becomes evidence. Its footer restates the duty rather than assuming the heading was read: Suggested codes are not evidence until you confirm them.

Each suggestion carries a per-item Rationale disclosure — a collapsible on the row, not one switch for the screen — and each rationale argues from the corpus rather than from itself: This sentence lists three physical places the card is kept and names no digital system. Twelve excerpts already coded paper-first describe the card's physical position. Confirming four suggestions moves the rail's count from 9 to 13 and stamps each new card Suggested by TraceLab · confirmed by NR just now, so the provenance of a machine-originated excerpt survives its promotion.

2 · A proposal is provisional until a person keeps it

On the Theme Canvas, five clusters are the researcher's own. The sixth — Whatever gets said at the door — was proposed by the machine, and it is drawn differently: a dashed border, a Suggested tag on its name, and a bar beneath it reading TraceLab grouped 3 excerpts it read as the same idea. Nothing is saved until you keep it. with Undo and Keep theme. Keeping it makes the border solid and leaves the attribution on the cluster header, permanently. The canvas is a canvas and not a kanban: clusters sit at free positions on a dotted field, with no columns, no lanes and no fixed order.

3 · The machine notices when its own input has gone stale

Suggestions were computed against a corpus that has since moved. Rather than serving them silently or invalidating them without asking, the band says Two excerpts changed since these were suggested. and puts Refresh suggestions on the same line. A warning whose remedy is elsewhere is a way of transferring blame; a warning that carries its own action is a state.

4 · A failure is scoped to the step that failed, and says what survived

Edge 1 is the drafting service being unreachable. It renders inside the draft step card and nowhere else: an oxblood cause line (TraceLab could not reach the drafting service.), then the sentence that actually matters — Your excerpts, themes and notes are saved. Nothing was lost. — then two real exits, Try drafting again and Write it myself, directly above the slot the draft would have filled. The footnote names what still works: The evidence rail still works. Citations can be added by hand from any excerpt.

The context bar, the title, the editor and all six evidence cards stay live behind it, and a test asserts that they do. The same discipline applies when the operation succeeds: the step card reports Step 2 of 3 · grouping claims with a Stop drafting control, because an operation a person cannot stop is an operation that is happening to them.

Two excerpts, opposite readings

The prototype's happy path is rigged to reach a disagreement. Claim 2 of the draft carries two excerpts that contradict each other, so reconciliation is exercised rather than skipped — the hardest screen in the product is on the main line, not on a branch.

Both excerpts are coded app-as-record. One opening lead says I do not open the app until lunch. The card tells me where to start. Another says The app is what I open first. The card is just what is left over from the night before, and half the time it is not filled in. They cannot both be summarised away, and the reconciliation card does not try: Decide what each one does to the claim. Nothing is deleted either way.

Each excerpt gets its own three-way answer — Supports this claim, Contradicts this claim, Not related to this claim — so the researcher's judgement is recorded per source rather than as a verdict on the pair. Saving keeps the loser: Counter-example · S-04 · 08:50 — one opening lead reads the app first. is published beneath the claim, and the nav badge clears.

Declining is a first-class outcome. Leave unresolved stamps the claim Unresolved · 2 excerpts disagree. This will publish with the claim., holds the navigation badge at 1, and writes the decision into the activity log with a name against it. Publishing a disagreement is a research position; a tool that only permits tidy claims is quietly asking for false ones.

The terminal action is not the machine's to take

Publish is rendered disabled and visible with Pending approval immediately beside it — the block is explained at the point of refusal, not in a tooltip. It unlocks only when both named reviewers have submitted an approval; one reviewer requesting changes does not unlock it, and a test asserts exactly that. The confirmation dialog then states what publishing means in the study's own terms: It becomes readable by everyone in Corveth Research, with its 6 citations and 1 counter-example.

A quote is not evidence on its own

A citation in a research insight is a promise that a real person said something in a real context. Reproducing the sentence keeps half of that promise. TraceLab keeps the other half.

Citations are inline typographic badges in the prose — never chat turns, never a footnote pile at the bottom. Activating one opens the rail on the quote inside its surrounding turn, under the label SHOWN IN CONTEXT: the cited sentence is tinted, and the sentences either side of it are shown at full strength, including the ones that complicate it. The footnote says why: This quote is shown inside the surrounding turn, so you can judge how it was taken.

Support arrives one quote at a time behind a 1 of 4 pager, rather than as a wall of eleven excerpts that nobody reads. The end of a claim's support is stated rather than hidden — at 4 of 4 the Next control is disabled and visible. And every quote offers Open S-09 in coding, which re-enters the transcript view with that excerpt focused. It is the audit path: from a published claim back to the passage, in one action, without leaving the product.

The published page keeps its own provenance: byline, Approved by Tomas Ekwall and June Bakhshi · version 1, a View history popover with two versions, and a source index listing each cited session with its excerpt count. Nothing about the page pretends the insight arrived without hands on it.

One trigger per transition, decided in advance

Every step below advances on a single named control that the product actually exposes. Nothing advances on a timer, a guess or a test hook — which is why the same list drives the end-to-end suite, and why a green run is evidence that a researcher can complete the arc rather than evidence that a script can.

Coverage is walked, not claimed

Every state is filed on one of two surfaces before it is built. Public states are the ones a researcher reaches by using the product. QA-only states — skeletons, a 240-turn session, a cluster holding 24 chips, a validation reset, the print stylesheet — live in Storybook and the test suite and never appear in the shipped bundle.

Eight of the 40 public states belong to the three edge states, because the edges are part of the research experience and are held to the same finish as the happy path. A permanent audit harness drives all 40 through product controls, photographs each one and runs six DOM detectors over it, at 1440 × 900 with six states repeated at 1280 × 800 and twelve at 768 × 1024: 58 renders, 40 / 40 reached, 0 unreachable.

overlap

0

overflow & clipping

0

truncation

0

touch target

0

contrast

0

dead space

9

The nine dead-space findings are held rather than hidden, each with its own measurement: four are empty or short data (a no-results search, a three-row tab, and S-12 twice), four are geometry the screen spec fixes — the 768 drawer is a drawer, not a column — and one is a three-claim document shorter than its window, 15 px over the limit. Not one is closable by composition, and every one of them is closable by fabrication. No threshold was moved and no selector was added to an exclusion list to reach the other five zeros.

The suites, and what each one is for

Accessibility, measured rather than asserted

axe-core ran over the six views and the three edge states at all three viewports — 27 scans — with the WCAG 2.0 / 2.1 / 2.2 A and AA tag sets, scoped to the product surface, with no disableRules() and no waiver list anywhere in the spec. Result: 0 violations of any impact, not merely zero serious and critical.

The scan waits for fonts painted, images decoded and the 160 ms route fade landed before it runs. That is a correctness fix rather than a relaxation: axe computes contrast from composited colour, and a scan taken mid-fade reported a serious color-contrast on a frame nobody sees.

Contrast, read off composited pixels

The audit measures contrast on the pixels as painted — 280 runs, 0 indeterminate — which is how it caught the one finding a DOM walk cannot. text-muted clears AA on paper at 5.02:1, but the product tints the current navigation item and the open source row with 6 % blue, and tints the failure card with 6 % oxblood. Composited, the same colour reads 4.2:1 and 4.49:1.

The palette did not change. Three runs took a darker ink; blue, oxblood and paper kept every tint, chip, border and accent role they already held. A brand colour that fails in one composite is not a brand problem, it is a composite problem.

Keyboard operation, and what is not claimed

Radix carries the dialog, popover, dropdown, tabs and radio-group semantics — focus trapping, Escape, roving focus — and four keyboard and AT defects were found and repaired rather than waived: a role="button" on an <article> that wrapped its own focusable chips; a control nested inside a control on the focused evidence card; a Radix modal dropdown that made the whole app aria-hidden with focusable content inside it, fixed with modal={false} rather than by hiding the rule; and two scroll regions — the transcript and the cited-evidence list — that a keyboard user could not scroll at all, because every coded span is a <mark> and nothing inside them was focusable. The reconciliation radios were given real 24 / 44 px targets while keeping the 16 px painted mark.

What is not claimed: this suite has no keyboard-only walk of the critical flow. Every state the audit reaches, it reaches through a control resolved by role and accessible name, which is a strong signal and not the same proof. It is recorded here as an absence rather than dressed up as a presence.

Sources, and what was not taken

The stack, as installed

The three-role type system is the brand: anything a participant said is set in a serif, anything the product says about itself is set in Geist, and anything that identifies a piece of evidence — S-09 · 04:12 — is set in mono, so a source ID is recognisable as an object even when it appears in running text.

Specimen — the fixture that refused to invent a transcript

/* rail-empty (S-12) — COPY_DECK §2.10. No S-12 transcript body exists in the
 * deck; the reading column is kept present but no P-12 turns are fabricated.
 *
 * The reading column carries its OWN empty state. COPY_DECK §2.10 gives this
 * state a header, the rail's empty copy and nothing else, so the column rendered
 * as 722 px of blank paper under a header reading `recorded 26 min`: a source
 * that is recorded and blank looks broken rather than empty, and the rail's
 * `Select a passage in the transcript to code it.` asks for something the column
 * cannot offer. The two lines below add no fact the deck does not already carry
 * — S-12 exists, its recording is 26 minutes (`headerSub`), and the source index
 * lists it as `0 excerpts` / `Uncoded`. No P-12 turn is invented to close the
 * void; that is the fabrication VISUAL_ACCEPTANCE_STANDARD §4⑥ exists to stop. */
export const CODING_EMPTY = {
  headerTitle: 'Apprentice shift · six bays',
  headerSub: 'S-12 · P-12 · apprentice · 8 months · recorded 26 min',
  railTab: 'Evidence 0',
  emptyLine: 'No excerpts on S-12 yet.',
  emptyBody: 'Select a passage in the transcript to code it.',
  emptyControl: 'Suggest codes',
  footerCount: '0 excerpts coded',
  transcriptEmptyLine: 'No transcript for S-12 yet.',
  transcriptEmptyBody:
    'The 26-minute recording is stored. There is nothing to select until it has been transcribed.',
};

Specimen — both exits from a contradiction, written the same way

function resolveConflict() {
  setConflict('resolved');
  setNavBadge(0);
  announce(CONFLICT.resolvedStatus);
  setExtraActivity({ text: CONFLICT.resolvedActivity.replace(' — just now', ''), time: 'just now' });
}
function leaveUnresolved() {
  setConflict('unresolved');
  setNavBadge(1);
  announce(CONFLICT.unresolvedStatus);
  setExtraActivity({ text: CONFLICT.unresolvedActivity.replace(' — just now', ''), time: 'just now' });
}

Neither branch removes an excerpt from anything. They differ in the badge count, the announcement and the line written into the activity log — which is the whole claim of Folio C, in the only place where a claim like that can be checked.