§1 · The setup

Three systems. One QA team. Every release night looks the same.

Each system has its own console, login wall and quirks. Integration bugs hide in the seams.

§1 · The math

1,284 cases. Run by hand. Every release.

Across the three systems, the manual QA catalogue looks like this — and grows quarter on quarter as new flows ship.

Counts above are from the POC's demo fixtures — placeholder numbers we wired in for the runner. Worth replacing with CIFF's real catalogue counts before the readout.

§1 · The thesis

The QA spreadsheet is a programming language.
We just hadn't compiled it yet.

Manual test cases already describe inputs, steps, and expected results. That's a spec. The only thing missing is a runtime that understands it — and a way to feed the results back.

If the spreadsheet is the spec, the browser is the compiler, and the run is the assertion — we don't need to rewrite QA, we need to plug it in.

A five-stage pipeline, end to end.

Upload a sheet. Walk away. Come back to results, screenshots, and a diagnosis on every failure.

§2 · What we built

One portal. Three integrations. Zero scripts to hand-write.

QA lands on a personal dashboard: total / passed / failed / skipped, modules with pass-rate bars, and the last five runs at a glance.

§2 · Upload → generate

Drop a spreadsheet. We figure out the rest.

import { test, expect } from '@playwright/test';

test('TC-015 reject grant submission', async ({ page }) => {
  await page.goto('https://cif.fluxx.io/login');
  await page.getByLabel('Username').fill(process.env.FLUXX_USER);
  await page.getByLabel('Password').fill(process.env.FLUXX_PASS);
  await page.getByRole('button', { name: 'Sign in' }).click();

  await page.goto('https://cif.fluxx.io/grants/4821');
  await page.getByRole('button', { name: 'Reject' }).click();
  await page.getByLabel('Reason').fill('Out of programme scope');
  await page.getByRole('button', { name: 'Confirm rejection' }).click();

  await expect(page.getByTestId('grant-status')).toHaveText('Rejected');
});

Column-mapper figures out which column is "title", which is "steps", which is the test ID — even when the QA file has been re-saved nine times.

§2 · The run

QA can watch tests run. So can their PM.

§2 · A real failure caught

TC-015: Reject grant submission — failed

Async race condition. The assertion fired before the API responded. The kind of bug a manual QA might miss because the screen "looks right" after a second.

Three failure modes the runner has already surfaced: async race, brittle locator, real backend bug. We file the third with the dev team — the first two, we fix in seconds.

§2 · Claude's fix

From diagnosis to a one-line patch.

  await page.getByRole('button', { name: 'Confirm rejection' }).click();
+ await page.waitForResponse(r =>
+   r.url().includes('/api/grants/') && r.url().includes('/reject')
+ );
  await expect(page.getByTestId('grant-status')).toHaveText('Rejected');

§2 · So far

What the POC shipped.

What if every manual case became a runnable spec?

Not just the three flows we hand-picked for the POC. The whole catalogue — 1,284 cases — converted, suite-ised, and rerun on every PR.

Numbers on this slide are illustrative projections — not measured. They assume average case complexity and 2× concurrency. Real numbers will land somewhere in this ballpark once we have a baseline week of runs.

§3 · The conversion loop

Engineer + AI + runner — a feedback loop that gets faster every week.

§3 · What changes

Three teams. Three new defaults.

§3 · Roadmap

Six months to a self-healing regression suite.

No big-bang migration. Each milestone ships value on its own. If we stop at M3, CIFF still keeps the 200-case library.

CIFF × DHWANI · MAY 2026 · v0.1