What we kept finding was missing
A chronology of corrections to the research instrument. This is not a changelog: it is the list of the times we discovered that the platform could not see something we believed it could.
It is worth reading for one reason. A project whose thesis is that informal commerce produces no structured data spent six months running into that same problem inside its own platform — and each finding carries a date and a commit hash. 512 commits since 12 April 2026; 76 of them titled fix.
A model adopted and abandoned
Mercadito launched as a delivery marketplace charging a 5–8 % commission. On 24 August 2026 (dbdc382) it stopped operating deliveries: orders composed on a digital menu now leave for the business’s own WhatsApp, and the platform never touches the money.
The commission code was not deleted. It remains in the repository behind a feature flag, and that is the primary evidence of the proposal: a model that was adopted and abandoned, with a date and a hash, is worth more than any claim about it.
Counters were not data
The platform had counted menu opens and orders since the beginning — but only as running totals. They say how many and never when, and a cumulative total cannot be un-aggregated afterwards. Every day without a time series was a day lost for good.
On 28 September 2026 (ecbe979) that changed: one row per menu open and per order sent, with a timestamp, an origin, a 24-hour random session id and an amount band — plus the business’s one-tap confirmation that the order actually arrived. Without that confirmation the order counter was measuring intentions and we were calling them sales.
The instrument itself was blind
The most instructive part, and the one worth telling plainly: four times in four days the series turned out to be blind to an entire channel. All three were corrected the same day they were found — which is the only moment when correcting a series costs nothing.
| Date | What the series could not see |
|---|---|
| 28 Sep | The native app sends the same beacon, but origin had no such value: all of its traffic fell into “unknown”. |
| 28 Sep | A printed QR code and a link pasted into WhatsApp both arrive without a referrer. New codes are now printed with a marker; the ones already stuck on tables are not, and that limitation is declared. |
| 1 Oct | Facebook and Instagram — the largest channel — fell into “unknown” or “link”. Facebook’s in-app browser rarely sends a referrer at all. |
| 1 Oct | The QR code each business downloads from its own panel did not carry the marker — only the printed pack did. We would have handed out marked codes while businesses kept posting their own. |
The fourth deserves a note of its own for who caught it: not the code and not a review, but field knowledge — “every business profile already shows its QR”. It is the only one of the four that could not have been found by reading the repository: it took knowing how the platform is used in real life. Verified by decoding the PNG served in production.
Two corrections of the same kind, elsewhere:
8bb59a7— The menu search ignored category names: searching “hamburguesa” in a menu with an entire hamburger section returned “no results”. It lost the order and would have filled the empty-search table with catalogue gaps that do not exist.b081891— The courier board’s rate limit sat beforevalidation, so anyone mistyping their phone number burned their attempts correcting their own mistake.
What the data corrected about us
The last stretch is not bugs. These are assumptions of ours that the data contradicted, on 1 October 2026:
- The catalogue was not where we were looking. Products are generic and shared; a business’s catalogue lives in the price table — which was already a history, so catalogue growth had been datable all along and nobody had noticed.
- “A single order” came from the wrong table. 51 orders left through the menu to WhatsApp: 452 menu opens, 11.3 % conversion, 18 of 33 businesses with at least one.
- Catalogue size predicts nothing. Correlation with menu opens: +0.07. The largest catalogue (168 products) has 4 views and zero orders.
- The platform’s business type is not the trade. 49 of 52 are “market”, which only says what they use. The catalogue-size hypothesis cannot be tested with that variable: it is not that it comes out negative — there is nothing to cross it with.
- Assisted loading turned out to be constant. All 33 catalogues were loaded with help, so there is no cross to run. And that says more: if setting it up was never the business’s own effort, then the 19 who loaded nothing did not even send the photographs.
Extracted from the project’s private repository; full commit history available to evaluators on request. Related: the research statement and the pilot’s public notice.