Analysis that shows its working

The problem

The engine would recommend entering at market precisely when price had run furthest from any support, so the more extended a move already was, the more readily it chased. The client photographed it proposing an entry at the top of a vertical rally.

What we did

The rule read backwards against risk without being obviously wrong: the retestable check asked whether the entry zone sat within a set distance of a level, and the failure branch fell through to entering at market rather than declining to act. That case now proposes no trade. The verdict is untouched, so the reading stays bullish and still explains why, but the instruction becomes wait, carried through the app by an existing field rather than a new one. It is applied symmetrically to shorts.

The outcome

The product stopped conflating a market observation with a trade recommendation, which is the distinction its credibility rests on.

The problem

The analysis is written in one place and drawn in another, and nothing stopped a chart label from showing a rounded, stale or invented figure that contradicted the paragraph beneath it. There is no natural point where a human notices 61.8 percent quietly becoming 62 percent, and a reader who spots the two disagreeing has no way to know which to trust.

What we did

A test asserts that every digit appearing in a chart label traces back to a value the engine computed, including the per-level fibonacci captions, which are the largest cluster of numbers on the chart. It was validated by mutation: hard-coding a touch count and rounding a fibonacci caption both fail it.

The outcome

The chart has nothing of its own to say. The guarantee is mechanical rather than a convention someone eventually forgets.

The problem

Running the charting feature end to end on Cardano revealed that five of six fibonacci levels were drawn off-screen, and a sixth had vanished entirely. Neither defect appeared in the unit tests and neither logged anything. The geometry was right; the view was wrong.

What we did

The macro fibonacci was anchored across 300 daily candles while the chart opened on a 90-day window, and the label separation pass only ever pushed labels downward, so levels arriving bottom-to-top cascaded off the pane. The timeframe snap became profile-aware, so a daily spot analysis opens on a one-year chart wide enough to contain what it describes, and label positions are now resolved and sorted before anything is drawn.

The outcome

All six levels render in view, every drawn level is labelled, and the fix carries a regression test that fails without the sort.

The problem

The daily AI-cost view reported figures that could not both be true, such as eleven AI calls against a cost of 0.0016 dollars. It was the single number used to watch AI spend against subscription revenue, so being quietly wrong mattered more than being obviously broken.

What we did

Cached results were settling under the original provider's name, so a cache hit, which makes no API call and costs nothing, was counted as one. It is fixed at both ends: the route now settles cache hits under their own provider, and the view no longer trusts the provider field alone to mean a call was made.

The outcome

Deliberately redundant, because the number is money, and a view that is only correct while every caller remembers a convention will eventually be wrong.