Transparency, the first pillar.

Of the four pillars Mirox is built on — transparency, safety, intelligence, EU-native — transparency is the one that decides whether the other three are believed. A safety claim from a black box is a marketing assertion. An intelligence claim from a black box is a brand promise. Only transparency is verifiable on the bid you ran yesterday.

This is the long form of why we built it that way.

The category lives behind glass

Pacvue, Perpetua, Teikametrics. Three good tools, run by people who know Amazon advertising well. None of them can answer the question "why did the bid on keyword X rise from €1.20 to €1.84 between Tuesday and Friday." Not because the answer doesn't exist inside the system — it does — but because surfacing it was never a product priority.

The historical reason is reasonable. PPC software earned its early reputation by beating manual operators on outcomes. The pitch was trust us, we beat the baseline. Trust was a fair price for performance when performance was the only thing being measured.

The audience changed. The marketing director who used to be the buyer is now the seller-to-CFO. The CFO does not measure performance. The CFO measures defensibility. "It beat the baseline" is no longer the question on the table. "Show me what changed and why" is.

A black box cannot answer that. The current category lost a layer of buyer trust over the last three years and is still mostly pretending it didn't happen.

What transparency means at Mirox, concretely

For one bid, on one keyword, in one campaign, Mirox writes the following to the trace:

• The proposing agent. The system has sixteen specialised agents. One proposed the bid. Zero or more contributed. The trace names them — Tactician proposing, Strategist contributing stage, Sentinel contributing supply, Semantic contributing relevance.
• The Bayesian conversion estimate. A BetaBinomial pCVR computed per marketplace per category. The trace shows the value, the prior, the data window the prior is drawn from.
• The relevance score. Asymmetric semantic similarity between search term and product detail page. The score, the embedding model version, whether it cleared the relevance gate.
• The supply multiplier. If Sentinel applied a throttle for low days-of-cover, the trace shows the DOC at decision time, the multiplier, and the stale-feed cap if the SP-API feed was older than one hour.
• The alternatives ruled out. Tactician considers a candidate set. The selected bid is one of them. The others are listed, each with the reason it lost — under-bid relative to first-page minimum, over-bid relative to a tripped circuit breaker, off-policy relative to current stage.
• The safety gates that fired or held. ROAS circuit breaker, daily spend anomaly, SP-API freshness, optimistic-locking version, recovery floor, marketplace gate. Pass or trip. Threshold in force at the time.
• The export. All of the above, downloadable as a CSV with the original Amazon IDs preserved.

That is on the record for every bid. Not a sample. Not the ones the system finds interesting. Every one.

The CFO test

The test we use internally to decide whether a feature ships is what we call the CFO test. Can the marketing director defend this decision to the CFO without rebuilding the reasoning from scratch?

A bid that moved without a trace fails the test. The marketing director either guesses, blames the tool, or sets up a 45-minute meeting with the analyst who can reconstruct the reasoning. None of those are defensible answers.

A bid with a trace passes immediately. The marketing director pulls the CSV, walks the CFO through the inputs, the alternatives, the gates that were checked. The conversation that used to take 45 minutes takes four.

The CFO test is not a brand-marketing artefact. It is the single design constraint that has shaped more product decisions at Mirox than any other.

Defensive transparency, offensive transparency

The case for explainability is usually framed defensively — "what happens when something goes wrong." That frame is correct but undersells the second case.

When the AI is performing well, the trace teaches you what your account looks like through the model's eyes. A marketing director who reads twenty traces a week learns more about their own account than one who reads a hundred dashboard summaries. The trace is a calibration tool first, an audit tool second.

We have watched founding-cohort sellers cancel agency contracts not because the agency was wrong but because the trace surfaced a structural account decision the agency had never explained, and once the seller saw it they realised they could decide it themselves. Transparency makes the seller more competent. That is the offensive case.

What you do not get from a transparent system

Two things, worth saying plainly.

You do not get a shorter dashboard. A transparent system has more data on the screen, not less. The trace is a deep artefact — six fields per bid, sixteen agents that might be involved — and the dashboard reflects that. Sellers who want a tile that says "ACoS green / yellow / red" will not love our default view.

You do not get a faster sales cycle. Selling a transparent tool to a buyer who was scarred by a black-box tool last year takes longer than selling a black-box tool to a buyer who has not yet been scarred. The Shadow Mode tier exists in part because the proof has to happen on the buyer's real account before the conversation moves forward.

We accept both costs. They are the same cost — being honest about how much information the seller deserves.

Where this pillar leads

The other three pillars exist because transparency on its own is insufficient. Transparency without safety means you watch the AI burn your spend in slow motion with a full audit trail. Transparency without intelligence means you watch a rule engine cosplaying as AI, fully documented. Transparency without an EU-native architecture means a US tool with extensive logging that still cannot tell you why the German campaign is bleeding.

But transparency is the first pillar because nothing else is verifiable without it. Read the safety post and the intelligence post next.

If you want to see a real decision trace before the public Shadow tier opens in Q3 2026, the founding cohort is open through summer 2026.