Most automation is built to maximise output. Colosseum is built to maximise restraint. The reasons we do not publish are part of the product. This essay defends that position and shows the data — what gets stopped, by whom, and why — for the last quarter.
The output-maximisation default
Every content automation pitch you have read this year is a pitch for output. More posts, more variants, more drafts, more channels. The unit of value is the published artefact. The unit of cost is the seat licence. The growth lever is volume.
The lever works for a quarter or two. Then a platform’s policy team notices, the algorithm’s quality filter notices, the audience notices. Volume becomes a tax. The pitch that won the deal becomes the reason the deal does not renew.
We have lived this cycle and we have decided not to play it again.
What restraint looks like in code
The audit log is the visible part. The invisible part is in the agents themselves.
- Feasibility rejects proposals that cost more than they will return. Most ideas are this.
- Decision advances only the top-ranked proposals from each batch. Most batches advance one or none.
- Legal Compliance blocks any draft that lacks a required disclosure, regardless of operator urgency.
- Niche Resonance withholds a pattern from another niche until a 72-hour read-out from a sandbox publish confirms transfer.
- Community does not respond at all to comments that fall outside the documented response rules — they go to a human.
In aggregate, the system stops more proposals than it advances. We design for this. We track the stop-to-advance ratio as a quality metric: too few stops means the guardrails are weak; too many means the operator’s brief is unproductive.
The data, last quarter (illustrative — production data lands in the next Safety Report)
The numbers below are from the v1 sandbox period. Production data appears in the next quarterly Safety Report.
| Decision | Count | % |
|---|---|---|
| Advanced to publish | 4,127 | 31% |
| Withdrew (superseded by a better proposal) | 3,544 | 27% |
| Rejected (failed Feasibility, Legal Compliance, content filter, or rate-limit) | 4,291 | 32% |
| Escalated to human | 1,389 | 10% |
| Total decisions | 13,351 | 100% |
The headline: 31% of proposals reach publication. The other 69% are stopped, withdrawn, or escalated. None of the 69% costs the operator anything in platform reputation, because none of them was published.
Why this is the architecture, not the marketing
Restraint built into a system is durable. Restraint built into a marketing message is not. We have seen many companies promise “ethical AI” or “responsible automation” without changing the underlying architecture. The promise is rhetorical; the architecture is still output-maximising.
Our position is the opposite: the architecture is restraint-maximising; the marketing is honest about it. The home page shows the ratio. The audit log proves it row by row. The operator dashboard makes the ratio adjustable per niche. The team is paid against the ratio holding.
If a future operator asks us to dial down the guardrails, we will tell them we cannot. The system is not configurable that way. Some of the rules — the global ones — are not configurable by anyone, including the team. Those are documented at /safety section 7.
The case, summarised
A content system that cannot say no is not an editorial system. A team that cannot show why it said no is not an editorial team. The audit log is how we say it. The architecture is what makes the saying possible. The home page is where we show the ratio.
The reasons we do not publish are part of the product. We mean it.
The Safety architecture is at /safety. The seven canonical home-page audit rows are at /. Comments via [email protected].