Ratios Are Questions, Not Grades
A ratio looks like a grade because it arrives wearing arithmetic.
That is the trap.
In an AI assisted product system, a ratio is usually not a verdict. It is a question with a numerator, a denominator, and a grudge against lazy interpretation.
A signal pass can report 98 observed items, 16 newly inserted observations, and 39 opportunity candidates. Another can report 126 observed items, 0 newly inserted observations, and 44 opportunity candidates. A third can report 94 observed items, 48 newly inserted observations, and 10 opportunity candidates.
The childish reading is to rank them. Higher insertion rate, better. Higher opportunity count, better. Lower cost, better. Nice little scoreboard. Very clean. Also a decent way to teach the system the wrong manners.
The useful reading is slower: what kind of product judgment does each ratio expose?
A ratio preserves the argument
A count is often too polite. It says something happened, then leaves before the argument starts.
A ratio makes the argument harder to dodge. Observed to inserted asks whether the system is seeing novelty or mostly touching familiar material. Inserted to opportunity asks whether accepted evidence is turning into candidate work, or whether the candidate layer is operating on a different pool than the newest inputs. Cost to request volume asks whether the work was cheap in a meaningful way, or merely cheap enough to avoid scrutiny.
Promptara Lab keeps returning to this distinction because small systems can now generate a lot of tidy numbers before anyone has had coffee. The dangerous part is not that the numbers are wrong. The dangerous part is that they are accurate and still not enough.
A run that sees 126 items and inserts 0 new observations is not automatically failed. It may be deduplicating correctly. It may be revisiting known material. It may be connected to an opportunity layer that still has candidate inventory from earlier accepted evidence. Or it may be doing something less useful. The ratio does not answer that. It tells the operator where to look.
That is already better than a total.
This is close to the argument in Counts Flatten the Mess: totals are tidy because they have thrown away shape. Ratios are useful when they put some of that shape back.
Good ratios keep their nouns
The fastest way to ruin a ratio is to blur the nouns.
Observation is not insertion. Insertion is not opportunity. Opportunity is not publication. Publication is not distribution. Distribution is not demand. Demand is not revenue. Anyone who compresses that chain into a single success percentage is doing dashboard origami, not operations.
A useful ratio keeps both sides named:
- observed items to newly inserted observations
- newly inserted observations to opportunity candidates
- requests to token volume
- cost to request volume
- traffic to same day actions
- queued messages to messages actually taken
None of these should be treated as universal health scores. They are lenses. Each one tests a different part of the product system.
This is where many automation dashboards get a little silly. They want one heroic number that makes the operator feel informed. The product needs the opposite: several smaller ratios that are humble enough to say, please inspect this branch.
A traffic report that says the strongest traffic to action signal was none is not a business obituary. It is a diagnostic branch. Maybe the offer is unclear. Maybe the audience was casual. Maybe the measurement surface is incomplete. Maybe the product simply did not earn the action that day. The ratio cannot decide which one is true. It can prevent the team from pretending the answer is already known.
The denominator is a product decision
Every ratio hides a decision inside the denominator.
Cost per request sounds simple until the requests are doing different kinds of work. Opportunity candidates per observation sounds simple until one product surface needs broad exploration and another needs strict filtering. Insertions per observed item sounds simple until the source has a lot of repeat language and the system is supposed to avoid storing duplicates.
This is why one global score becomes nonsense quickly. A tidy cross asset comparison can punish the exact behavior a system was designed to perform. A cautious filter looks lazy. A noisy discovery pass looks productive. A narrow product with stable source language looks stale beside a broad product with more fresh phrasing.
Arithmetic does not fix bad product framing. It just gives it a haircut.
The denominator should be chosen for the decision at hand. If the question is novelty, divide by observations. If the question is review load, divide by candidates. If the question is cost discipline, divide by the unit of work that creates operator burden, not merely the unit that appears on a usage bill. If the question is demand, do not let pageviews cosplay as action.
This is also why One Threshold Is a Lie keeps aging annoyingly well. A threshold without product context is just a preference pretending to be governance.
Ratios should slow the machine down
The point of a ratio is not to make the machine feel more scientific. The point is to slow down the next bad assumption.
A good operating ratio should create one of three responses:
- Inspect the inputs.
- Inspect the acceptance rule.
- Inspect the next action the system wants to create.
If it does none of those, it is probably decoration.
For the agentic framework under the hood of Promptara Lab, this is a design taste issue as much as a telemetry issue. The system should not merely say that work completed. It should expose enough shape for a builder to know whether the next move is to continue, pause, tighten the filter, loosen the filter, review the candidate pile, or ignore the noise.
That does not require a giant command center. It requires boring nouns, honest denominators, and a refusal to let one percentage sit on a throne.
Promptara Lab is built around that kind of operator surface: small product systems, explicit telemetry, and enough friction to keep automation from grading its own homework. The public notes live at Promptara Lab, where the interesting part is usually not the largest number. It is the ratio that asks the more annoying question.
A ratio is not a grade.
It is a little interrogation device. Use it accordingly.
