Evidence grades: why likes corrupt learning

A reel can do numbers and still be a mistake.

If your loop cannot say that out loud, it is not measuring. It is flattering you. Step five of an online agentic closed loop is measure honestly. Likes are a weather report. They are not a verdict on strategy. A system that cannot grade evidence will treat a lucky spike as a constitution amendment. That is how brands become joke accounts with a logo.

Likes ≠ proof. Write it on the wall if you must. Then build grades, because walls do not stop a model.

How corruption actually happens

The loop is simple, and wrong:

1. Post ships (maybe it even had a so-what; maybe it did not).
2. It gets a burst of easy emotion — a dunk, a dog, a ratio.
3. The system, hungry for a signal, files “this is what works.”

A working grade, not a dashboard religion

You do not need a PhD. You need honesty that can survive a client.

Grade A — decision evidence. A named buyer did a named thing you asked: booked, applied, bought, referred, complained (yes, complaints are evidence). Platform-sourced, not guessed. Sample size you would defend.

Grade B — qualified attention. Replies that show understanding of the offer; saves on a proof post; traffic to a page that can convert. Still not a sale. Better than a heart.

Grade C — reach without recognition. Impressions, views, follower ticks. Weather. Useful as weather. Illegal as strategy.

Grade F — contaminated signal. Giveaways, rage-bait, bots, a celebrity accident. Do not learn from F. Do not let the agent propose a brand change from F. File it as contamination.

When a platform does not report a figure, you show a dash — never a filled-in guess. That is the habit on the measurement record. Zero is a number. A quiet week reported as a quiet week is what makes a good week believable.

NIST wants evaluation, not vibes. The ACCC wants claims you can prove. “Our engagement is up” is not a claim you should make to a client unless you can show the grade.

What the agent is allowed to learn

Propose a change of tactic from Grade A and B, inside the permission slip. Never a change of truth. Observation is not authority. Step six of the loop is learn without hijacking. Step seven is back to brand truth.

If you only remember one operational rule: the system must be able to say “this did numbers and we will not do it again.” That sentence is adult. The rest is a caption factory with analytics.

Restart uses the same grades on a career: applause is C; a paid brief is A. gailmacleod.ai. Soft door: app.theagencyiq.ai.

Frequently asked questions

Why do likes corrupt AI learning?

Because a loop that treats easy emotion as success will optimise for easy emotion. The brand drifts toward whatever is cheapest to react to.

What is an evidence grade?

A label on a result: A decision evidence, B qualified attention, C reach without recognition, F contaminated. Learn from A/B. Do not rewrite the brand from C/F.

Should we stop reporting likes?

No. Report them as weather. Do not let them veto strategy, and never let them rewrite brand truth. 4. Next week’s plan moves toward dunks, dogs, ratios. 5. Brand truth is still in a PDF nobody opens. That is learning as hijack. Lions Law forbids it at the other end of the loop: you may not let performance rewrite recognition. Lions Law. Followers already stopped being the scoreboard — why followers stopped mattering. Likes are the same trap at a smaller denomination. LinkedIn’s marketing desk has been blunt in 2026 about spend that looks busy and does not move the business. Vanity is not a B2B-only disease. It is any loop that cannot tell a grade from a count.