Seven defects, and six were invisible from the design alone.

One gate asked two questions and sat a level too high

v1's G9 asked whether you review output against criteria set in advance and whether those criteria are written down. Those are different things at different maturities. Blum reviews everything — it'll give me a draft and then I'll change it and I'll just send it — but has nothing written. One gate, two verdicts, so the gate failed and pinned him at Level 3 despite passing every delegation marker emphatically.

The principle, independent of Blum: Level 4 is defined by retaining judgment while delegating tasks. Formalising that judgment so it can be handed over is a compounding practice — it belongs beside friction logging and distribution at Level 5, not beside scheduling at Level 4.

A Level 2 gate asked you to rate a capability, and could collapse the ladder alone

v1's G2 asked whether you notice when output is wrong in your own area — reliably, and I know where it tends to fail. That is a self-rating of an internal capability, which is precisely what the model's own design note tells you not to do: ask for observable facts, not agreement statements. v1 violated its own rule in its second question.

The sensitivity check made it worse: because G2 gates Level 2, an error there drops any score to Level 1 regardless of everything above it. The most load-bearing item in the instrument rested on its least observable evidence.

The principle: at the bottom of a ladder, gates should test a low-bar behaviour that is either happening or not. Discrimination belongs higher up, where there is more to discriminate between.

A gate's wording excluded half of what it was meant to catch

v1's G7 asked whether AI work runs on a schedule. Carson's setup is emphatically delegated — Watchdog sweeps every customer account and returns the top three things going wrong; Land PR fires when an agent reports a PR done, runs up to two review loops, and records a narrated walkthrough. Neither is on a schedule. Land PR is event-triggered, and Watchdog he fires by hand when he feels the need.

The principle: the corpus's own loop taxonomy names four triggers — heartbeat, cron, hook, and goal. Only two of them are schedules. A gate testing “work starts without you” that recognises just one trigger type is measuring the implementation, not the property.

Two co-equal axes measured with four questions against nine

The whole point of the two-axis structure is that the axes compound — the source study's finding is that having both capabilities is worth more than the sum of having each. But v1 measured fluency with 4 items and work-system design with 9. Fluency scores moved in 12.5-point jumps, and the imbalance flag needed a 34-point gap it could rarely reach.

v1 acknowledged this in a note — “most observable markers live on the S axis, which is why the assessment asks more about it” — which was honest and also a description of a defect rather than a justification for one. The fix is not padding with unobservable questions; it is finding observable fluency markers, which exist.

Two gates were asking the same question

Once G9 was split, its written-criteria half and v1's G12 — “you could hand someone else a written definition of what a good result looks like” — were the same gate under two names, both at Level 5, both on the fluency axis. Scoring Blum against both produced the same verdict from the same evidence twice, which inflates whichever axis carries the duplicate.

The principle: duplicated gates are silent weighting. If a construct matters double, weight it deliberately rather than by accident of drafting.

The headline number hid the shape

Strict gating is the right rule — it is what makes this a diagnostic that names one next action rather than a score. But “Level 3” alone threw away the most informative thing about the Blum profile, which was that every Level 5 gate was already met. That lived in a flag below the fold.

v2 reports a level and a reach: the highest level whose own gates are all met, ignoring gaps beneath it. When they differ, the headline says so.

An inverted ladder produced a confidently wrong answer

Nothing stops someone failing both Level 2 gates while meeting every gate at Level 4 — an employee who inherited a configured system, or anyone who read a low gate more strictly than a high one. Strict gating puts them at Level 1, which is right. But everything wrapped around that number was wrong: the page told a person with scheduled tasks writing to a system of record that nothing accumulates — what you worked out today will not be there tomorrow, and advised them to start saving their prompts.

The principle: a level is a summary, and a summary that contradicts the answers it summarises is worse than none. Where the profile and the level's narrative disagree, the narrative yields — not the gating.

The fix separates two causes the instrument had been conflating. Some gates are true by construction if a higher one is: a defined checkpoint on delegated output (G9) means you check output before using it (G2); written review criteria (G13) mean the review happens (G9). When the higher holds and the lower doesn't, that is far likelier a misread question than a real gap, and it is now reported as one — with both answers quoted back, because the level turns on it. When no implication applies and the inversion is real, it gets its own diagnosis: automation you did not build — delegation inherited without the habits underneath it. That is not a contradiction. It is the most exposed profile on the scale, and it had no name before.