The bar moved last year, and most engineers found out the hard way. A mid-level engineer (L4) in 2026 is expected to ship code that runs. A senior (L5) is expected to ship code a reviewer will trust six months later, after the model that wrote half of it has been deprecated. The difference is no longer about scope or quantity. It is about the kind of evidence the engineer can leave behind. This post is about what that evidence looks like, what the rubric now weights, and the four conversations that actually move a promotion packet.
Why this matters
Promotion packets used to argue from output. Tickets closed, projects shipped, lines changed. AI assistance flattened that argument: a productive L4 and a productive L5 ship similar volumes now. Calibration committees noticed. The weight quietly shifted onto the dimensions a model cannot ghostwrite — the reasoning trail behind decisions, the rejections that did not make it into the diff, and the architecture choices that survive a rewrite. The engineers who understand that shift get promoted. The ones who keep arguing from velocity stall at L4 for an extra cycle.
What does the rubric say has changed at L5 in 2026?
The rubric did not get harder. It got narrower. The dimensions that used to make up most of an L4 score — correctness, spec fidelity, shipping cadence — have compressed at the top. Almost every L4 in a serious calibration pool now writes code that passes tests and meets the spec. The discriminator moved upward into defensibility, AI-rejection rate, and replay clarity. The chart below shows the typical rubric profile at each level, taken from our 2026 calibration cohort. L4 and L5 score nearly the same on correctness. The gap opens elsewhere.
The two largest gaps are defensibility (32 points) and AI-rejection rate (29 points). Correctness moves only 7 points. The shape of the diagram is the actual shape of the conversation now. A reviewer who says "your code is fine, but I cannot tell you to L5" is almost always pointing at one of those two dimensions, even if they do not name them that way.
The radar below tells the same story from a different angle. It plots skill emphasis — what a panel weights when deciding the level — across six axes that overlap less with the rubric. The L5 profile leans heavily on defensibility, architecture sense, and scope ambition. The L4 profile is more balanced, which is exactly what gets it stuck.
Why has defensibility become the biggest gap?
Three years ago, the diff carried the reasoning. You typed it, so you remembered why. In 2026, the diff is the output of a conversation that nobody keeps. The model proposes a retry budget of three; the engineer accepts it; six weeks later, on-call asks why three. The engineer who can answer — because they tested the upstream timeout against the SLO before accepting — is operating at L5. The engineer who shrugs is operating at L4, regardless of how much code they ship.
This is not a soft skill. It is the only skill that scales when the cost of writing code goes to zero. When everyone can produce the artifact, the value moves to whoever can argue for it under pressure. Reviewers learned this faster than candidates did. Most L4s we coach have never seen the rubric they are being scored against, and read their pushback as taste. It is not taste. It is a structural shift in what evidence the panel is looking for.
The expectation table below makes the L4 / L5 difference concrete. Same dimension, same panel, different bar.
| Dimension | L4 expectation | L5 expectation | Most common gap |
|---|---|---|---|
| Defensibility | Can explain what the code does. | Can explain what was rejected and why, without the chat log. | Engineer kept the diff, not the reasoning. |
| AI-rejection rate | Accepts most model suggestions if tests pass. | Rejects model suggestions on principled grounds and can name the principle. | Engineer rejects on vibes, not on a criterion they can teach. |
| Architecture sense | Owns a module. Touches a service. | Names the invariant a service depends on. Owns the failure mode. | Engineer can describe code structure but not system contracts. |
| Scope ambition | Finishes the ticket as written. | Renegotiates the ticket when the ticket is wrong. | Engineer treats spec as fixed rather than as evidence to question. |
| Mentorship | Answers questions when asked. | Sees the question one level up and rewrites the team's defaults. | Engineer helps individuals but does not raise the bar around them. |
| Replay clarity | Has a clean commit history. | Has a record of decisions a stranger can audit. | Engineer documents code, not choices. |
What are the four conversations that move a promo packet?
Promo packets are decided in four conversations, not one. Each one has a failure mode that AI assistance makes easier to fall into.
The first is the scope conversation with your manager. The failure mode at L4 is to accept the scope as written and execute it well. The failure mode at L5 is the same — except now the panel reads "executed the scope well" as L4 behavior. The L5 move is to come back two weeks in with "the spec assumed X, but the data shows Y, so I propose we cut these two parts and add this one." Owned scope changes are the strongest L5 signal we see.
The second is the architecture conversation with a senior reviewer. The L4 failure mode is to defend the design that landed. The L5 failure mode is to defend a design you do not understand because the model proposed it. The fix is to enter every architecture review with one rejected alternative and one invariant the design protects. Both should be in your own words.
The third is the AI-rejection conversation. This is new. It is the moment a teammate asks "why did you not just take the model's suggestion." An L4 answers "it didn't look right." An L5 answers "it would have violated our idempotency contract on retries — the suggested key collapsed two tenants into one." The principle is named. That single behavior, repeated across a year, is what panels remember.
The fourth is the promo conversation itself, where you describe a project to a calibration committee that was not in any of the rooms. AI assistance makes this the most dangerous one. The committee can sense, within a paragraph, whether the engineer is reading their own reasoning or paraphrasing the model's. The fix is to bring three decisions you owned end-to-end and can defend without notes. Three is enough. One is not.
The v2.6 rubric methodology behind the L4-vs-L5 gap chart.
From Codritium Research
The Defensibility Rubric, 2026
The public scoring rubric behind every Codritium benchmark and certification: six weighted dimensions, twelve task categories, inter-rater agreement, and a frozen reference quarter.
Common questions
Does my employer use this rubric?
Probably not by name. The rubric we publish is a research instrument calibrated against panel decisions at FAANG-class shops. Most internal leveling guides at those companies share the same shape — defensibility is rising, raw correctness is compressing — even when the document calls them something else. Read your own leveling guide first; the rubric is a lens, not a replacement.
How long does L4 to L5 take in 2026?
Median in our cohort is 2.4 calibration cycles after the engineer starts running deliberate replay reviews. Without replay practice, the median is 4.1 cycles. The biggest variable is not headcount or scope — it is how soon the engineer starts treating the reasoning trail as the deliverable.
Is AI helping or hurting promo chances?
Both, depending on what the engineer does with it. AI raises the floor on what gets shipped, which compresses the correctness gap and makes it harder to stand out on output. It also raises the temptation to accept model decisions without owning them, which directly tanks the defensibility score. Engineers who use AI as a sparring partner improve faster. Engineers who use it as a ghostwriter stall.
What if I don't get the visibility I need?
Visibility at L5 is mostly self-produced. The two highest-leverage moves are writing a short post-mortem of a decision you owned (200 words, in the team channel) and volunteering to be the reviewer on someone else's architecture doc. Both are cheap, both produce paper trails, both compound.
Should I try to skip a level?
Skip-level attempts succeed when the engineer has already been operating at the target level for a cycle and the packet is mostly bookkeeping. They fail almost universally when the engineer is trying to grow into the level inside the packet itself. The fastest path is usually two solid cycles, not one risky one.
Do open-source contributions still count?
They count as evidence, not as a multiplier. A merged PR to a known project is a fine artifact. The actual evidence the panel wants is the same evidence as for internal work: what was the decision, what got rejected, why, and how would you revise it now.
Where to go next
- The Defensibility Rubric, 2026 — the scoring instrument behind the gap charts above
- Defensibility: The Engineering Skill AI Cannot Replace — the underlying skill the L5 rubric leans on
- Codritium challenges — Hard-tier tasks that score against this rubric