02

The Pillars: What "Good" Looks Like

You cannot call something "under" performance without a written definition of "at" performance. • Score a coverage map, not a quiz.

Overview

The pillars are the heart of the system. They are the small set of things you actually evaluate — at hire and every review after. This part defines them for an infrastructure / platform / SRE-shaped role, but the shape generalizes: a technical group (the substance a person must know and operate) and an engineering-mindset group (how they apply it and operate at the seams). Adapt the technical list to your domain; keep the mindset list almost verbatim, because it is what actually separates levels.

The technical pillars are a coverage map

Do not score technical depth as a checklist of tools named. Score it as a map of coverage and judgment: real depth in several domains, working knowledge across most, and the taste to say where each domain is the wrong tool. For a platform role, twelve domains cover the ground:

#DomainWhat "strong" sounds like
01Container orchestrationScheduling, runtime, networking, RBAC — and when not to reach for it
02GitOps / continuous deliveryDeclarative delivery, drift, safe rollback; the failure modes of auto-sync
03Policy-as-codeEnforce vs audit, mutation vs validation; policy as guardrail, not theatre
04Service meshmTLS, traffic splitting; when the mesh pays for itself vs native primitives
05ObservabilitySLOs tied to user journeys; which dashboard lies; load-bearing vs vanity metrics
06Secrets managementRotation, blast radius of a leak, least-privilege issuance
07Ingress / gatewaysL7 routing, TLS, multi-tenant exposure
08Multi-tenancyPooled vs siloed, noisy-neighbor control, per-tenant blast radius
09Developer platformsGolden paths, self-service, staying out of the flow; platform-as-product taste
10Supply-chain securityImage provenance, signing, CI/CD trust boundaries
11Infrastructure-as-codeModule design, state, drift, reversible changes
12Autoscaling & costEvent- and node-autoscaling, cost vs latency, scale-to-zero, bin-packing

Reading the signal is the point: deep in four or five domains, fluent across most, and able to name where a domain is overrated reads as senior. Name-dropping all twelve with no tradeoff or consequence reads as junior. The domains are anchors for a conversation, not a trivia sheet.

The mindset pillars are what separate levels

These are scored on top of the technical work, in every round and every review. Two of them are through-line pillars, evaluated everywhere: autonomy and end-to-end ownership.

PillarWhat "good" looks like
Engineering fluencyCame to the domain through writing code and still reads and writes it — ships tooling, not just configuration
First-principles reasoningStarts from constraints, failure modes, tradeoffs; names non-goals; at the top, reframes the problem and names the binding constraint
Reversibility & blast radiusWrites the rollback before touching anything; speaks in canary, kill-threshold, failover
Debugging fluencyWorks the layers in order with a hypothesis and expected evidence at each; cheapest decisive check first
Simplicity / via negativaRemoves toil, services, alerts before adding; can name a subtraction that made things better
AI in the workflowUses AI daily and verifies it; treats output as a hypothesis; can name what they would not trust it to do in production
Autonomy (through-line)Sets direction without being told; makes reversible calls without asking; escalates the right ones
End-to-end ownership (through-line)Owns something that broke in production — who got paged, what changed; closes loops and follow-ups
Decision-makingEvidence- and tradeoff-first; connects choices to revenue, velocity, and risk
Strategic/tactical balanceMany small reversible bets plus a few big high-conviction ones — nothing stuck in the timid middle
Cross-functional collaborationOperates at the seams; resolves peer-team tension with evidence, not escalation

The barbell, the via negativa, and the paged owner

Three mindset pillars deserve their own note because they are the ones people most often fake.

The barbell is a specific claim about risk-taking: the best operators make many small, cheap, reversible experiments and a few large, high-conviction bets, and almost nothing in the medium-risk middle. Ask for one of each. A person who only has medium-sized, hedged bets rarely moves anything.

Via negativa is the discipline of subtraction. The strongest engineers can point to a service they deleted, an alert they removed, a runbook they made unnecessary — and say what got better and whom they had to convince. Addition is easy; removal takes judgment and courage.

The paged owner is the ownership test: ask for a decision the person personally owned that later broke in production. Who got paged? What did they change? "We had an incident" is a story about a team. "I shipped the change that caused it, I got paged, and here is what I changed in how we deploy" is a story about an owner.

Write the "why it matters" next to each pillar

A pillar list that only says what you want reads as a demand. Next to each pillar, write one line of why it matters to the person's own growth. "Ownership — because the engineers who carry things end-to-end are the ones who get trusted with bigger things." This tiny addition turns the rubric from a stick into a development map, and it is the same document you will later hand someone whose ownership is slipping.