« Phase 00 · Warmup · Track Overview

Lab 01 — Platform Reference Model & Budget Calculator

The problem

Your Platform Product Owner needs three numbers by Thursday: what SLO can we offer, what does an agent action cost, and what stops a bad action. Today those answers are opinions. By the end of this lab they are functions with tests.

You will build the arithmetic layer that every later phase leans on — availability composition, error budgets, burn-rate alerting, latency budgets, loop reliability, cost — and then the piece that turns arithmetic into architecture: an admission pipeline that runs all five layers' checks against a proposed action and reports every layer that would have denied it.

What you build

Key concepts

Files

Run

pip install -r requirements.txt
pytest test_lab.py -v                       # your lab.py — red until you implement
LAB_MODULE=solution pytest test_lab.py -v   # the reference — must be green
python solution.py                          # the worked example

Success criteria

• All 103 tests green against your lab.py.
• weakest_links is deterministic under ties (sort by unavailability desc, then name asc).
• A budget consumed to exactly its limit reports freeze, not reliability-focus — you used EPSILON, not 0.0.
• A burn rate mathematically equal to 14.4 fires. (1 - 0.999 is 0.0009999999999998899; a naive >= misses the boundary.)
• A parallel latency group contributes its max: the 3-second budget commits 2 370 ms, not 2 490 ms.
• fits_fallback(630) is true and fits_fallback(631) is false — the headroom boundary is inclusive.
• The dual-control check counts distinct approvers and excludes the agent itself.
• The injected-payment scenario returns denials from four distinct layers.

How this maps to the real stack

The honest limits. This lab models steady-state availability. It says nothing about mean time to recovery, correlated multi-hour outages, or the human factors that dominate real incidents. It also assumes every component's availability is known — in practice, getting trustworthy per-component SLI data is most of the work, and the arithmetic is the easy part.

Extensions

1. Make the model a DAG. Replace the flat component list with a graph of nodes and edges, support fan-out (a request that calls three tools in parallel and needs two), and compute availability by enumerating minimal cut sets.
2. Continuous budget consumption. Replace BudgetLedger.consume with a stream of (timestamp, good, total) buckets and compute a rolling-window budget, so policy_state changes over time. Then implement the alert policy against the same buckets.
3. Queueing. Add a utilization field to LatencyStage and inflate its contribution by the M/M/1 factor \( 1/(1-\rho) \) — then watch the budget become infeasible at 80% utilization, which is the real reason capacity planning exists.
4. Measured p per tool. Feed the loop-reliability model from a table of per-tool success rates and compute a task's success probability from its actual tool sequence.
5. Emit evidence. Have AdmissionPipeline.evaluate return an audit record (decision, all denials, the policy inputs, a version stamp) and hash-chain successive records. That is Phase 10 in miniature.

Interview / resume bullets

• "Built the platform's availability and error-budget model: composed a six-component request path, identified retrieval as convertible from a serial to a degradable dependency, and raised modelled read-path availability from 99.10% to 99.66% — 6 h 28 m to 2 h 27 m of monthly downtime — without adding hardware."
• "Derived our alerting thresholds from a stated budget-burn tolerance rather than copying 14.4, and implemented multi-window multi-burn-rate rules so pages require a sustained burn confirmed by a live short window."
• "Introduced cost per successful action as the platform's unit economic, which reframed evaluation investment as a cost-reduction programme: raising task success from 70% to 90% cut effective unit cost 22%."
• "Made 'defence in depth' measurable: the admission path reports every layer that would deny an action, and we require at least two independent denials for any money-moving tool."