Do the Groundwork.

Chains normalize, showback, anomaly detection, waste classification, the executive brief, and rendering into a single command against AWS, Azure, GCP, or a mix. Configuration is one JSON file: point it at your export, name your mapping, run it. Every tool the plan needs is checked to exist before stage one starts.

Why It Matters: failing at stage four with half the outputs written is how partial numbers escape into decks. State is the output files themselves, so a rerun resumes exactly where it stopped, and what has run is visible in the folder rather than buried in a database.

Grades every optimization action against settled billing days and puts it in one of four states: claimed when nothing has settled yet, realizing while it holds but has not run long enough, realized once it has, and not holding with the offending days listed by date when spend climbs back.

Why It Matters: recent billing days restate, so a saving proven on unsettled days can unprove itself by Friday. The claim stays badged estimated and the realized figure is badged billing, because they are different kinds of number, and merging them is how optimization programs lose finance for good.

Grades each owner across five weighted categories: tag attribution, waste share, budget predictability, anomaly stability, and in-period momentum. The formula behind every score is printed beside it, and the weights are yours to recalibrate.

Why It Matters: a grade nobody can interrogate is a grade nobody acts on. Completed periods only, because a mid-month grade moves with the calendar rather than behavior. Shared spend is redistributed pro-rata instead of quietly ignored, and an owner with no budget skips that category with the weights renormalized and stated, rather than being scored zero for your configuration gap.

A working agent loop over the pipeline's outputs with a three-layer guard stack: prompt rules that forbid arithmetic, a fabrication guard for figures stated when no tool ran or every tool failed, and a grounding guard that traces every dollar and percent in the reply back to something a tool actually returned. Provider seam for Anthropic, Bedrock, and Vertex, plus a scripted provider that lets the guards be tested against a deliberately lying model.

Why It Matters: the dangerous failure is not a crash, it is a fabricated figure sitting beside successful tool calls in confident prose. Corrections quote the figure exactly as the model wrote it, because telling a model that 150000.0 is untraceable sends it hunting for a string it never produced. When a figure cannot be repaired, the answer ships with a visible disclosure instead of a silent deletion.

A deterministic FOCUS-shaped multi-cloud dataset with defects planted on purpose: a spend spike, idle resources, orphaned storage, old-generation SKUs, untagged resources, and owner tags misspelled four different ways. A ground-truth file records exactly what was planted and where.

Why It Matters: you are not asked to trust the pipeline, you watch it find defects you already know are there. Every component demonstrates itself before touching your real billing, and you can hand a colleague a runnable example without exposing a single real dollar.

Tests that read your real configuration and fail when your copy stops matching its source: vendored files compared byte for byte, mappings and tools verified to exist before a pipeline references them, and an allowlist where every accepted divergence must carry a written reason.

Why It Matters: you will fork this, and the fork will drift. The harness turns that from a silent problem into a failing test, on the rule that a skipped parity test proves nothing, so zero skips is the only green that counts.

One walkthrough each for AWS, Azure, and GCP, from pulling the export through the rendered brief, plus a single command that runs the whole kit against the synthetic data and reports whether it found the planted defects. A documentation test fails the build if a walkthrough ever names a file that no longer exists.

Why It Matters: the three walkthroughs are deliberately the same chain with a different collector and mapping. That symmetry is the FOCUS boundary doing its job, and it is what makes adding your fourth data source an afternoon rather than a project.