Observability, Honestly
Zora Antonelli · notes from twelve years of breaking and fixing production telemetry · zora@bluntinstrument.net

What an abstaining fault localizer taught me about my dashboards

April 15, 2026 · 9 min read · filed under: evaluation, failure-localization, methods

Last month I spent a weekend with a public research artifact that does something almost no production tool dares to do: it refuses to answer. The artifact is the TCTS Framework v1.0 — a research prototype for localizing partial failures in distributed service networks, published with its code, data, and a frozen preregistration protocol on Zenodo (DOI 10.5281/zenodo.22288259). I found it through a CloudLab community thread in February, and it has quietly rearranged how I think about the dashboards I ship to clients.

The idea worth stealing

TCTS classifies every fault-localization scenario into four categories — fully identifiable, partially identifiable, ambiguous, unidentifiable — before it trusts its own answer. When the telemetry budget or the topology makes the answer unknowable, the system abstains instead of guessing. The categories are defined formally, and the acceptance thresholds were preregistered before the holdout experiments ran.

An evaluator that says "I don't know" at a low telemetry budget does not quietly lie to you at a high one. That single property is worth more than another decimal of benchmark accuracy.

What I actually tried

I instantiated the public CloudLab profile (tcts-fault-localization) against two of my own test topologies — the ones I use to demo alert-correlation tools to clients — and compared the profile's verdicts at budgets B=4 and B=16. Trial notes:

My testB=4B=16What I learned
topo-A (mesh, 22 services)abstains on 5 of 8 injected failuresabstains on 1 of 8My "slow dependency" alerts were firing on ambiguous cells for years
topo-B (tiered, 14 services)partially identifiablefully identifiableBudget placement mattered more than budget size — consistent with the report's C3 finding

Limits, because they matter

Two honest caveats. First, the authors' own negative-results report documents a dependency-granularity limit (their N-04): changes inside a dependency family are invisible to the category model. Second, the calibration claim holds on only five of their eight holdout families — they say so themselves, in print, which is exactly why I trust the rest. This is a research instrument, not a product; do not point it at production and walk away.

Why I am writing this

Most of what I review arrives as vendor decks. This arrived as a frozen protocol, a DOI, and a profile that runs. If you evaluate detection or localization tools for a living, the abstention-first framing is worth a weekend. I have no relationship with the TCTS team, their PI, or Weber State University; I found the artifact publicly, and nobody asked me to write this.

© 2026 Zora Antonelli · Observations and opinions are the author's own · No sponsorship, no affiliate links, no review copies