GO TO A SECTION

↑ ↓ to navigate · Enter to openExplore all sections →
FDEInterviews logo
RecentResolveAI technical report

SRE-1: Dominating the Pareto Frontier of Production Investigations with a 27B Model

A 27B open-weight model trained with RL to find root causes of production alerts. Five merged runs matched the best frontier model tested at 37% less cost. How, and what to check.

Breakdown by FDE Interviews · 13 min read · Updated

About the original · 2026
AUTHORS
ResolveAI Labs
FROM
ResolveAI
PUBLISHED
Nov 2026 · ResolveAI technical report
IN THIS BREAKDOWN · 10 SECTIONS

You can read 19% of this breakdown.

01

The paper in one minute

When a production alert fires, someone has to work out what broke and why, quickly, across logs, metrics, traces, code, deploys and tickets. ResolveAI sells an agent that does this investigation, and on 18 November 2026 it published a technical report on SRE-1: a 27-billion-parameter model, post-trained with reinforcement learning from the open Qwen3.8-27B weights, built only for this job.

The model receives an alert and a map of the environment, queries more than 400 tools across about 60 integrations, and answers with a causal chain: this change caused that effect, which caused the alert, each link citing evidence. A judge panel compares the chain with a reference chain and calls a score of 0.8 or more a pass, a threshold calibrated with on-call engineers.

On 49 difficult live production cases, each run five times, one SRE-1 investigation passes 38.8% of the time against 6.9% for the base model. Combining five investigations into one answer raises that to 71%, the same pass rate as Fable 5.2 at its highest effort, the strongest frontier model the report measured, for 37% less cost. The report's own phrase is that SRE-1 "dominates the Pareto frontier" of pass rate against cost: at every budget from its cheapest configuration upward, no model it tested did better.

The parts worth studying are less about the model than about the environment: how to train on incidents that have no written answer, and how to stop a model from finding the answer in commits and tickets written after the fact.

THE IDEA, VISUALLY
1,503 LIVE-TOOL INCIDENTS, FOUR SOURCES (TABLE 4) live incidents own production, read-only 23% incident replays preserved telemetry 34% synthetic built by agent swarms 35% licensed vendor systems 8% blue = real faults; gray = planted faults, answers known WHAT MAKES THE REWARD TRUSTWORTHY time pinning tools see only data from before the alert reference chains multi-agent; experts rated 26 of 27 as good leakage checks 63 of 1,663 candidates excluded for leakage 13 judges one random judge per group of 32 rollouts TRAINING Qwen3.8-27B, modified CISPO with a ratio mask [0.2, 5] 200 steps x 2,048 rollouts; step 110 selected as SRE-1 harness randomized in half the groups; 1/4 offline synthesis 7.4 days on up to 512 GPUs, two clouds 49 LIVE CASES x 5 RUNS: PASS RATE (TABLE 8) base model, 1 run 6.9% SRE-1, 1 run, 1x cost 38.8% SRE-1, 5 merged, 7.3x 71.0% Fable 5.2 xhigh, ~12x 71.0% intervals are wide: 49 cases, not 245 independent runs OFFLINE SYNTHESIS, 73 PROBLEMS (TABLE 3) SRE-1 51.7% Opus 5.5 xhigh 44.5% Fable 5.2 xhigh 39.0% base model 23.5% mean scores barely differ: 0.749 vs 0.743 for Opus 5.5 READ WITH THE REPORT'S OWN LIMITS Test cases over-represent alerts the frontier-backed production system scored lower on. One task only: root cause from an alert, single pass. Ensembles take longer than one investigation. Company technical report, not peer reviewed.
Left: SRE-1's training incidents come from four sources, every tool is pinned to the moment the alert fired, reference answers are written by a multi-agent system, and one of 13 judges scores each group of rollouts. Right: pass rates on 49 live production cases (Table 8) and on offline synthesis of saved evidence (Table 3), with the limits the report states itself.
Research notes: the problem, approach and results
6.9% to 38.8%

single-investigation pass rate, base model to SRE-1, 49 live cases (abstract, Table 8)

71% at -37%

five-run ensemble pass rate and cost against Fable 5.2 at xhigh effort (abstract)

63 of 1,663

candidate incidents excluded for leaking their answer (Table 5)

THE PROBLEM

Investigating every production alert with the strongest frontier models is accurate enough to help but too expensive to run on every alert, and too slow to put on the critical path.

THE IDEA

Post-train an open-weight 27B model with reinforcement learning in realistic, leak-free incident environments, and teach it to synthesize several investigations into one answer.

THE RESULT

One investigation passes 38.8% of hard live cases against 6.9% for the base model; five synthesized investigations pass 71%, matching the strongest frontier model tested at 37% lower cost.

02

Why it matters to an FDE

Three conversations FDEs have with customers run straight through this report. The first is build versus buy. Customers now ask whether they should train their own small model instead of paying for a frontier one, and SRE-1 is the best public evidence that a narrow, high-volume task can justify it. It is also evidence of what that costs: environments, answer keys, leakage checks, a judge panel and a training run on up to 512 GPUs. Our question on training your own model for incident analysis works through that decision.

The second is evaluation. Replaying past incidents is the obvious way to test an investigation agent and the obvious way to fool yourself, because the fix is already written down somewhere. The report describes the leak concretely, and its answer, time pinning, applies to any agent you evaluate on historical data. Our question on replay evaluations that leak the answer builds the fix in code.

The third is judges. The report describes a policy that learned to game its grader in four distinct ways, and the defenses: citation checks, score caps and a randomly drawn judge from 13 models across providers. If you grade agents with an LLM judge, this section is worth reading closely.

Read it as what it is: a company's technical report about its own product, not a peer-reviewed paper. It is unusually specific, it reports its ablations including failures, and it names a limitation of its own test set that favors its model.

THE FULL BREAKDOWN

Understand the mechanism. Try it on a real problem.

Unlock this breakdown and 297 more papers.

Premium opens every paper in full, plus every answer, concept, course and guide on the site. One payment for six months of access.

6 months · One payment · No auto-renewal

Get Premium · ₹2,000 / $25
  • ◇1 architecture diagram
  • </>2 runnable code blocks
  • ▦3 results tables
  1. 03Key takeaways

    Six lessons on environments, leakage, judges and cost that transfer beyond incident response

    Free account
  2. 04How it works

    Follow one training step from incident sampling to judged, filtered, masked updates

    Premium
  3. 05The numbers that matter

    Pass rates, costs and ablations, with the denominator and interval behind each one

    Premium
  4. 06Try it yourself

    Reconcile the run's arithmetic, see why groups go silent, and measure what 49 cases can resolve

    Premium
  5. 07Where it breaks

    Test-set selection, one narrow task, vendor authorship, environment fidelity and frontier drift

    Premium
  6. 08Using it on a deployment

    Evaluate an AI SRE vendor, or plan your own domain model, using this report's checklist

    Premium
  7. 09Interview angles

    Questions on small domain models, leak-free evaluation, judge gaming and ensemble economics

    Premium
Back to the paper library →