3amBench v0.2.0
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitignore +0 -6
- README.md +176 -24
- docs/calibration-status.md +40 -0
- manifest.jsonl +30 -30
- null/afx-e1-s01/solve.sh +3 -0
- null/afx-e1-s02/solve.sh +3 -0
- null/afx-e1-s03/solve.sh +3 -0
- null/afx-e1-s04/solve.sh +3 -0
- null/afx-e1-s05/solve.sh +3 -0
- null/afx-e1-s06/solve.sh +3 -0
- null/afx-e1-s07/solve.sh +3 -0
- null/afx-e1-s08/solve.sh +3 -0
- null/afx-e1-s09/solve.sh +3 -0
- null/afx-e2-s01/solve.sh +3 -0
- null/afx-e2-s02/solve.sh +3 -0
- null/afx-e2-s03/solve.sh +3 -0
- null/afx-e2-s04/solve.sh +3 -0
- null/afx-e2-s05/solve.sh +3 -0
- null/afx-e2-s06/solve.sh +3 -0
- null/afx-e2-s07/solve.sh +3 -0
- null/afx-e2-s08/solve.sh +3 -0
- null/afx-e2-s09/solve.sh +3 -0
- null/afx-e3-s01/solve.sh +3 -0
- null/afx-e3-s02/solve.sh +3 -0
- null/afx-e3-s03/solve.sh +3 -0
- null/afx-e3-s04/solve.sh +3 -0
- null/afx-e3-s05/solve.sh +3 -0
- null/afx-e3-s06/solve.sh +3 -0
- null/afx-e3-s07/solve.sh +3 -0
- null/afx-e3-s08/solve.sh +3 -0
- null/afx-e3-s09/solve.sh +3 -0
- null/afx-e4-s01/solve.sh +3 -0
- null/afx-e4-s02/solve.sh +3 -0
- null/afx-e4-s03/solve.sh +3 -0
- null/afx-e4-s04/solve.sh +3 -0
- null/afx-e4-s05/solve.sh +3 -0
- null/afx-e4-s06/solve.sh +3 -0
- null/afx-e4-s07/solve.sh +3 -0
- null/afx-e4-s08/solve.sh +3 -0
- null/afx-e4-s09/solve.sh +3 -0
- null/afx-e5-s01/solve.sh +3 -0
- null/afx-e5-s02/solve.sh +3 -0
- null/afx-e5-s03/solve.sh +3 -0
- null/afx-e5-s04/solve.sh +3 -0
- null/afx-e5-s05/solve.sh +3 -0
- null/afx-e5-s06/solve.sh +3 -0
- null/afx-e5-s07/solve.sh +3 -0
- null/afx-e5-s08/solve.sh +3 -0
- null/afx-e5-s09/solve.sh +3 -0
- openenv/alertforge_env/README.md +1 -1
.gitignore
CHANGED
|
@@ -4,10 +4,4 @@ __pycache__/
|
|
| 4 |
.venv/
|
| 5 |
.DS_Store
|
| 6 |
dist/
|
| 7 |
-
.claude/
|
| 8 |
-
.claude-flow/
|
| 9 |
-
CLAUDE.md
|
| 10 |
-
.mcp.json
|
| 11 |
-
.swarm/
|
| 12 |
-
.hive-mind/
|
| 13 |
*.egg-info/
|
|
|
|
| 4 |
.venv/
|
| 5 |
.DS_Store
|
| 6 |
dist/
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
*.egg-info/
|
README.md
CHANGED
|
@@ -28,49 +28,162 @@ atomic checks, so the reward is dense.
|
|
| 28 |
|
| 29 |
| | |
|
| 30 |
|---|---|
|
| 31 |
-
| Tasks | 30 (5 workflows × 3 tiers × 2 seeds), plus unlimited fresh ones from the generator / OpenEnv server |
|
| 32 |
-
| Checks per task | 59-189 (easy 59-67, medium 112-139, hard 167-189) |
|
| 33 |
-
| Reward | `reward` in [0, 1] plus 15 diagnostic keys, identical
|
| 34 |
-
| Grader | promtool 3.5.0 + amtool 0.28.1 + stdlib Python + PyYAML; 0.7-7 s per grade
|
| 35 |
-
| Formats | Harbor tasks (this repo) and an OpenEnv server with per-step reward (`openenv/alertforge_env`) |
|
| 36 |
|
| 37 |
## Quick start
|
| 38 |
|
| 39 |
```bash
|
| 40 |
# oracle (expect mean 1.0) and no-op (expect 0.0)
|
| 41 |
-
uvx harbor run --repo https://huggingface.co/datasets/openenvforge/3ambench@v0.
|
| 42 |
-
uvx harbor run --repo https://huggingface.co/datasets/openenvforge/3ambench@v0.
|
| 43 |
# your model on the 5-task smoke set
|
| 44 |
-
uvx harbor run --repo https://huggingface.co/datasets/openenvforge/3ambench@v0.
|
|
|
|
|
|
|
| 45 |
```
|
| 46 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
Tasks set `[agent] network_mode = "no-network"`, because this repo ships the solutions (see Anti-hacking).
|
| 48 |
Harbor enforces that on Linux Docker hosts and sandboxed providers. **Docker Desktop on macOS cannot enforce it**
|
| 49 |
(its VM kernel lacks `CONFIG_NFT_FIB_INET`), and Harbor refuses the task. For local smoke runs on a Mac, use
|
| 50 |
`scripts/harbor_local.sh`, which runs a temp copy without the network override. Those runs are not leaderboard-safe.
|
| 51 |
|
| 52 |
-
**OpenEnv (multi-turn, per-step reward)
|
|
|
|
| 53 |
```python
|
| 54 |
from alertforge_env.client import AlertForgeEnv
|
| 55 |
-
with AlertForgeEnv(base_url="
|
| 56 |
obs = env.reset(seed=7, split="train", workflow="missed-page-postmortem", tier="medium")
|
| 57 |
obs = env.step({"tool": "read_file", "path": "README.md"})
|
| 58 |
obs = env.step({"tool": "write_file", "path": "rules/service-health.yml", "content": "..."})
|
| 59 |
print(obs.reward, obs.done) # parity mode: rewards sum to the Harbor reward
|
| 60 |
```
|
| 61 |
|
| 62 |
-
**Fresh tasks (training, private held-out sets):** `python generator/generate.py --master-seed <secret>`
|
|
|
|
| 63 |
|
| 64 |
## Try it
|
| 65 |
|
| 66 |
- **Replay runs in the browser:** [3amBench Replay](https://huggingface.co/spaces/openenvforge/3ambench-replay)
|
| 67 |
-
steps through recorded runs (tool calls, file diffs, the reward curve,
|
|
|
|
| 68 |
Its data comes from `scripts/record_reference_runs.py`, `scripts/record_agent_runs.sh` and `scripts/export_runs.py`;
|
| 69 |
the page itself is in `space/`.
|
| 70 |
- **Run the environment locally** and open the web interface at http://localhost:8000/web:
|
| 71 |
```bash
|
| 72 |
-
docker run -p 8000:8000 -e ENABLE_WEB_INTERFACE=true ghcr.io/devesh-maheshwari/3ambench-env:0.
|
| 73 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 74 |
|
| 75 |
## What a task looks like
|
| 76 |
|
|
@@ -120,8 +233,9 @@ missing `continue` gives Jaccard ½, and a correct burn alert on top of broken S
|
|
| 120 |
| `req_alerts`, `req_repairs`, `req_recording`, `req_routing`, `req_inhibit` | per-category scores |
|
| 121 |
| `fire_rate`, `silent_rate`, `label_rate`, `check_pass_rate` | raw diagnostic pass rates (non-zero for a no-op) |
|
| 122 |
| `syntax_ok`, `tamper` | all files load / a forbidden construct was used (tamper zeroes the reward) |
|
|
|
|
| 123 |
|
| 124 |
-
## Reward spread (measured
|
| 125 |
|
| 126 |
| Policy | reward | notes |
|
| 127 |
|---|---|---|
|
|
@@ -136,8 +250,7 @@ Adversaries (all 30 tasks, local gate): always-fire, rename, catch-all route, in
|
|
| 136 |
Receiver nulling, over-broad inhibition, band thresholds, per-pod alerts, `for`-only fixes, deleting untouched
|
| 137 |
rules → `outcome` 0 and reward ≤ 0.4.
|
| 138 |
|
| 139 |
-
Real-model
|
| 140 |
-
failure gallery, dense-vs-outcome GRPO curve).
|
| 141 |
|
| 142 |
## How it was built: Skill2Env, but procedural
|
| 143 |
|
|
@@ -172,9 +285,9 @@ equivalent of Skill2Env's freeze boundary.
|
|
| 172 |
14.4x burn-rate rule, and `slo-architect` covers burn-rate alerting. We downloaded ten SRE-adjacent Skill2Env
|
| 173 |
tasks. The closest one (`task_slo-architect_lgre556n`) grades burn-rate policy by comparing YAML fields, and none
|
| 174 |
run promtool or amtool.
|
| 175 |
-
- [camel-ai/seta-env](https://github.com/camel-ai/seta-env) task 1114: Alertmanager routing and inhibition
|
| 176 |
-
|
| 177 |
-
carries.
|
| 178 |
- Community rule sets: [samber/awesome-prometheus-alerts](https://github.com/samber/awesome-prometheus-alerts),
|
| 179 |
[kubernetes-mixin](https://github.com/kubernetes-monitoring/kubernetes-mixin); Google SRE Workbook ch. 5.
|
| 180 |
|
|
@@ -185,17 +298,56 @@ equivalent of Skill2Env's freeze boundary.
|
|
| 185 |
Unsupported constructs fail closed.
|
| 186 |
- Seven alert templates host the defects, not a large upstream corpus.
|
| 187 |
- The routing chain case uses the agent's static labels, not labels observed in `ALERTS`.
|
| 188 |
-
-
|
|
|
|
| 189 |
- **Contamination:** SkillHub skills (and any skill-augmented agent) already contain near-identical burn-rate
|
| 190 |
recipes, and this repo ships solutions. Evaluate on a private-seed build, and report whether the agent had
|
| 191 |
`SKILL.md`.
|
| 192 |
- English only.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 193 |
|
| 194 |
## Repository layout
|
| 195 |
|
| 196 |
-
`tasks/` (Harbor tasks
|
| 197 |
-
|
| 198 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 199 |
|
| 200 |
## License and citation
|
| 201 |
|
|
@@ -205,6 +357,6 @@ Apache-2.0; see `NOTICE.md` for attribution of adapted community rules (CC BY 4.
|
|
| 205 |
@misc{3ambench2026,
|
| 206 |
title = {3amBench: a behaviorally graded, dense-reward RL environment for Prometheus alerting as code},
|
| 207 |
year = {2026},
|
| 208 |
-
note = {Harbor dataset and OpenEnv environment; generator package alertforge v0.
|
| 209 |
}
|
| 210 |
```
|
|
|
|
| 28 |
|
| 29 |
| | |
|
| 30 |
|---|---|
|
| 31 |
+
| Tasks | 30 core (5 workflows × 3 tiers × 2 seeds) and 45 expert (5 families × 9 seeds, see [Expert tier](#expert-tier-v02)), plus unlimited fresh ones from the generator / OpenEnv server |
|
| 32 |
+
| Checks per task | 59-189 (easy 59-67, medium 112-139, hard 167-189); expert 128-232 |
|
| 33 |
+
| Reward | `reward` in [0, 1] plus 15 diagnostic keys (core) or 26 (expert), identical within a tier |
|
| 34 |
+
| Grader | promtool 3.5.0 + amtool 0.28.1 + stdlib Python + PyYAML; deterministic; core 0.7-7 s, expert about 0.5-12 s per grade (one process) |
|
| 35 |
+
| Formats | Harbor tasks (this repo has the 30 core tasks; the Hugging Face dataset has all 75) and an OpenEnv server with per-step reward for the core tier (`openenv/alertforge_env`) |
|
| 36 |
|
| 37 |
## Quick start
|
| 38 |
|
| 39 |
```bash
|
| 40 |
# oracle (expect mean 1.0) and no-op (expect 0.0)
|
| 41 |
+
uvx harbor run --repo https://huggingface.co/datasets/openenvforge/3ambench@v0.2.0 -d 3ambench@0.2.0 -a oracle -n 4
|
| 42 |
+
uvx harbor run --repo https://huggingface.co/datasets/openenvforge/3ambench@v0.2.0 -d 3ambench@0.2.0 -a nop -n 4
|
| 43 |
# your model on the 5-task smoke set
|
| 44 |
+
uvx harbor run --repo https://huggingface.co/datasets/openenvforge/3ambench@v0.2.0 -d 3ambench-mini@0.2.0 -a terminus-2 -m <model>
|
| 45 |
+
# expert tier: the 15 leaderboard tasks (all 45: -d 3ambench-expert@0.2.0)
|
| 46 |
+
uvx harbor run --repo https://huggingface.co/datasets/openenvforge/3ambench@v0.2.0 -d 3ambench-expert-lb@0.2.0 -a terminus-2 -m <model>
|
| 47 |
```
|
| 48 |
|
| 49 |
+
The core tasks in `3ambench@0.2.0` are the `3ambench@0.1.1` tasks with their fictional company domains moved to
|
| 50 |
+
reserved `.example` names (and the version stamp); checks, scenarios and grader are unchanged, and every recorded
|
| 51 |
+
v0.1.1 run regrades to the same reward, so v0.1.1 numbers stay comparable.
|
| 52 |
+
|
| 53 |
Tasks set `[agent] network_mode = "no-network"`, because this repo ships the solutions (see Anti-hacking).
|
| 54 |
Harbor enforces that on Linux Docker hosts and sandboxed providers. **Docker Desktop on macOS cannot enforce it**
|
| 55 |
(its VM kernel lacks `CONFIG_NFT_FIB_INET`), and Harbor refuses the task. For local smoke runs on a Mac, use
|
| 56 |
`scripts/harbor_local.sh`, which runs a temp copy without the network override. Those runs are not leaderboard-safe.
|
| 57 |
|
| 58 |
+
**OpenEnv (multi-turn, per-step reward)**, against a local server started as in [Try it](#try-it). The client
|
| 59 |
+
needs `pip install "openenv-core==0.3.0"` and a clone of this repo on `PYTHONPATH=openenv`:
|
| 60 |
```python
|
| 61 |
from alertforge_env.client import AlertForgeEnv
|
| 62 |
+
with AlertForgeEnv(base_url="http://localhost:8000").sync() as env:
|
| 63 |
obs = env.reset(seed=7, split="train", workflow="missed-page-postmortem", tier="medium")
|
| 64 |
obs = env.step({"tool": "read_file", "path": "README.md"})
|
| 65 |
obs = env.step({"tool": "write_file", "path": "rules/service-health.yml", "content": "..."})
|
| 66 |
print(obs.reward, obs.done) # parity mode: rewards sum to the Harbor reward
|
| 67 |
```
|
| 68 |
|
| 69 |
+
**Fresh tasks (training, private held-out sets):** `python generator/generate.py --master-seed <secret> --out <dir>`
|
| 70 |
+
(needs promtool and amtool; without `--out` it overwrites this repo's tasks).
|
| 71 |
|
| 72 |
## Try it
|
| 73 |
|
| 74 |
- **Replay runs in the browser:** [3amBench Replay](https://huggingface.co/spaces/openenvforge/3ambench-replay)
|
| 75 |
+
steps through recorded runs (tool calls, file diffs, the reward curve, which checks pass after each step, and
|
| 76 |
+
for expert runs which tickets were solved).
|
| 77 |
Its data comes from `scripts/record_reference_runs.py`, `scripts/record_agent_runs.sh` and `scripts/export_runs.py`;
|
| 78 |
the page itself is in `space/`.
|
| 79 |
- **Run the environment locally** and open the web interface at http://localhost:8000/web:
|
| 80 |
```bash
|
| 81 |
+
docker run -p 8000:8000 -e ENABLE_WEB_INTERFACE=true ghcr.io/devesh-maheshwari/3ambench-env:0.2.0
|
| 82 |
```
|
| 83 |
+
The image is built for linux/amd64 and linux/arm64. To build it yourself from a clone:
|
| 84 |
+
`docker build -f openenv/alertforge_env/server/Dockerfile -t 3ambench-env .` (then run `3ambench-env`).
|
| 85 |
+
|
| 86 |
+
## Leaderboard (core, 6 tasks: af-001, af-009, af-013, af-017, af-021, af-029)
|
| 87 |
+
|
| 88 |
+
<!-- LEADERBOARD:core -->
|
| 89 |
+
| Agent | Model | Runs | Mean reward | Solved |
|
| 90 |
+
|---|---|---|---|---|
|
| 91 |
+
| claude-code | claude-opus-5-5 | 12 | 1.000 | 12/12 |
|
| 92 |
+
| codex | gpt-6-astra | 12 | 1.000 | 12/12 |
|
| 93 |
+
| terminus-2 | gpt-oss-120b | 13 | 0.436 | 3/13 |
|
| 94 |
+
|
| 95 |
+
37 runs on 6 tasks.
|
| 96 |
+
<!-- /LEADERBOARD -->
|
| 97 |
+
|
| 98 |
+
The runs were recorded on the v0.1.1 tasks; their saved workspaces, read with the renamed domains, regrade to the
|
| 99 |
+
same rewards on the 0.2.0 tasks. Local Harbor runs on Docker Desktop (agent network enabled, so not sealed). Frontier agents saturate this subset;
|
| 100 |
+
the dense reward separates weaker models (gpt-oss-120b scores between 0.03 and 0.37 on every unsolved run instead of 0).
|
| 101 |
+
v0.2 adds the expert tier below. Every run can be replayed step by step in the
|
| 102 |
+
[replay Space](https://huggingface.co/spaces/openenvforge/3ambench-replay).
|
| 103 |
+
|
| 104 |
+
## Expert tier (v0.2)
|
| 105 |
+
|
| 106 |
+
A core task is a short list of change requests on a small repo. An expert task is a week of a team's
|
| 107 |
+
observability queue on a large legacy one: 45 tasks, five families of nine seeds, each at a different fictional
|
| 108 |
+
company with its own services, teams and history. The starting repo has 62-85 alerting rules in 14-22 files, SLI
|
| 109 |
+
recording rules, an SLO catalog, an ownership file, ADRs, runbooks, a CI script, and an Alertmanager config shared
|
| 110 |
+
with two other Prometheus servers (41-53 receivers, legacy `match:` routes next to `matchers:`). The queue holds two
|
| 111 |
+
or three symptom tickets and two to seven routine requests.
|
| 112 |
+
|
| 113 |
+
| Family | The week |
|
| 114 |
+
|---|---|
|
| 115 |
+
| E1 quiet pager | Nobody was paged: an outage where the Down alert could not fire once the targets went away, an error-ratio page that never fired, a replica-lag alert stuck in pending, a postmortem action item asking for a traffic-floor page |
|
| 116 |
+
| E2 storm after a deploy | Too many pages: one per pod during a database failover, false pages from canaries and rollouts, a cluster outage that paged every service owner |
|
| 117 |
+
| E3 reorg + routing migration | A team split that routing never caught up with: alerts that still carry the old team, a team with no route, SRE still everyone's first pager, a channel being archived, warnings muted by unrelated criticals |
|
| 118 |
+
| E4 latency SLO rollout | A latency SLO going live: SLI recording rules that are wrong during a rollout, slow requests that never page, burn alerts to add |
|
| 119 |
+
| E5 legacy cleanup after an on-call survey | The work arrives as raw answers to an on-call survey: a consumer stall that never fired, a nightly batch that stopped without a ticket, a disk prediction that pages for the backup job, dead runbook links, a legacy service to remove. Some answers need nothing from this repo, and some alerts have to stay as they are |
|
| 120 |
+
|
| 121 |
+
What makes it realistic:
|
| 122 |
+
|
| 123 |
+
- Tickets read like tickets. A symptom ticket says what happened in the reporter's words ("17 pages for one
|
| 124 |
+
database failover", "KafkaConsumerStalled has never fired") and points at an incident packet: PagerDuty and Slack
|
| 125 |
+
exports, a postmortem, `query_range` pulls of the night that `bin/range2test` turns into promtool tests. It never
|
| 126 |
+
says which rule is wrong.
|
| 127 |
+
- Policy lives in the repo, not in the ticket: SLOs and windows in the catalog, the paging policy in the README,
|
| 128 |
+
ownership in `teams/ownership.yaml`, the burn-rate policy in an ADR. A new hire would have to look there too.
|
| 129 |
+
- The repo has a past: two routing syntaxes, a team split half done, rules that moved with their team, decoy alerts
|
| 130 |
+
that look broken and aren't, alerts other teams own.
|
| 131 |
+
- No two tasks share a company, an org chart or a queue, so a fix can't be copied from a sibling task.
|
| 132 |
+
|
| 133 |
+
**Grading** is behavioral and dense, as in the core tier. Promtool replays every hidden scenario group once (plus a
|
| 134 |
+
re-run where a deadline window needs one) and the grader reads which alerts fire, minute by minute; routing goes through amtool and inhibition through the strict
|
| 135 |
+
Alertmanager simulator. A symptom ticket is an outcome check: when the incident replays, the right page reaches the
|
| 136 |
+
right team, and it stays quiet on normal traffic, deploys, flaps and the noise that started the ticket. Alerts are
|
| 137 |
+
counted after inhibition, so muting everything doesn't make a quiet pager look fixed. Routine requests (routes,
|
| 138 |
+
inhibitions, label moves, burn alerts, recording rules, annotations, legacy removals) are checked the same way.
|
| 139 |
+
Untouched rules, receivers and routes have to keep working; a rewritten but equivalent rule is replayed, not
|
| 140 |
+
compared as text. Each requirement scores s in [-0.25, 1] against the starting repo, and `reward` = 0.6 · solved +
|
| 141 |
+
0.4 · progress as in [Reward](#reward); expert progress also loses 0.75 weight units for each leave-as-is alert the
|
| 142 |
+
run changes, and `solved` requires all of them intact. A symptom ticket weighs 5-10 against 2-3 for a routine request, so the
|
| 143 |
+
tickets carry 55-82% of a task's weight. `diagnosis` is the weighted mean s over the symptom tickets and `routine`
|
| 144 |
+
the same over the rest: among runs that don't solve a task, they tell finding the cause apart from doing the easy
|
| 145 |
+
part of the queue. Expert tasks have 128-232 checks and the same anti-hacking defenses, plus tamper checks on reads
|
| 146 |
+
of `ALERTS`.
|
| 147 |
+
|
| 148 |
+
**Calibration, honestly:** frontier agents solve most expert tasks, in minutes (3-5 minutes per task in our pilot
|
| 149 |
+
runs). The tier is not built to rank frontier models against each other. It is aimed at separating mid-size and
|
| 150 |
+
open models, where the dense reward and the diagnosis/routine split show how far a run got, and at RL training on a
|
| 151 |
+
realistic repo. A harder frontier tier is in development; there is no frontier-hardness result yet.
|
| 152 |
+
|
| 153 |
+
### Leaderboard (v0.2.0, expert: the 15 tasks of `3ambench-expert-lb@0.2.0`, 3 per family)
|
| 154 |
+
|
| 155 |
+
<!-- LEADERBOARD:expert -->
|
| 156 |
+
| Agent | Model | Runs | Tasks | Mean reward | Solved | Diagnosis | Routine |
|
| 157 |
+
|---|---|---|---|---|---|---|---|
|
| 158 |
+
| claude-code | claude-opus-5-5 | 14 | 14/15 | 1.000 | 14/14 | 1.000 | 1.000 |
|
| 159 |
+
| codex | gpt-6-astra | 15 | 15/15 | 0.960 | 14/15 | 1.000 | 1.000 |
|
| 160 |
+
| claude-code | claude-fable-5-1 | 14 | 14/15 | 0.956 | 13/14 | 1.000 | 1.000 |
|
| 161 |
+
| codex | gpt-5.6-sol | 15 | 15/15 | 0.912 | 13/15 | 0.967 | 1.000 |
|
| 162 |
+
| claude-code | claude-sonnet-5 | 14 | 14/15 | 0.762 | 9/14 | 0.903 | 1.000 |
|
| 163 |
+
| claude-code | claude-haiku-4-5-20251001 | 14 | 14/15 | 0.346 | 2/14 | 0.473 | 0.927 |
|
| 164 |
+
| terminus-2 | gpt-oss-120b | 10 | 10/15 | 0.065 | 0/10 | 0.071 | 0.345 |
|
| 165 |
+
|
| 166 |
+
96 runs on 15 tasks; 2 runs flagged invalid are not counted.
|
| 167 |
+
<!-- /LEADERBOARD -->
|
| 168 |
+
|
| 169 |
+
The 15 tasks are `3ambench-expert-lb@0.2.0`: the development pilot set (3 per family,
|
| 170 |
+
`scripts/expert_calibrate.py --pilot`), not a calibrated or held-out selection. These are unbalanced local
|
| 171 |
+
development runs on Docker Desktop (agent network enabled, so not sealed), not a held-out model comparison, and the
|
| 172 |
+
rows cover different tasks (the Tasks column), so compare them with care. The two historical `afx-e5-s08` trials
|
| 173 |
+
(Opus and Astra) are excluded because their survey evidence differs from the current task. Astra and Sol were re-run
|
| 174 |
+
on the current task, so they cover all 15 tasks; Opus, Fable, Sonnet and Haiku have no valid `afx-e5-s08` run yet
|
| 175 |
+
and cover 14. gpt-oss-120b covers 10: the provider budget cap stopped it before afx-e2-s03, afx-e3-s04, afx-e4-s04
|
| 176 |
+
and afx-e5-s03, and it has no `afx-e5-s08` run. Every saved output was regraded with the released task's own
|
| 177 |
+
grader. Six runs score higher than Harbor's verifier gave them at the time, because of the grader fixes listed in
|
| 178 |
+
the CHANGELOG; the final 0.2.0 changes (reserved domains, ingress series, the two grader fixes) change no recorded
|
| 179 |
+
score. Infrastructure failures are excluded when present. Every run, with the
|
| 180 |
+
tickets it solved, can be replayed in the [replay Space](https://huggingface.co/spaces/openenvforge/3ambench-replay).
|
| 181 |
+
|
| 182 |
+
To build the 45 tasks yourself: `pip install .` from a clone, then `alertforge expert --out <dir> --seeds 1-9`
|
| 183 |
+
(promtool and amtool on the path, or in `AF_PROMTOOL` and `AF_AMTOOL`). It runs each task's acceptance gate as well,
|
| 184 |
+
and it reproduces the released tasks byte for byte.
|
| 185 |
+
|
| 186 |
+
See [calibration status](docs/calibration-status.md) for the calibration protocol and what is still open.
|
| 187 |
|
| 188 |
## What a task looks like
|
| 189 |
|
|
|
|
| 233 |
| `req_alerts`, `req_repairs`, `req_recording`, `req_routing`, `req_inhibit` | per-category scores |
|
| 234 |
| `fire_rate`, `silent_rate`, `label_rate`, `check_pass_rate` | raw diagnostic pass rates (non-zero for a no-op) |
|
| 235 |
| `syntax_ok`, `tamper` | all files load / a forbidden construct was used (tamper zeroes the reward) |
|
| 236 |
+
| `diagnosis`, `routine` | expert tier only: weighted mean `s` over the symptom tickets, and over the rest of the queue. Tickets carry at least 55% of an expert task's weight, so among runs that don't solve a task these say whether the progress came from diagnosing or from routine edits |
|
| 237 |
|
| 238 |
+
## Reward spread (measured on the v0.1.0 tasks, before the 0.1.1 fairness fixes)
|
| 239 |
|
| 240 |
| Policy | reward | notes |
|
| 241 |
|---|---|---|
|
|
|
|
| 250 |
Receiver nulling, over-broad inhibition, band thresholds, per-pod alerts, `for`-only fixes, deleting untouched
|
| 251 |
rules → `outcome` 0 and reward ≤ 0.4.
|
| 252 |
|
| 253 |
+
Real-model results are in the two leaderboards above; a dense-vs-outcome GRPO training curve is not published yet.
|
|
|
|
| 254 |
|
| 255 |
## How it was built: Skill2Env, but procedural
|
| 256 |
|
|
|
|
| 285 |
14.4x burn-rate rule, and `slo-architect` covers burn-rate alerting. We downloaded ten SRE-adjacent Skill2Env
|
| 286 |
tasks. The closest one (`task_slo-architect_lgre556n`) grades burn-rate policy by comparing YAML fields, and none
|
| 287 |
run promtool or amtool.
|
| 288 |
+
- [camel-ai/seta-env](https://github.com/camel-ai/seta-env) task 1114: an Alertmanager routing and inhibition task. Its
|
| 289 |
+
tests run amtool only for `check-config` and check routes and inhibitions by reading the YAML. 3amBench's routing
|
| 290 |
+
checks are end-to-end: the chain case routes the labels the agent's own alert carries.
|
| 291 |
- Community rule sets: [samber/awesome-prometheus-alerts](https://github.com/samber/awesome-prometheus-alerts),
|
| 292 |
[kubernetes-mixin](https://github.com/kubernetes-monitoring/kubernetes-mixin); Google SRE Workbook ch. 5.
|
| 293 |
|
|
|
|
| 298 |
Unsupported constructs fail closed.
|
| 299 |
- Seven alert templates host the defects, not a large upstream corpus.
|
| 300 |
- The routing chain case uses the agent's static labels, not labels observed in `ALERTS`.
|
| 301 |
+
- Frontier agents solved every run on the two easy tasks measured (af-001, af-013), so the easy tier likely
|
| 302 |
+
saturates for them.
|
| 303 |
- **Contamination:** SkillHub skills (and any skill-augmented agent) already contain near-identical burn-rate
|
| 304 |
recipes, and this repo ships solutions. Evaluate on a private-seed build, and report whether the agent had
|
| 305 |
`SKILL.md`.
|
| 306 |
- English only.
|
| 307 |
+
- Expert tier (v0.2 candidates): hidden scenarios replay through promtool, whose samples sit exactly on the
|
| 308 |
+
evaluation ticks. A target that leaves service discovery is marked stale on the next sample; a real
|
| 309 |
+
Prometheus 3.5 server writes those stale markers about two scrape intervals after the last scrape, so an
|
| 310 |
+
`absent()` Down alert pages 1-2 minutes later in production than in the replay. The graders accept
|
| 311 |
+
`absent()` Down alerts with `for` up to 10m, which in production page 11-13 minutes after the targets went
|
| 312 |
+
away, against the 10 minutes the task README asks for. Window corners that depend on the scrape and
|
| 313 |
+
evaluation phases are graded as promtool sees them: `[90s]` on a 1m scrape passes in the replay, but on a
|
| 314 |
+
real server it is empty for about half of all phase pairs.
|
| 315 |
+
- Expert tier: when a counter stops (consumer offsets, requests), a stated deadline is counted from the last
|
| 316 |
+
sample that still increased, the earliest the stop can have happened. When a series goes away, it is counted
|
| 317 |
+
from the minute its stale marker is written (see above for how that differs from a real server).
|
| 318 |
+
- Expert tier: a routing fix that sits below an unfixed catch-all route earns nothing on its own route checks,
|
| 319 |
+
because the catch-all takes the alerts first. That is Alertmanager's first-match routing, kept on purpose: a
|
| 320 |
+
run that misses the catch-all also loses the route tickets behind it.
|
| 321 |
+
- Expert tier: `cluster` is a target label on `up` only; request counters and histograms in the hidden data
|
| 322 |
+
don't carry it.
|
| 323 |
+
- Expert tier: every hidden scenario holds what the world's targets export at healthy levels: request counters for
|
| 324 |
+
every service whose pods are up; the histograms, pool gauges and cAdvisor series the repo's rules read about those
|
| 325 |
+
services; the ingress controller's request counters (`nginx_ingress_controller_requests`) for every HTTP service;
|
| 326 |
+
and the exporter series a ticket needs. Targets that go down or leave discovery get stale markers. The ingress
|
| 327 |
+
counts the same requests and status codes as the app's own counters, and while a service has no target up it
|
| 328 |
+
answers 502 (targets down) or 503 (none left) for the traffic clients keep sending. Not simulated: the ingress
|
| 329 |
+
latency histogram and config-reload gauge, and fleet exporters (node, Postgres, Redis, Kafka, JVM) beyond those a
|
| 330 |
+
ticket needs. The repo's fleet alerts and its ingress latency and reload alerts therefore never fire in the replay,
|
| 331 |
+
and an alert that reads only those series cannot be graded on them. A scenario still starts without history: at
|
| 332 |
+
its first evaluation `rate()` has no value yet, so a floor rule that reads a missing counter as zero traffic sees
|
| 333 |
+
zero there, and its `for` is what keeps it quiet.
|
| 334 |
+
- Expert tier: an untouched alert counts as intact when it behaves the same. Its text may change (formatting,
|
| 335 |
+
operand order, a static label equal to what its expression already yields, an identical second copy) as long as
|
| 336 |
+
its `for`, thresholds, windows and functions stay and it fires with the same labels at every minute of every
|
| 337 |
+
hidden scenario. An alert that fires in none of them can only be kept as written.
|
| 338 |
+
- Expert tier: an alert needs the labels the task's README names, with their values; it may carry others, which are
|
| 339 |
+
judged by where the alert is delivered, not by their presence.
|
| 340 |
|
| 341 |
## Repository layout
|
| 342 |
|
| 343 |
+
`tasks/` (Harbor tasks: this repository holds the 30 core tasks; the Hugging Face dataset adds the 45 expert `afx-*`
|
| 344 |
+
tasks that `registry.json` also lists, so run those with `--repo https://huggingface.co/datasets/openenvforge/3ambench@v0.2.0`
|
| 345 |
+
or build them with `alertforge expert`) · `registry.json` (`3ambench`, `3ambench-mini`, `3ambench-expert`,
|
| 346 |
+
`3ambench-expert-lb`, all 0.2.0) · `manifest.jsonl` (one row per core task) · `partial/`, `null/` (reference
|
| 347 |
+
policies) · `skill/`, `workflows/` · `generator/generate.py` + `src/alertforge/` (the generator and grader source;
|
| 348 |
+
`src/alertforge/expert/` builds the expert tier) · `openenv/alertforge_env/` (OpenEnv server) · `space/` (static
|
| 349 |
+
replay viewer) · `scripts/` (`export_runs.py` and `leaderboard.py` produce the replay data and the leaderboard tables,
|
| 350 |
+
from a GitHub clone; the dataset does not ship `space/`) · `tests/` (pytest).
|
| 351 |
|
| 352 |
## License and citation
|
| 353 |
|
|
|
|
| 357 |
@misc{3ambench2026,
|
| 358 |
title = {3amBench: a behaviorally graded, dense-reward RL environment for Prometheus alerting as code},
|
| 359 |
year = {2026},
|
| 360 |
+
note = {Harbor dataset and OpenEnv environment; generator package alertforge v0.2.0}
|
| 361 |
}
|
| 362 |
```
|
docs/calibration-status.md
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Calibration status (v0.2.0)
|
| 2 |
+
|
| 3 |
+
The expert tier separates mid-size and open models clearly, but it does not separate frontier agents: they solve
|
| 4 |
+
almost every expert task, usually in a few minutes. v0.2.0 makes no frontier-difficulty claim.
|
| 5 |
+
|
| 6 |
+
## What the v0.2.0 leaderboard is
|
| 7 |
+
|
| 8 |
+
- One attempt per task and model on the 15 tasks of `3ambench-expert-lb@0.2.0` (3 per family), run locally through
|
| 9 |
+
Harbor on Docker Desktop with the agent's network enabled. These are development runs, not a sealed, held-out
|
| 10 |
+
comparison.
|
| 11 |
+
- Every stored run was graded with the release grader. The fictional company domains were renamed to reserved
|
| 12 |
+
`.example` names after the runs; stored workspaces and trajectories are read through that rename, and all 171
|
| 13 |
+
stored runs (96 expert, 73 core, plus 2 excluded) regrade to the same scores they had before it.
|
| 14 |
+
- Coverage differs by model and the table says so: two `afx-e5-s08` attempts (Opus and Astra) are excluded because
|
| 15 |
+
that task's survey wording was corrected after they ran; Astra and Sol were rerun on the corrected task, the other
|
| 16 |
+
models were not. gpt-oss-120b ran on 10 tasks because its hosted API had a spending cap.
|
| 17 |
+
- Every failure in these runs was triaged before it was counted. Failures caused by the task or the grader
|
| 18 |
+
(an unstated requirement, scenario data that punished a valid fix, ambiguous wording) were fixed in the
|
| 19 |
+
generator or grader and the stored runs regraded; only genuine model mistakes count as failures.
|
| 20 |
+
|
| 21 |
+
## Rules for counting results
|
| 22 |
+
|
| 23 |
+
- Mechanical acceptance gates come before any model measurement: the reference solution scores 1, an alternative
|
| 24 |
+
valid solution scores 1, doing nothing scores 0, a partial solution lands in its band, and every scripted cheat
|
| 25 |
+
stays in its band.
|
| 26 |
+
- A lower score caused by ambiguity, grader errors, missing evidence or infrastructure is not a model failure and not
|
| 27 |
+
a hardening result.
|
| 28 |
+
- `scripts/expert_calibrate.py` requires a balanced pilot before any difficulty verdict: two declared models, three
|
| 29 |
+
preselected tasks per family and three valid attempts per model and task. Missing cells produce `collect`, not a
|
| 30 |
+
verdict. Duplicate and invalid runs never count as extra attempts, and historical results are regraded against
|
| 31 |
+
the current task artifacts before use. Report models separately: a pooled average can hide saturation by the
|
| 32 |
+
stronger model.
|
| 33 |
+
- Before fresh measurements, freeze task and grader hashes, seeds, model and harness settings, reasoning and
|
| 34 |
+
sampling settings, resource, tool and network limits, attempt counts and the retry and exclusion rules, and keep
|
| 35 |
+
every raw trajectory, final workspace, result and exclusion.
|
| 36 |
+
|
| 37 |
+
## Frontier tier
|
| 38 |
+
|
| 39 |
+
A harder frontier tier is in development. It is held out of this release so that it can be measured on tasks no
|
| 40 |
+
model has seen; it will ship as its own version once a calibrated measurement supports it.
|
manifest.jsonl
CHANGED
|
@@ -1,30 +1,30 @@
|
|
| 1 |
-
{"task_id": "af-001-slo-onboarding-easy-s1", "workflow": "slo-onboarding", "tier": "easy", "archetype": "operate_configure", "primary_verifier_pattern": "regression_suite", "persona": "platform_engineer", "tone": "terse", "expertise": "novice", "rules_format": "plain", "services": ["chat-gw"], "n_requirements": 5, "n_checks": 66, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.167133, "null_reward": 0.0, "grader_seconds": 1.
|
| 2 |
-
{"task_id": "af-002-slo-onboarding-easy-s2", "workflow": "slo-onboarding", "tier": "easy", "archetype": "operate_configure", "primary_verifier_pattern": "regression_suite", "persona": "platform_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "plain", "services": ["loyalty-api"], "n_requirements": 5, "n_checks": 66, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.258651, "null_reward": 0.0, "grader_seconds": 1.
|
| 3 |
-
{"task_id": "af-003-slo-onboarding-medium-s1", "workflow": "slo-onboarding", "tier": "medium", "archetype": "author_to_constraints", "primary_verifier_pattern": "behavioral_simulation", "persona": "site_reliability_engineer", "tone": "neutral", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["loyalty-svc", "ingest-gw"], "n_requirements": 9, "n_checks": 139, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.254341, "null_reward": 0.0, "grader_seconds":
|
| 4 |
-
{"task_id": "af-004-slo-onboarding-medium-s2", "workflow": "slo-onboarding", "tier": "medium", "archetype": "operate_configure", "primary_verifier_pattern": "regression_suite", "persona": "platform_engineer", "tone": "terse", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["orders-worker", "media-api"], "n_requirements": 9, "n_checks": 138, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.250639, "null_reward": 0.0, "grader_seconds": 2.
|
| 5 |
-
{"task_id": "af-005-slo-onboarding-hard-s1", "workflow": "slo-onboarding", "tier": "hard", "archetype": "author_to_constraints", "primary_verifier_pattern": "behavioral_simulation", "persona": "site_reliability_engineer", "tone": "urgent-pager", "expertise": "practitioner", "rules_format": "plain", "services": ["shipping-api", "chat-edge", "cart-edge"], "n_requirements": 14, "n_checks": 176, "req_counts": {"req_alerts": 3, "req_repairs": 5, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.212933, "null_reward": 0.0, "grader_seconds": 4.
|
| 6 |
-
{"task_id": "af-006-slo-onboarding-hard-s2", "workflow": "slo-onboarding", "tier": "hard", "archetype": "author_to_constraints", "primary_verifier_pattern": "behavioral_simulation", "persona": "site_reliability_engineer", "tone": "terse", "expertise": "expert", "rules_format": "prometheusrule", "services": ["fraud-svc", "wishlist-api", "inventory-svc"], "n_requirements": 14, "n_checks": 181, "req_counts": {"req_alerts": 3, "req_repairs": 5, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.208439, "null_reward": 0.0, "grader_seconds":
|
| 7 |
-
{"task_id": "af-007-alert-storm-cleanup-easy-s1", "workflow": "alert-storm-cleanup", "tier": "easy", "archetype": "harden_secure", "primary_verifier_pattern": "adversarial_corpus", "persona": "on_call_engineer", "tone": "terse", "expertise": "novice", "rules_format": "plain", "services": ["orders-svc"], "n_requirements": 5, "n_checks": 64, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.285102, "null_reward": 0.0, "grader_seconds": 1.
|
| 8 |
-
{"task_id": "af-008-alert-storm-cleanup-easy-s2", "workflow": "alert-storm-cleanup", "tier": "easy", "archetype": "harden_secure", "primary_verifier_pattern": "adversarial_corpus", "persona": "on_call_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "plain", "services": ["catalog-api"], "n_requirements": 5, "n_checks": 64, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.262312, "null_reward": 0.0, "grader_seconds": 1.
|
| 9 |
-
{"task_id": "af-009-alert-storm-cleanup-medium-s1", "workflow": "alert-storm-cleanup", "tier": "medium", "archetype": "harden_secure", "primary_verifier_pattern": "adversarial_corpus", "persona": "on_call_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "prometheusrule", "services": ["report-worker", "billing-svc"], "n_requirements": 10, "n_checks": 124, "req_counts": {"req_alerts": 2, "req_repairs": 4, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.253641, "null_reward": 0.0, "grader_seconds": 2.
|
| 10 |
-
{"task_id": "af-010-alert-storm-cleanup-medium-s2", "workflow": "alert-storm-cleanup", "tier": "medium", "archetype": "repair_debug", "primary_verifier_pattern": "regression_suite", "persona": "site_reliability_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "plain", "services": ["checkout-worker", "cart-edge"], "n_requirements": 10, "n_checks": 116, "req_counts": {"req_alerts": 2, "req_repairs": 4, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.230701, "null_reward": 0.0, "grader_seconds": 1.
|
| 11 |
-
{"task_id": "af-011-alert-storm-cleanup-hard-s1", "workflow": "alert-storm-cleanup", "tier": "hard", "archetype": "harden_secure", "primary_verifier_pattern": "adversarial_corpus", "persona": "on_call_engineer", "tone": "terse", "expertise": "practitioner", "rules_format": "plain", "services": ["inventory-svc", "ledger-gw", "ratings-svc"], "n_requirements": 14, "n_checks": 176, "req_counts": {"req_alerts": 2, "req_repairs": 6, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.208481, "null_reward": 0.0, "grader_seconds":
|
| 12 |
-
{"task_id": "af-012-alert-storm-cleanup-hard-s2", "workflow": "alert-storm-cleanup", "tier": "hard", "archetype": "harden_secure", "primary_verifier_pattern": "adversarial_corpus", "persona": "on_call_engineer", "tone": "neutral", "expertise": "expert", "rules_format": "prometheusrule", "services": ["fraud-api", "promo-svc", "auth-edge"], "n_requirements": 14, "n_checks": 182, "req_counts": {"req_alerts": 2, "req_repairs": 6, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.182729, "null_reward": 0.0, "grader_seconds":
|
| 13 |
-
{"task_id": "af-013-missed-page-postmortem-easy-s1", "workflow": "missed-page-postmortem", "tier": "easy", "archetype": "repair_debug", "primary_verifier_pattern": "behavioral_simulation", "persona": "incident_commander", "tone": "neutral", "expertise": "novice", "rules_format": "plain", "services": ["pricing-gw"], "n_requirements": 5, "n_checks": 62, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.239609, "null_reward": 0.0, "grader_seconds": 1.
|
| 14 |
-
{"task_id": "af-014-missed-page-postmortem-easy-s2", "workflow": "missed-page-postmortem", "tier": "easy", "archetype": "data_query", "primary_verifier_pattern": "exact_state", "persona": "site_reliability_engineer", "tone": "terse", "expertise": "novice", "rules_format": "plain", "services": ["coupon-svc"], "n_requirements": 5, "n_checks": 59, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.283333, "null_reward": 0.0, "grader_seconds": 0.
|
| 15 |
-
{"task_id": "af-015-missed-page-postmortem-medium-s1", "workflow": "missed-page-postmortem", "tier": "medium", "archetype": "repair_debug", "primary_verifier_pattern": "behavioral_simulation", "persona": "incident_commander", "tone": "urgent-pager", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["coupon-api", "orders-svc"], "n_requirements": 10, "n_checks": 137, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 2, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.260517, "null_reward": 0.0, "grader_seconds": 3.
|
| 16 |
-
{"task_id": "af-016-missed-page-postmortem-medium-s2", "workflow": "missed-page-postmortem", "tier": "medium", "archetype": "repair_debug", "primary_verifier_pattern": "behavioral_simulation", "persona": "incident_commander", "tone": "urgent-pager", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["orders-edge", "tax-edge"], "n_requirements": 10, "n_checks": 132, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 2, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.236505, "null_reward": 0.0, "grader_seconds": 2.
|
| 17 |
-
{"task_id": "af-017-missed-page-postmortem-hard-s1", "workflow": "missed-page-postmortem", "tier": "hard", "archetype": "data_query", "primary_verifier_pattern": "exact_state", "persona": "site_reliability_engineer", "tone": "urgent-pager", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["ledger-api", "payments-edge", "billing-api"], "n_requirements": 14, "n_checks": 173, "req_counts": {"req_alerts": 2, "req_repairs": 5, "req_recording": 2, "req_routing": 3, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.21644, "null_reward": 0.0, "grader_seconds":
|
| 18 |
-
{"task_id": "af-018-missed-page-postmortem-hard-s2", "workflow": "missed-page-postmortem", "tier": "hard", "archetype": "repair_debug", "primary_verifier_pattern": "behavioral_simulation", "persona": "incident_commander", "tone": "urgent-pager", "expertise": "expert", "rules_format": "prometheusrule", "services": ["feed-edge", "search-edge", "geo-gw"], "n_requirements": 14, "n_checks": 167, "req_counts": {"req_alerts": 2, "req_repairs": 5, "req_recording": 2, "req_routing": 3, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.213061, "null_reward": 0.0, "grader_seconds": 3.
|
| 19 |
-
{"task_id": "af-019-latency-slo-easy-s1", "workflow": "latency-slo", "tier": "easy", "archetype": "author_to_constraints", "primary_verifier_pattern": "differential_equivalence", "persona": "performance_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "plain", "services": ["ledger-gw"], "n_requirements": 5, "n_checks": 61, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.167133, "null_reward": 0.0, "grader_seconds": 1.
|
| 20 |
-
{"task_id": "af-020-latency-slo-easy-s2", "workflow": "latency-slo", "tier": "easy", "archetype": "author_to_constraints", "primary_verifier_pattern": "differential_equivalence", "persona": "performance_engineer", "tone": "terse", "expertise": "novice", "rules_format": "plain", "services": ["wishlist-api"], "n_requirements": 5, "n_checks": 67, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.250133, "null_reward": 0.0, "grader_seconds": 1.
|
| 21 |
-
{"task_id": "af-021-latency-slo-medium-s1", "workflow": "latency-slo", "tier": "medium", "archetype": "author_to_constraints", "primary_verifier_pattern": "differential_equivalence", "persona": "performance_engineer", "tone": "urgent-pager", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["export-edge", "refund-gw"], "n_requirements": 9, "n_checks": 137, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.213043, "null_reward": 0.0, "grader_seconds":
|
| 22 |
-
{"task_id": "af-022-latency-slo-medium-s2", "workflow": "latency-slo", "tier": "medium", "archetype": "author_to_constraints", "primary_verifier_pattern": "differential_equivalence", "persona": "performance_engineer", "tone": "urgent-pager", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["invoice-api", "notify-worker"], "n_requirements": 9, "n_checks": 132, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.261538, "null_reward": 0.0, "grader_seconds":
|
| 23 |
-
{"task_id": "af-023-latency-slo-hard-s1", "workflow": "latency-slo", "tier": "hard", "archetype": "author_to_constraints", "primary_verifier_pattern": "differential_equivalence", "persona": "performance_engineer", "tone": "terse", "expertise": "expert", "rules_format": "plain", "services": ["ingest-worker", "chat-gw", "tax-edge"], "n_requirements": 14, "n_checks": 178, "req_counts": {"req_alerts": 3, "req_repairs": 5, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.2056, "null_reward": 0.0, "grader_seconds": 4.
|
| 24 |
-
{"task_id": "af-024-latency-slo-hard-s2", "workflow": "latency-slo", "tier": "hard", "archetype": "author_to_constraints", "primary_verifier_pattern": "differential_equivalence", "persona": "performance_engineer", "tone": "urgent-pager", "expertise": "expert", "rules_format": "prometheusrule", "services": ["wishlist-gw", "fraud-worker", "stock-gw"], "n_requirements": 14, "n_checks": 179, "req_counts": {"req_alerts": 3, "req_repairs": 5, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.177933, "null_reward": 0.0, "grader_seconds":
|
| 25 |
-
{"task_id": "af-025-team-reorg-migration-easy-s1", "workflow": "team-reorg-migration", "tier": "easy", "archetype": "operate_configure", "primary_verifier_pattern": "regression_suite", "persona": "platform_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "plain", "services": ["loyalty-gw"], "n_requirements": 5, "n_checks": 61, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.208333, "null_reward": 0.0, "grader_seconds":
|
| 26 |
-
{"task_id": "af-026-team-reorg-migration-easy-s2", "workflow": "team-reorg-migration", "tier": "easy", "archetype": "operate_configure", "primary_verifier_pattern": "regression_suite", "persona": "platform_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "plain", "services": ["ingest-worker"], "n_requirements": 5, "n_checks": 61, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.208333, "null_reward": 0.0, "grader_seconds":
|
| 27 |
-
{"task_id": "af-027-team-reorg-migration-medium-s1", "workflow": "team-reorg-migration", "tier": "medium", "archetype": "transform_convert", "primary_verifier_pattern": "exact_state", "persona": "site_reliability_engineer", "tone": "neutral", "expertise": "novice", "rules_format": "plain", "services": ["pricing-worker", "inventory-worker"], "n_requirements": 9, "n_checks": 112, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.267826, "null_reward": 0.0, "grader_seconds":
|
| 28 |
-
{"task_id": "af-028-team-reorg-migration-medium-s2", "workflow": "team-reorg-migration", "tier": "medium", "archetype": "operate_configure", "primary_verifier_pattern": "regression_suite", "persona": "platform_engineer", "tone": "neutral", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["report-svc", "export-svc"], "n_requirements": 9, "n_checks": 112, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.27513, "null_reward": 0.0, "grader_seconds":
|
| 29 |
-
{"task_id": "af-029-team-reorg-migration-hard-s1", "workflow": "team-reorg-migration", "tier": "hard", "archetype": "transform_convert", "primary_verifier_pattern": "exact_state", "persona": "site_reliability_engineer", "tone": "terse", "expertise": "expert", "rules_format": "prometheusrule", "services": ["payments-gw", "fraud-svc", "profile-edge"], "n_requirements": 14, "n_checks": 181, "req_counts": {"req_alerts": 3, "req_repairs": 5, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.202898, "null_reward": 0.0, "grader_seconds": 5.
|
| 30 |
-
{"task_id": "af-030-team-reorg-migration-hard-s2", "workflow": "team-reorg-migration", "tier": "hard", "archetype": "transform_convert", "primary_verifier_pattern": "exact_state", "persona": "site_reliability_engineer", "tone": "neutral", "expertise": "expert", "rules_format": "plain", "services": ["media-edge", "identity-svc", "payments-api"], "n_requirements": 14, "n_checks": 189, "req_counts": {"req_alerts": 3, "req_repairs": 5, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.253119, "null_reward": 0.0, "grader_seconds":
|
|
|
|
| 1 |
+
{"task_id": "af-001-slo-onboarding-easy-s1", "workflow": "slo-onboarding", "tier": "easy", "archetype": "operate_configure", "primary_verifier_pattern": "regression_suite", "persona": "platform_engineer", "tone": "terse", "expertise": "novice", "rules_format": "plain", "services": ["chat-gw"], "n_requirements": 5, "n_checks": 66, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.167133, "null_reward": 0.0, "grader_seconds": 1.38, "generator_version": "0.2.0"}
|
| 2 |
+
{"task_id": "af-002-slo-onboarding-easy-s2", "workflow": "slo-onboarding", "tier": "easy", "archetype": "operate_configure", "primary_verifier_pattern": "regression_suite", "persona": "platform_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "plain", "services": ["loyalty-api"], "n_requirements": 5, "n_checks": 66, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.258651, "null_reward": 0.0, "grader_seconds": 1.25, "generator_version": "0.2.0"}
|
| 3 |
+
{"task_id": "af-003-slo-onboarding-medium-s1", "workflow": "slo-onboarding", "tier": "medium", "archetype": "author_to_constraints", "primary_verifier_pattern": "behavioral_simulation", "persona": "site_reliability_engineer", "tone": "neutral", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["loyalty-svc", "ingest-gw"], "n_requirements": 9, "n_checks": 139, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.254341, "null_reward": 0.0, "grader_seconds": 2.96, "generator_version": "0.2.0"}
|
| 4 |
+
{"task_id": "af-004-slo-onboarding-medium-s2", "workflow": "slo-onboarding", "tier": "medium", "archetype": "operate_configure", "primary_verifier_pattern": "regression_suite", "persona": "platform_engineer", "tone": "terse", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["orders-worker", "media-api"], "n_requirements": 9, "n_checks": 138, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.250639, "null_reward": 0.0, "grader_seconds": 2.99, "generator_version": "0.2.0"}
|
| 5 |
+
{"task_id": "af-005-slo-onboarding-hard-s1", "workflow": "slo-onboarding", "tier": "hard", "archetype": "author_to_constraints", "primary_verifier_pattern": "behavioral_simulation", "persona": "site_reliability_engineer", "tone": "urgent-pager", "expertise": "practitioner", "rules_format": "plain", "services": ["shipping-api", "chat-edge", "cart-edge"], "n_requirements": 14, "n_checks": 176, "req_counts": {"req_alerts": 3, "req_repairs": 5, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.212933, "null_reward": 0.0, "grader_seconds": 4.78, "generator_version": "0.2.0"}
|
| 6 |
+
{"task_id": "af-006-slo-onboarding-hard-s2", "workflow": "slo-onboarding", "tier": "hard", "archetype": "author_to_constraints", "primary_verifier_pattern": "behavioral_simulation", "persona": "site_reliability_engineer", "tone": "terse", "expertise": "expert", "rules_format": "prometheusrule", "services": ["fraud-svc", "wishlist-api", "inventory-svc"], "n_requirements": 14, "n_checks": 181, "req_counts": {"req_alerts": 3, "req_repairs": 5, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.208439, "null_reward": 0.0, "grader_seconds": 5.17, "generator_version": "0.2.0"}
|
| 7 |
+
{"task_id": "af-007-alert-storm-cleanup-easy-s1", "workflow": "alert-storm-cleanup", "tier": "easy", "archetype": "harden_secure", "primary_verifier_pattern": "adversarial_corpus", "persona": "on_call_engineer", "tone": "terse", "expertise": "novice", "rules_format": "plain", "services": ["orders-svc"], "n_requirements": 5, "n_checks": 64, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.285102, "null_reward": 0.0, "grader_seconds": 1.09, "generator_version": "0.2.0"}
|
| 8 |
+
{"task_id": "af-008-alert-storm-cleanup-easy-s2", "workflow": "alert-storm-cleanup", "tier": "easy", "archetype": "harden_secure", "primary_verifier_pattern": "adversarial_corpus", "persona": "on_call_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "plain", "services": ["catalog-api"], "n_requirements": 5, "n_checks": 64, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.262312, "null_reward": 0.0, "grader_seconds": 1.11, "generator_version": "0.2.0"}
|
| 9 |
+
{"task_id": "af-009-alert-storm-cleanup-medium-s1", "workflow": "alert-storm-cleanup", "tier": "medium", "archetype": "harden_secure", "primary_verifier_pattern": "adversarial_corpus", "persona": "on_call_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "prometheusrule", "services": ["report-worker", "billing-svc"], "n_requirements": 10, "n_checks": 124, "req_counts": {"req_alerts": 2, "req_repairs": 4, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.253641, "null_reward": 0.0, "grader_seconds": 2.3, "generator_version": "0.2.0"}
|
| 10 |
+
{"task_id": "af-010-alert-storm-cleanup-medium-s2", "workflow": "alert-storm-cleanup", "tier": "medium", "archetype": "repair_debug", "primary_verifier_pattern": "regression_suite", "persona": "site_reliability_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "plain", "services": ["checkout-worker", "cart-edge"], "n_requirements": 10, "n_checks": 116, "req_counts": {"req_alerts": 2, "req_repairs": 4, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.230701, "null_reward": 0.0, "grader_seconds": 1.89, "generator_version": "0.2.0"}
|
| 11 |
+
{"task_id": "af-011-alert-storm-cleanup-hard-s1", "workflow": "alert-storm-cleanup", "tier": "hard", "archetype": "harden_secure", "primary_verifier_pattern": "adversarial_corpus", "persona": "on_call_engineer", "tone": "terse", "expertise": "practitioner", "rules_format": "plain", "services": ["inventory-svc", "ledger-gw", "ratings-svc"], "n_requirements": 14, "n_checks": 176, "req_counts": {"req_alerts": 2, "req_repairs": 6, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.208481, "null_reward": 0.0, "grader_seconds": 5.23, "generator_version": "0.2.0"}
|
| 12 |
+
{"task_id": "af-012-alert-storm-cleanup-hard-s2", "workflow": "alert-storm-cleanup", "tier": "hard", "archetype": "harden_secure", "primary_verifier_pattern": "adversarial_corpus", "persona": "on_call_engineer", "tone": "neutral", "expertise": "expert", "rules_format": "prometheusrule", "services": ["fraud-api", "promo-svc", "auth-edge"], "n_requirements": 14, "n_checks": 182, "req_counts": {"req_alerts": 2, "req_repairs": 6, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.182729, "null_reward": 0.0, "grader_seconds": 6.33, "generator_version": "0.2.0"}
|
| 13 |
+
{"task_id": "af-013-missed-page-postmortem-easy-s1", "workflow": "missed-page-postmortem", "tier": "easy", "archetype": "repair_debug", "primary_verifier_pattern": "behavioral_simulation", "persona": "incident_commander", "tone": "neutral", "expertise": "novice", "rules_format": "plain", "services": ["pricing-gw"], "n_requirements": 5, "n_checks": 62, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.239609, "null_reward": 0.0, "grader_seconds": 1.19, "generator_version": "0.2.0"}
|
| 14 |
+
{"task_id": "af-014-missed-page-postmortem-easy-s2", "workflow": "missed-page-postmortem", "tier": "easy", "archetype": "data_query", "primary_verifier_pattern": "exact_state", "persona": "site_reliability_engineer", "tone": "terse", "expertise": "novice", "rules_format": "plain", "services": ["coupon-svc"], "n_requirements": 5, "n_checks": 59, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.283333, "null_reward": 0.0, "grader_seconds": 0.74, "generator_version": "0.2.0"}
|
| 15 |
+
{"task_id": "af-015-missed-page-postmortem-medium-s1", "workflow": "missed-page-postmortem", "tier": "medium", "archetype": "repair_debug", "primary_verifier_pattern": "behavioral_simulation", "persona": "incident_commander", "tone": "urgent-pager", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["coupon-api", "orders-svc"], "n_requirements": 10, "n_checks": 137, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 2, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.260517, "null_reward": 0.0, "grader_seconds": 3.0, "generator_version": "0.2.0"}
|
| 16 |
+
{"task_id": "af-016-missed-page-postmortem-medium-s2", "workflow": "missed-page-postmortem", "tier": "medium", "archetype": "repair_debug", "primary_verifier_pattern": "behavioral_simulation", "persona": "incident_commander", "tone": "urgent-pager", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["orders-edge", "tax-edge"], "n_requirements": 10, "n_checks": 132, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 2, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.236505, "null_reward": 0.0, "grader_seconds": 2.48, "generator_version": "0.2.0"}
|
| 17 |
+
{"task_id": "af-017-missed-page-postmortem-hard-s1", "workflow": "missed-page-postmortem", "tier": "hard", "archetype": "data_query", "primary_verifier_pattern": "exact_state", "persona": "site_reliability_engineer", "tone": "urgent-pager", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["ledger-api", "payments-edge", "billing-api"], "n_requirements": 14, "n_checks": 173, "req_counts": {"req_alerts": 2, "req_repairs": 5, "req_recording": 2, "req_routing": 3, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.21644, "null_reward": 0.0, "grader_seconds": 5.36, "generator_version": "0.2.0"}
|
| 18 |
+
{"task_id": "af-018-missed-page-postmortem-hard-s2", "workflow": "missed-page-postmortem", "tier": "hard", "archetype": "repair_debug", "primary_verifier_pattern": "behavioral_simulation", "persona": "incident_commander", "tone": "urgent-pager", "expertise": "expert", "rules_format": "prometheusrule", "services": ["feed-edge", "search-edge", "geo-gw"], "n_requirements": 14, "n_checks": 167, "req_counts": {"req_alerts": 2, "req_repairs": 5, "req_recording": 2, "req_routing": 3, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.213061, "null_reward": 0.0, "grader_seconds": 3.88, "generator_version": "0.2.0"}
|
| 19 |
+
{"task_id": "af-019-latency-slo-easy-s1", "workflow": "latency-slo", "tier": "easy", "archetype": "author_to_constraints", "primary_verifier_pattern": "differential_equivalence", "persona": "performance_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "plain", "services": ["ledger-gw"], "n_requirements": 5, "n_checks": 61, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.167133, "null_reward": 0.0, "grader_seconds": 1.97, "generator_version": "0.2.0"}
|
| 20 |
+
{"task_id": "af-020-latency-slo-easy-s2", "workflow": "latency-slo", "tier": "easy", "archetype": "author_to_constraints", "primary_verifier_pattern": "differential_equivalence", "persona": "performance_engineer", "tone": "terse", "expertise": "novice", "rules_format": "plain", "services": ["wishlist-api"], "n_requirements": 5, "n_checks": 67, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.250133, "null_reward": 0.0, "grader_seconds": 1.75, "generator_version": "0.2.0"}
|
| 21 |
+
{"task_id": "af-021-latency-slo-medium-s1", "workflow": "latency-slo", "tier": "medium", "archetype": "author_to_constraints", "primary_verifier_pattern": "differential_equivalence", "persona": "performance_engineer", "tone": "urgent-pager", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["export-edge", "refund-gw"], "n_requirements": 9, "n_checks": 137, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.213043, "null_reward": 0.0, "grader_seconds": 3.34, "generator_version": "0.2.0"}
|
| 22 |
+
{"task_id": "af-022-latency-slo-medium-s2", "workflow": "latency-slo", "tier": "medium", "archetype": "author_to_constraints", "primary_verifier_pattern": "differential_equivalence", "persona": "performance_engineer", "tone": "urgent-pager", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["invoice-api", "notify-worker"], "n_requirements": 9, "n_checks": 132, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.261538, "null_reward": 0.0, "grader_seconds": 3.07, "generator_version": "0.2.0"}
|
| 23 |
+
{"task_id": "af-023-latency-slo-hard-s1", "workflow": "latency-slo", "tier": "hard", "archetype": "author_to_constraints", "primary_verifier_pattern": "differential_equivalence", "persona": "performance_engineer", "tone": "terse", "expertise": "expert", "rules_format": "plain", "services": ["ingest-worker", "chat-gw", "tax-edge"], "n_requirements": 14, "n_checks": 178, "req_counts": {"req_alerts": 3, "req_repairs": 5, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.2056, "null_reward": 0.0, "grader_seconds": 4.95, "generator_version": "0.2.0"}
|
| 24 |
+
{"task_id": "af-024-latency-slo-hard-s2", "workflow": "latency-slo", "tier": "hard", "archetype": "author_to_constraints", "primary_verifier_pattern": "differential_equivalence", "persona": "performance_engineer", "tone": "urgent-pager", "expertise": "expert", "rules_format": "prometheusrule", "services": ["wishlist-gw", "fraud-worker", "stock-gw"], "n_requirements": 14, "n_checks": 179, "req_counts": {"req_alerts": 3, "req_repairs": 5, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.177933, "null_reward": 0.0, "grader_seconds": 7.44, "generator_version": "0.2.0"}
|
| 25 |
+
{"task_id": "af-025-team-reorg-migration-easy-s1", "workflow": "team-reorg-migration", "tier": "easy", "archetype": "operate_configure", "primary_verifier_pattern": "regression_suite", "persona": "platform_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "plain", "services": ["loyalty-gw"], "n_requirements": 5, "n_checks": 61, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.208333, "null_reward": 0.0, "grader_seconds": 1.2, "generator_version": "0.2.0"}
|
| 26 |
+
{"task_id": "af-026-team-reorg-migration-easy-s2", "workflow": "team-reorg-migration", "tier": "easy", "archetype": "operate_configure", "primary_verifier_pattern": "regression_suite", "persona": "platform_engineer", "tone": "urgent-pager", "expertise": "novice", "rules_format": "plain", "services": ["ingest-worker"], "n_requirements": 5, "n_checks": 61, "req_counts": {"req_alerts": 1, "req_repairs": 1, "req_recording": 1, "req_routing": 1, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.208333, "null_reward": 0.0, "grader_seconds": 1.2, "generator_version": "0.2.0"}
|
| 27 |
+
{"task_id": "af-027-team-reorg-migration-medium-s1", "workflow": "team-reorg-migration", "tier": "medium", "archetype": "transform_convert", "primary_verifier_pattern": "exact_state", "persona": "site_reliability_engineer", "tone": "neutral", "expertise": "novice", "rules_format": "plain", "services": ["pricing-worker", "inventory-worker"], "n_requirements": 9, "n_checks": 112, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.267826, "null_reward": 0.0, "grader_seconds": 2.72, "generator_version": "0.2.0"}
|
| 28 |
+
{"task_id": "af-028-team-reorg-migration-medium-s2", "workflow": "team-reorg-migration", "tier": "medium", "archetype": "operate_configure", "primary_verifier_pattern": "regression_suite", "persona": "platform_engineer", "tone": "neutral", "expertise": "practitioner", "rules_format": "prometheusrule", "services": ["report-svc", "export-svc"], "n_requirements": 9, "n_checks": 112, "req_counts": {"req_alerts": 2, "req_repairs": 3, "req_recording": 1, "req_routing": 2, "req_inhibit": 1}, "oracle_reward": 1.0, "partial_reward": 0.27513, "null_reward": 0.0, "grader_seconds": 2.47, "generator_version": "0.2.0"}
|
| 29 |
+
{"task_id": "af-029-team-reorg-migration-hard-s1", "workflow": "team-reorg-migration", "tier": "hard", "archetype": "transform_convert", "primary_verifier_pattern": "exact_state", "persona": "site_reliability_engineer", "tone": "terse", "expertise": "expert", "rules_format": "prometheusrule", "services": ["payments-gw", "fraud-svc", "profile-edge"], "n_requirements": 14, "n_checks": 181, "req_counts": {"req_alerts": 3, "req_repairs": 5, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.202898, "null_reward": 0.0, "grader_seconds": 5.48, "generator_version": "0.2.0"}
|
| 30 |
+
{"task_id": "af-030-team-reorg-migration-hard-s2", "workflow": "team-reorg-migration", "tier": "hard", "archetype": "transform_convert", "primary_verifier_pattern": "exact_state", "persona": "site_reliability_engineer", "tone": "neutral", "expertise": "expert", "rules_format": "plain", "services": ["media-edge", "identity-svc", "payments-api"], "n_requirements": 14, "n_checks": 189, "req_counts": {"req_alerts": 3, "req_repairs": 5, "req_recording": 2, "req_routing": 2, "req_inhibit": 2}, "oracle_reward": 1.0, "partial_reward": 0.253119, "null_reward": 0.0, "grader_seconds": 9.48, "generator_version": "0.2.0"}
|
null/afx-e1-s01/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e1-s02/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e1-s03/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e1-s04/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e1-s05/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e1-s06/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e1-s07/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e1-s08/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e1-s09/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e2-s01/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e2-s02/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e2-s03/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e2-s04/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e2-s05/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e2-s06/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e2-s07/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e2-s08/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e2-s09/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e3-s01/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e3-s02/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e3-s03/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e3-s04/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e3-s05/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e3-s06/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e3-s07/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e3-s08/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e3-s09/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e4-s01/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e4-s02/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e4-s03/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e4-s04/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e4-s05/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e4-s06/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e4-s07/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e4-s08/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e4-s09/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e5-s01/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e5-s02/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e5-s03/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e5-s04/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e5-s05/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e5-s06/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e5-s07/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e5-s08/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
null/afx-e5-s09/solve.sh
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# null: no-op (equivalent to harbor -a nop)
|
| 3 |
+
exit 0
|
openenv/alertforge_env/README.md
CHANGED
|
@@ -13,7 +13,7 @@ tags: [openenv, rl-environment, harbor, prometheus, alertmanager, sre, dense-rew
|
|
| 13 |
|
| 14 |
# 3amBench / AlertForge: OpenEnv server
|
| 15 |
|
| 16 |
-
Multi-turn version of the [3amBench Harbor dataset](https://huggingface.co/datasets/
|
| 17 |
repo with tools, and every step is graded by the same `grade.py` Harbor uses (promtool + amtool replays of hidden
|
| 18 |
outage scenarios).
|
| 19 |
|
|
|
|
| 13 |
|
| 14 |
# 3amBench / AlertForge: OpenEnv server
|
| 15 |
|
| 16 |
+
Multi-turn version of the [3amBench Harbor dataset](https://huggingface.co/datasets/openenvforge/3ambench): the agent edits a Prometheus/Alertmanager
|
| 17 |
repo with tools, and every step is graded by the same `grade.py` Harbor uses (promtool + amtool replays of hidden
|
| 18 |
outage scenarios).
|
| 19 |
|