[BLOOM] — TRP1 FDE Programme, April 2026

[IbnuEyni]

BLOOM Team — Oracle Forge

Programme: TRP1 FDE April 2026
Model: google/gemini-3.1-pro-preview
Overall Score: 19/54 = 35.2% pass@1
Repo: https://github.com/IbnuEyni/oracle-forge

Architecture

3-layer KB context injection (~16KB per query):

• Layer 1: Schema hints via --use_hints
• Layer 2: Domain knowledge (join keys, field inventory)
• Layer 3: Corrections memory (auto-written after every failure)

Results

Dataset	Passed	Total	Score
bookreview	3	3	100%
stockindex	2	3	66.7%
crmarenapro	6	13	46.2%
yelp	4	7	57.1%
music_brainz_20k	1	3	33.3%
googlelocal	1	4	25.0%
agnews	1	4	25.0%
stockmarket	1	5	20.0%
GITHUB_REPOS	0	4	0.0%
PANCANCER_ATLAS	0	3	0.0%
PATENTS	0	3	0.0%
DEPS_DEV_V1	0	2	0.0%

[shreyashankar]

Hi @IbnuEyni — we can't re-validate this submission as-is because the JSON has only aggregate counts (overall, per-dataset {passed, total}). Please re-emit according to the instructions in the README — an array of {"dataset", "query", "run", "answer"} entries, one per run, covering every query across all 12 datasets with at least 5 runs each. Once the file is in I'll verify and post the Pass@1 here.