Known limitations¶
A single place for caveats that are otherwise scattered across guides and audience pages. Read this before treating any exported artifact as more authoritative than it is.
Proposition bridge covers deterministic templates only¶
REFERS_TO, PRICES, COMPLEMENT, and EXACTLY_ONE edges (plus
Proposition payloads on OUTCOME nodes) are emitted by the deterministic
proposition compiler for:
- match moneylines / draw / team-to-advance
- group-winner markets
- stage-of-elimination markets
- world-cup-winner markets
- nation/team reaches-stage markets (
Nation to Reach Final/Semifinals/…,Team to advance to Knockout Stages)
Match-result EXACTLY_ONE partitions require the draw market in addition to
the two team moneylines — team-only moneylines do not claim exclusivity
because a soccer match can still draw. See
Logical layer for the full predicate and edge
emission table.
Residual / unrecognized market types still produce EVENT / MARKET /
OUTCOME structure without formal propositions, so those outcomes remain
disconnected from topology entities. Check proposition_json on exported
nodes and derivation_type on edges before assuming a logical bridge exists.
Rule engine is intentionally small¶
Direct logical edges (IMPLIES, EQUIVALENT, MUTEX) come from a fixed
WC2026 rule registry — not from LLM judgment. Transitive IMPLIES closure is
on-demand via oddsgraph closure and is not written into
edges.parquet by default. See Logical layer
for the registry table and flag interactions.
Scope is WC2026 / Polymarket only¶
Deterministic templates, team alias/code tables, proposition predicates, and
the official bracket fragment (oddsgraph/data/wc2026_schedule.json) are all
built against the Polymarket WC2026 hourly-odds schema and FIFA's 2026 World
Cup structure. Nothing here generalizes to other tournaments or other
prediction-market platforms without new templates, alias tables, and a new
schedule export.
Residual LLM output is lower-confidence than deterministic/official paths¶
Events not covered by deterministic templates or the official bracket go
through chunked local LLM extraction. This path is inherently less reliable
than template-driven extraction. Always check inference_method and
confidence before trusting an edge or node, and prefer filtering with
--minimum-confidence for downstream use. See
Entity resolution and
Integrators.
Fine-tuning scripts are experimental¶
scripts/export_finetune_dataset.py, scripts/finetune_lora.py, and
scripts/eval_finetuned_model.py are not part of the stable CLI surface.
Flags and output paths may change without the same compatibility
expectations as oddsgraph's core commands. See
Fine-tuning.
No hosted service or bundled data¶
OddsFox Graph is Hypertrial-owned MIT software, not a hosted dataset or API. You must supply your own Pipeline golden mart export and local model weights; nothing here is fetched or served on your behalf.
Apple Silicon is the primary tested environment¶
Backend performance guidance (inprocess, mlx, llama-server tuning) is
based on Apple Silicon + Metal measurements. CPU-only / Linux paths work
(see Linux / CPU-only setup) but are
exercised mainly through CI correctness checks, not performance
benchmarking.