Public event data looks clean until the wording gets tested. OracleMangle finds ambiguous terms, weak source rules, and resolution-risk patterns so analysts and automated systems can route messy questions for review before they create operational drag.
Free delayed signals, $19 Starter API, $79 Pro API, and custom institutional research workflows.
Measured 25 July 2026 on the model actually serving traffic. Held-out set of resolved UMA markets (47 contested, 94 clean completed), scored with the exact prompt the API ships. At our published 0.4 flag threshold the scorer flagged 19 markets, of which 15 were genuinely contested. Adjusted to the ~10% real-world base rate that is roughly 45% precision, about 4.5x baseline, at 32% recall. Read that honestly: 19 flagged is a small sample — the 95% confidence interval on the hit rate runs 57% to 91% — and we catch under a third of contested outcomes, because we tune for precision when we flag rather than coverage. Every response returns the scoring model ID and threshold so you can re-run any of it.
A lightweight screening report for teams that want a human-readable review of event questions, source rules, and likely resolution friction.
The report ranks questions by ambiguity, explains the wording problem, and recommends whether to automate, rewrite, or manually review.
Ambiguous event questions create downstream review work, inconsistent labels, and brittle automated decisions. The earlier you catch them, the cheaper they are to route or rewrite.
Most pipelines treat event text as structured data too early. That misses whether the wording is subjective, underspecified, or dependent on sources that may disagree.
“Material,” “significant,” “substantial,” and “generally available” sound usable until a system has to apply them consistently.
OracleMangle scores event questions for resolution risk and returns a simple signal: automate, rewrite, or send to manual review.
These are the kinds of wording problems OracleMangle is built to catch before they become review problems.
Our model scored this at 75% resolution risk. The trigger was obvious: “significant” is subjective unless the workflow defines a measurable threshold.
Scored at 75% risk. “Materially breach” sounds clear until the source rule needs to be applied across real evidence.
“Fully launch” needs interpretation: private beta, public availability, paid access, regional rollout, or something else?
Risk score: 5%. Single authority, binary result, and a straight resolution path. This is what event data looks like when it can flow through automation cleanly.
Historical reference — not current product performance. Calibration on the 139K public event questions scored by our earlier pipeline (Gemini 2.5 Flash plus retrieval, 2025–2026). That model is no longer available and is not what the API serves today. Two further caveats we think you should weigh: the production scorer prioritises questions it considers worth reviewing, so these buckets are not a uniform random sample, and the rates below are measured on the same corpus the model scored rather than on held-out data. On a held-out sample that same model separates contested from clean outcomes far more weakly than this table implies. We publish it as methodology history; judge the live product on the held-out validation described above.
| Risk Bucket | Questions | Contested Outcomes | Contest Rate | vs Baseline |
|---|---|---|---|---|
| Clean (0-10%) | 88,410 | 702 | 0.8% | 0.7x |
| Medium (10-25%) | 44,469 | 548 | 1.2% | 1.1x |
| High (25-50%) | 6,392 | 324 | 5.1% | 4.4x |
| Extreme (50%+) | 213 | 20 | 9.4% | 8.2x |
OracleMangle is built to fit research and automation pipelines: quick scan, structured score, clear routing.
We analyze wording ambiguity, resolution source reliability, and historical precedent across a labelled history of 167K+ public event questions.
The output is a resolution-risk score that ranks questions by how likely they are to create contested interpretation or manual review.
Humans get Telegram examples and explanations. Automated workflows get structured API responses for routing and review logic.
If a question is ambiguous, route it for review before downstream systems treat it as clean data.
import requests
RISK_THRESHOLD = 0.25
response = requests.get(
"https://api.oraclemangle.com/v1/score",
headers={"X-API-Key": API_KEY},
params={"question": event_question},
timeout=5,
)
signal = response.json()
if signal["dispute_risk"] > RISK_THRESHOLD:
queue_manual_review()
else:
continue_workflow()Structured resolution risk for teams that need machine-readable routing before automation continues.
Each tier is designed for a clear stage: explore the signal, test delayed access, run active automation, then scale with custom support.
For exploring delayed examples before adding OracleMangle to a workflow.
For lightweight research workflows that can work with delayed or limited access.
For active API usage, higher limits, and production research workflows.
For teams that need custom limits, private data flows, or deeper integration support.
The fastest way to build trust is to show the dataset, the contested examples, and the mechanics behind the signal.
How to score event questions and route high-risk wording into manual review.
Read the API overviewThe five-question manual checklist researchers can use even without the product.
Read the guideA searchable reference library of ambiguous event questions and historical resolution friction.
Browse the dataset11 open-source Polymarket bot candidates and a 96 GB CC-BY-4.0 market-data tape, with honest results.
Explore the researchA plain-English explainer of the ambiguity, source reliability, and routing signals.
Read the explainerCheck our work rather than take our word. On a held-out set of resolved UMA markets (47 contested, 94 clean completed), scored with the exact prompt the API ships, the scorer flagged 19 and 15 of those were genuinely contested — roughly 45% precision against a ~10% base rate, at 32% recall. The honest caveats: 19 flagged is small, so the 95% confidence interval on that hit rate is 57–91%, and we miss most contested outcomes by design because we tune for precision when we do flag. The historical calibration table above describes an earlier pipeline measured in-sample on a non-random corpus; we have left it up as methodology history rather than quietly delete it, but it is not a claim about today's API, and on held-out data that older model separates contested from clean far more weakly than the table suggests. Every response returns its model ID and threshold, and the validation set comes from public UMA resolution data you can rebuild.
You can, and you should for important workflows. OracleMangle is for applying that discipline consistently across far more questions than a person wants to read line by line.
No. Free is useful for learning the signal, Starter is useful for lightweight API experiments, and Pro is useful once the API becomes part of a recurring workflow.
The signal is rooted in ambiguity, weak sources, and undefined thresholds. Different source, same class of wording problem.
Start free in Telegram, use Starter for lightweight API tests, or move straight to Pro for active automation.