# OracleMangle

> A high-performance safety rail for event ambiguity detection, source reliability, and resolution-risk scoring in research pipelines, info-markets, and data workflows.

## What it does

OracleMangle continuously monitors public event questions and evaluates them for semantic ambiguity, poorly defined thresholds, and source reliability issues. By identifying wording risks that lead to manual-review friction or contested outcomes, it allows automated workflows to safely route questions to appropriate processing channels.

- Platform: Public oracle schemas and event frameworks (e.g., UMA, etc.)
- Labelled history: 167,773 event questions tracked, 1,686 contested outcomes labeled
- Coverage: Continuous evaluation of active public event queues and custom question inputs
- Output: Contested outcome database, REST API for custom question-scoring, automated high-risk alerts, and workflow routing recommendations.

## The problem it solves

When automated decision engines or research workflows depend on public event databases, semantic ambiguity in question formulations can cause significant pipeline friction. If a question is contested due to vague terms or missing rules, processing queues can freeze and manual intervention is required. OracleMangle flags these issues pre-entry, enabling automated review routing.

Three patterns account for most contested outcomes:
1. Subjective thresholds — terms like "clearly", "significant", or "regime fall" used without precise, objective metrics.
2. Ambiguous key terms — words like "ceasefire", "collapse", or "officially" without specifying a single, authoritative reference source.
3. Source reliability issues — depending on news consensus or unverified media reports instead of official government documents or primary data feeds.

## Proof

Calibration (measured 2026-07-25 on the live model, held-out): held-out resolved UMA markets, 47 contested / 94 clean completed, scored with the exact production prompt. At the published 0.4 flag threshold the scorer flagged 19 markets, 15 of which were genuinely contested. Adjusted to the ~10% real-world base rate: ~45% precision (about 4.5x baseline) at 32% recall, AUC 0.630. Caveats stated plainly: n=19 flagged is a small sample (Wilson 95% CI on the hit rate 57%-91%), and recall is under a third because the threshold is tuned for precision rather than coverage. An earlier "8.2x baseline dispute rate" figure described a previous pipeline (Gemini 2.5 Flash plus retrieval) measured in-sample on a non-randomly-selected corpus; that model scores AUC 0.528 on this held-out set and the figure should not be cited as current product performance. Every API response returns the scoring model ID and the flag threshold so scores can be independently audited. 167,773 event questions tracked overall, 1,686 confirmed contested outcomes labeled.

Real examples of questions flagged before conflict/dispute:
- "Will Biden be declared senile by a doctor before August?" — flagged high risk, contested
- "Will Gaza ceasefire deal be reached before Dec 31 2024?" — flagged high risk, contested twice
- "Will Ukraine ceasefire be reached before 2025?" — flagged high risk, contested

Clean examples (rarely contested):
- Sports outcomes with single official source (FIFA, NBA)
- Federal Reserve rate decisions (FOMC statement is the source)
- Official government electoral college results

## How it works

1. Public event feeds are tracked continuously.
2. A daily queue selects new event formulations for QA scoring.
3. Each question is scored using a Gemini-powered model with RAG retrieval over our 167K+ labeled history of contested event questions.
4. Scores feed the daily watchlist, REST API, and automated alerts for review routing.

## Products

- **Free**: Telegram Bot (@OracleManglebot) — basic ambiguity checks for any open event question.
- **Starter ($19/mo)**: 1,000 API calls per day for lightweight research workflows.
- **Pro ($79/mo)**: 5,000 API calls per day, fresher API access, and webhook-ready integration.
- **Institutional (Custom)**: Custom limits, private data flows, integration/compliance review.
- **Ambiguity Screening Report ($99 one-off)**: Wording and source-reliability audit for up to 10 questions.
- **Advanced Workflow QA Audit ($499 one-off)**: Deep pipeline audit for up to 50 questions, rewrite recommendations, and 30 days Pro API access.

## API

Base URL: https://api.oraclemangle.com/v1

Endpoints:
- `GET /markets/{id}` — ambiguity score and risk metadata for a specific public question
- `GET /markets` — list of evaluated questions (filterable by risk level, source type)
- `GET /alerts` — recent high-risk ambiguity alerts
- `GET /score` — score any custom question on demand

Response includes: `dispute_risk` (float 0-1), `safe_to_trade` (bool, indicates whether wording is robust enough for automated settlement/routing), `question` (string)

## Content

Blog: https://oraclemangle.com/blog/

- Wording audits: identifying event question ambiguity before workflow pipelines stall.
- How to audit event question wording before deploying automated tracking.
- Analyzing 2024-2025's most contested event outcomes: three recurring patterns.
- How decentralized event oracles resolve conflicts and why semantic gaps trigger disputes.
- Integrating automated ambiguity scoring into NLP research pipelines (Python, REST API).

## Open Source

Page: https://oraclemangle.com/open-source.html

Two free research artifacts from the OracleMangle team:

- **polymarket-bot-lab** (https://github.com/oraclemangle/polymarket-bot-lab, Apache-2.0) — a lab of 11 candidate Polymarket trading bots plus research framework: CLOB clients, fee models, backtest engine, portfolio/risk gates, tests, fleet dashboard. Headline candidates: Playbook (sports mid-band NO-fade; post-fee ROI +5.32% on n=573 but 95% CI crosses zero) and Nimbus (weather range-fade; +11.07% ROI on a tiny 95-group live probe). Honest framing: research candidates, not proven edges. The Oracle bot (bot_b) is reference-only — it consumed OracleMangle's paid dispute-risk scorer over HTTP (GET /v1/score).
- **polymarket-canary-tape** (https://huggingface.co/datasets/oraclemangle/polymarket-canary-tape, CC-BY-4.0) — free market-data tape: 271,566,165 CEX trades (BTC/ETH/SOL), ~61M Polymarket WebSocket events, 23,750 market snapshots, 2026-04-30 → 2026-07-05, three zstd-compressed SQLite tapes (~96 GB raw) with a dual-vantage overlap window for latency/feed-consistency studies. Parquet samples preview in the HF Dataset Viewer.

## Contact

- Site: https://oraclemangle.com
- Telegram: https://t.me/OracleManglebot
- Twitter/X: https://x.com/OracleMangle
- Email: hello@oraclemangle.com