- Home
- Skills
- APIs & Backend
- Multi-Frame Consensus and Decision Gate for Recognizers
Multi-Frame Consensus and Decision Gate for Recognizers
Detection and OCR are the easy part of a camera recognizer; deciding when to act is where the harm happens.
$29
Works with the AI Agents you already use
Multi-Frame Consensus and Decision Gate for Recognizers
Example session with this skill installed
Process a track of 5 frames for a suspected license plate. Two frames show 'ABC128', three frames show 'ABC12B' with low confidence on the last char. Check against the signed hotlist.
- Read your context and instructions
- Compiled the multi-frame consensus
The tracker has fused 5 distinct frames, excluding 2 duplicates based on grayscale fingerprints. The look-alike rule was triggered for the 8/B ambiguity, resulting in a REVIEW status due to lack of a two-thirds integer majority.
Decision Result
| Field | Value |
|---|---|
| Decision | REVIEW |
| Text | ABC12B |
| Reasons | lookalike_ambiguity_detected, insufficient_consensus_ratio, possible_target_match |
| Review Required | True |
| Policy Version | 1.0.4 |
| Agreement Ratio | 3/5 (60%) |
| Track Purity | High |
| List Status | Verified (HMAC OK) |
Next steps
- Route this result to the manual verification queue for human operator confirmation of the final character.
- Review the
CharCandidateconfidence scores for the 6th position in the raw telemetry to calibrate futureGateConfigfloors. - Check if the
ABC128string exists in the hotlist; the gate flagged this asREVIEWbecause a weak read cannot dismiss a potential target.
Connects securely to your tools. The creator never sees your data.
About this skill
Detection and OCR are the easy part of a camera recognizer; deciding when to act is where the harm happens. This kit is the fusion and policy layer between any detector/reader pair and the action.
consensus.py: per-position votes across frames of the same object with one vote per distinct frame (near-duplicate frames excluded by a fingerprint distance so a static scene cannot manufacture agreement); look-alike pairs 0/O, 1/I, 5/S, 8/B and 2/Z resolved only by a strict majority AND two thirds of the electorate tested as integers (a >= 0.67 test failed 2-of-3 by rounding, so a three-frame track could never resolve such a position); the plurality length with a tie broken toward the longer read, because a clipped end is the dominant real failure and insertion order once overwrote a correct read with its truncation; a read one character short aligned at the clipped end so it votes without changing the length; a tracker that associates boxes by IoU or centre distance, keeps separate objects distinct, picks the sharpest observation per frame and computes second-mode support as a track-purity signal.
decision.py: four outcomes with reasons and a policy version. UNREADABLE for empty, short or low-confidence reads; REVIEW for any unresolved position, too few independent frames, a competing read, or a possible target with any weakness (never NO_MATCH on a possible target); NO_MATCH only for a clean, confident, multi-frame read off the list; MATCH only when the read is on a loaded, signed list and frames, agreement, confidence and purity all pass, and even then a human reviews. Frame agreement is compared as an exact fraction, so 4 of 6 meets two thirds. Thresholds have floors (and the one laxer-when-higher threshold a ceiling) checked on every call, so a configuration cannot weaken the policy after startup. Hotlist verifies an HMAC over canonical entries; a tampered list matches nothing and forces REVIEW on every read, so tampering can never dismiss a genuine target.
Reference: the outcome table, why the floors exist, what counts as independent evidence, the evaluation discipline (frozen identity-disjoint confirmation set, baseline first, no challenger that raises false target alerts). 10 tests. Python 3.9+, standard library, no network.
Requires pre-calculated sharpness and fingerprints. Format-specific string validation hooks are provided but empty. No built-in image processing functions.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 18 days ago
Requires