The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-2398
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-06T14:00:50.643054Z
status: published
lens: trace
sector: sports
headline: Faster, more accurate? A feasibility study on replacing human judges with artificial intelligence in video review for the Paris Olympics Taekwondo competition
dek: Introduction: This study explores the potential of artificial intelligence (AI) to enhance the accuracy and efficiency of video review systems in Taekwondo, addressing limitations in current human-based judgment processes during competitions. Methods: A total of 241 video review cases from the 2024 Paris Olympic Taekwondo competition were analyzed. AI-based judgments were generated using ChatGPT-4.5 and OpenPose deep learning models. The AI-generated penalty decisions were statistically compared to those made by…
gain_title: In a study of Paris 2024 taekwondo video reviews, ChatGPT-4.5 with OpenPose agreed almost perfectly with international referees (kappa 0.897) and cut average review time by about 81%.
problem_title: The same AI disagreed with referees in nine reviews, mostly minimal-contact head strikes or blocked views, and an expert panel classed two as outright AI errors.
trace_subject: AI-assisted video review of head-strike and penalty calls in Olympic taekwondo
gain_reading: In a study of Paris 2024 taekwondo video reviews, ChatGPT-4.5 with OpenPose agreed almost perfectly with international referees (kappa 0.897) and cut average review time by about 81%.
gain_evidence: the AI system reduced average review time by approximately 81% | Kappa value of 0.897
problem_reading: The same AI disagreed with referees in nine reviews, mostly minimal-contact head strikes or blocked views, and an expert panel classed two as outright AI errors.
problem_evidence: primarily in scenarios involving head strikes with minimal contact or visual occlusion
quick_read: Researchers ran ChatGPT-4.5 and the OpenPose pose-tracking model over 241 usable video review clips from the 2024 Paris Olympic taekwondo competition and compared its penalty rulings with those of the international video review referees. In the formal test of 177 clips, the AI and the referees gave the same verdict in 168 and differed in nine, an agreement of kappa 0.897. The AI also cut average review time by about 81%.

Reviews can hold up matches: the 255 reviews at the Games ranged from 47 seconds to more than 30 minutes, while athletes had to stand and wait without sitting or drinking water. The AI took 3 to 9 seconds per clip. The authors say it is not ready to replace referees and propose a hybrid system in which AI picks out key frames and flags simple calls, and a referee confirms the result.

The mismatches clustered in ambiguous head-strike calls. An expert panel judged two of the nine to be AI errors and the other seven to be controversial judgments that still images could not settle. The authors point to multi-angle video and real-time detection as the next steps, and to other sports such as boxing, judo and basketball as places to test the approach.
limitation: The authors say AI cannot yet replace human referees: it struggled with minimal contact, occlusion and ambiguous calls, and the work covers one event, one sport and frame-by-frame analysis rather than real-time, multi-angle review.
tag: Dual reading
key_points: Of 255 video reviews at the 2024 Paris Olympic taekwondo event, 241 clips of 3 to 19 seconds were usable; 177 from the last three days made up the formal test. | In the formal test, the AI and referees agreed on 74 'No' and 94 'Yes' verdicts, with 6 AI-'Yes' versus referee-'No' and 3 AI-'No' versus referee-'Yes' splits. | Before the main test, a pre-experiment on day one matched referees in 59 of 63 videos (93.65%); a new rule on slight toe contact with headgear fixed all four misses. | The AI reviewed each clip in 3 to 9 seconds, against a human minimum of 47 seconds, which the authors put at at least 80.9% faster. | Head-contact reviews made up 210 of 255 reviews (82%), and a successful head-strike review awards 3 points, so errors can change a result. | The panel classed two of the nine disagreements (a women's 57 kg and a men's +80 kg bout) as AI errors, linked to referees having looked at few camera angles; the other seven were judged controversial. | The authors suggest sending a review to a human when AI confidence falls below 90% and letting the AI rule automatically above it.
rundown: The AI was first given the latest World Taekwondo rules, then shown 300 labelled images (100 successful and 100 failed head strikes, plus out-of-bounds and fall examples) and given feedback on each answer. It then searched footage for frames in which the foot came within 10 pixels of the opponent's helmet. Judges' final hand signals were removed from the clips so they could not cue the AI.

The pre-experiment covered all 63 reviews from day one, and the AI matched the on-site judges in 59. All four misses were head strikes the judges ruled successful and the AI called failures. After experts confirmed these as AI errors and a rule on slight toe contact was added, the AI matched the judges on all four. The formal phase did not include this kind of correction.

In the formal test of 177 clips, the Pearson correlation between referee and AI judgments was 0.898 and the chi-square tests were significant (p < 0.001). Referees gave 80 'No' verdicts against 77 from the AI. Head-strike disagreements involved women's 57 kg (three), women's 67 kg (two) and men's +80 kg (two) bouts, with one each in men's 68 kg and women's +67 kg.

In the seven controversial cases, two showed possible blocking of the face by the defender's hands that the AI missed. In five, referees ruled no strike while the AI saw foot-to-helmet contact from some angles. The authors say frame-by-frame analysis alone struggles with ambiguous calls and that a referee's reading of motion and rules is still needed.

The authors argue AI could help most at municipal, provincial and regional events that lack top referees, where it could cut referee labour and training costs and help train young referees. At major events such as the Olympics they see it as support only, partly because fans value human judgment and may accept unfavourable AI rulings less readily.
sources:
- peer_reviewed | Frontiers in Sports and Active Living | https://doi.org/10.3389/fspor.2025.1632326 | 2025-08-18
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

0e0e1153b6b0c7ddbdd9c3bc616a426e51d570beef2cde2e71de1fd48a21c60d

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000