AI-Agent Platform

Yes — and the improvement survives the opponent correction. Raw numbers first: across 9 February and 8 April matches, the median time from losing the ball to a re-formed defensive shape falls from 6.8 s to 5.6 s. Because April’s calendar was softer, I grouped all 17 opponents by build-up style and compared within each group:

• Against short-passing build-up sides — 7.4 s in February, 5.9 s in April (11 matches, 312 loss-of-possession windows)
• Against direct, long-ball sides — 6.1 s in February, 5.4 s in April (6 matches, 141 windows)

The gain is larger against the sides that hold the ball — which is where a counterpressing change should show first. Recovery shape is measured from all 10 outfield skeletons at once, so substitutions and formation switches are already absorbed.

Caveat: the direct-opponent group holds only 3 April matches. Treat 5.4 s as provisional until two or three more of those fixtures are in — the short-passing number stands on firmer ground.

evidence · 17 matches · 453 recovery windows · split: 2 opponent-style groups

Mostly the double pivot. Ranked by the change in median re-entry time into the compact block:

• No. 6 — 1.9 s faster than in February; the largest single shift in the squad (78 windows)
• No. 8 — 1.4 s faster (74 windows)
• Back line — unchanged within ±0.2 s; it was never the target of the change

The 24 clearest before/after window pairs are queued in your review list, newest first.

Individual samples are small — 31 to 78 windows per player — so the ordering below the top two is not yet stable.

evidence · per-player split · 31–78 windows each · 24 linked pairs

148 pool players have enough processed footage to score against all three criteria. Three clear the profile threshold (0.85 combined similarity):

• Candidate A — 0.91 · first-two-metre burst in the top decile of the pool · turns barely slow under pressure · 1,840 min processed
• Candidate B — 0.88 · the steadiest late-match mechanics of the three: post-75 drift near zero · 1,120 min
• Candidate C — 0.86 · the cleanest turning shape in the pool · 640 min — the thinnest sample here

One reading note: these scores describe movement style, measured from skeleton windows. They say nothing about positioning, aerial timing, or decision-making — order the watchlist with them, then let the scouts do the watching.

Confidence: profiles stabilise at roughly 900 processed minutes. Candidate C sits at 640, so B and C could swap places as more of C’s footage lands.

evidence · 148 players scored · 83,400 windows · criteria: 3 · above threshold: 3

Overall he is close: across this week’s 4 sessions, movement similarity to his pre-injury baseline is 0.94, up from 0.83 in week 6. Most movement families have converged. Two deviations persist:

• Hard decelerations — right-side stance time runs 6% longer than left; his pre-injury norm was 2% (19 of 46 deceleration windows flagged)
• Top-speed strides — right stride length 3% short of his baseline above 7 m/s (8 of 52 sprint windows flagged)

Both gaps have narrowed every week since week 6 — the linked week-by-week view shows the two curves against his baseline band.

As always: this measures movement change against his own baseline. It is not a medical diagnosis, and the return decision stays with you.

evidence · baseline: 21 pre-injury sessions · week 9: 4 sessions · 1,412 windows compared · similarity: 0.94

It builds. In the first 20 minutes of a session, 3 of the 19 flagged windows occur; between minutes 20 and 40, 5; after minute 40, 11. The same late-session clustering appears in all four sessions this week — it reads as fatigue-linked rather than fixed.

The 19 flagged windows are linked with timestamps, in session order.

Within-session splits rest on small counts; the direction is consistent, the exact ratios will still move.

evidence · 46 deceleration windows · flagged: 19 · split: session minute 0–20 / 20–40 / 40+

Retrieval report: I searched 61,209 movement windows across 38 matches. 47 sequences match the overlap-plus-cut-inside pattern; requiring a set block — opposition back line formed and settled for at least 4 s before the overlap starts — removes 19 transition cases. 28 remain, spread over 14 matches:

• 9 of 28 are the early variant — the winger’s cut starts before the right-back reaches the halfway line; these stretch the block the most
• Frequency is stable home and away — 15 at home, 13 away

All 28 sequences are in your review list, cued to start six seconds before the overlap begins. The 19 excluded transition cases are kept in a separate list, in case you want them after all.

Count honesty: retrieval is similarity-based. At the current threshold (0.85) you get these 28; at 0.80 the pattern gains 11 borderline sequences — listed separately, not mixed in.

evidence · 38 matches · 61,209 windows searched · retrieved: 28 · borderline at looser threshold: 11

It is improving, with one flat spot. Across your 11 usable sessions:

• Stride symmetry at speed — the left–right difference fell from 4.9% in April to 2.8% now
• Hard stops — steadier month over month: flagged deceleration windows are down from 9 to 3 per session
• First two steps — unchanged since April; if you want one thing to work on, this is it

Each number links to your own clips, best and weakest side by side — the fastest way to see what 2.8% looks like.

Capture note: 3 of your 14 recordings were too dark or too shaky for reliable skeletons and are left out. The trend holds on the 11 clean ones — evening pitch lighting is the usual culprit.

evidence · 11 of 14 sessions usable · 3,120 windows · trend: improving