Final video assets, README header

README.md

@@ -1,7 +1,5 @@
 <div align="center">
 
-<img src="docs/devpost/thumbnail.png" alt="GRIDPRINT: the power grid is the witness" width="820">
-
 # GRIDPRINT
 
 **Checks that a video is real by reading the power grid in the flicker of the room lights.**

docs/VIDEO_SCRIPT.md

@@ -1,9 +1,9 @@
 # GRIDPRINT submission video: script
 
-Generated by `video/make_pitch.py`. Voice: edge-tts `en-US-AndrewNeural` (rate +0%). Duration: 195.4 s. <!-- src: results/video/duration.json#duration_s -->
+Generated by `video/make_pitch.py`. Voice: edge-tts `en-US-AndrewNeural` (rate +0%). Duration: 223.2 s. <!-- src: results/video/duration.json#duration_s -->
 Style: ASD-STE100 Simplified Technical English (short sentences, active voice, one idea per sentence).
 
-Sources on screen: live AWS app screenshots (`video/assets/live/`, captured with headless Chrome), `data/public/fluorescent_beat_effect.mp4` (Wikimedia Commons, CC BY-SA 3.0), engine outputs in `results/video/engine_runs/`, agent trace `results/agent_traces/clean__s1001.json`, benchmark `results/bench/2026-10-27/`. <!-- src: results/bench/2026-10-27/cost.json#configs -->
+Sources on screen: live AWS app screenshots (`video/assets/live/`, captured with headless Chrome), `data/public/fluorescent_beat_effect.mp4` (Wikimedia Commons, CC BY-SA 3.0), engine outputs in `results/video/engine_runs/`, agent trace `results/agent_traces/clean__s1001.json`, benchmark `results/bench/2026-10-27/`, real iPhone 13 clip `data/real/phone2.mov` with engine run `results/video/engine_runs/real-phone2/` and live job `cd65b6ea0c20`. <!-- src: results/bench/2026-10-27/cost.json#configs -->
 
 ## 01 THE PROBLEM / hook  (00:00:00 - 00:00:07)
 
@@ -54,53 +54,65 @@ Sources on screen: live AWS app screenshots (`video/assets/live/`, captured with
 - Then it slides that track along the grid reference log.
 - One clear match gives a recording time. A cut or an insertion breaks the match.
 
-## 04 THE PRODUCT / home  (00:01:27 - 00:01:36)
+## 04 REAL EVIDENCE / realclip  (00:01:27 - 00:01:36)
+
+- We filmed a real room with a normal phone.
+- This is an iPhone 13 video, recorded today under indoor mains light in India.
+
+## 04 REAL EVIDENCE / realrun  (00:01:36 - 00:01:45)
+
+- The OpenCV 5 engine measured the full clip.
+- It found no clean flicker in this clip.
+- So GRIDPRINT said INCONCLUSIVE. It did not guess.
+
+## 04 REAL EVIDENCE / reallive  (00:01:45 - 00:01:56)
+
+- We also ran the same clip live on AWS.
+- Then the live agent asked the user one question: which camera made this video?
+- A missing signal never counts against a video.
+
+## 05 THE PRODUCT / home  (00:01:56 - 00:02:05)
 
 - GRIDPRINT runs live on AWS today. This is the web app.
 - You upload a video. You can add a claimed grid zone and a claimed recording time.
 
-## 04 THE PRODUCT / cases  (00:01:36 - 00:01:42)
+## 05 THE PRODUCT / cases  (00:02:05 - 00:02:12)
 
 - Ready-made example cases run through the same live pipeline.
 - One click starts a full analysis job.
 
-## 04 THE PRODUCT / job  (00:01:42 - 00:01:55)
+## 05 THE PRODUCT / job  (00:02:12 - 00:02:24)
 
 - Here is a finished job. The verdict is VERIFIED, and the report is final.
 - The page shows the estimated recording time and its uncertainty.
 - Each verdict comes with reason codes in plain words.
 
-## 04 THE PRODUCT / trace  (00:01:55 - 00:02:02)
+## 05 THE PRODUCT / trace  (00:02:24 - 00:02:32)
 
 - An agent plans each analysis.
 - It calls engine tools through an MCP server, and it logs every step.
 
-## 04 THE PRODUCT / reviewer  (00:02:02 - 00:02:09)
+## 05 THE PRODUCT / reviewer  (00:02:32 - 00:02:39)
 
 - A CONTRADICTED result is never final on its own.
 - A human reviewer must approve it first.
 
-## 04 THE PRODUCT / limits  (00:02:09 - 00:02:15)
-
-- The limitations page tells users what GRIDPRINT cannot do.
-- Clear limits are part of the product.
-
-## 05 BUILT ON AWS / arch  (00:02:15 - 00:02:48)
+## 06 BUILT ON AWS / arch  (00:02:39 - 00:03:20)
 
-- A video arrives through API Gateway and a load balancer.
-- An API service on Fargate stores it in S3 with KMS keys, and puts a job on an SQS queue.
-- Workers on Graviton4 run the OpenCV 5 engine in fifteen stages.
+- A video arrives through API Gateway and an Application Load Balancer.
+- An API service on AWS Fargate stores it in S3 with KMS keys, and puts a job on an SQS queue.
+- Workers on Amazon ECS with Graviton4 run the OpenCV 5 engine in fifteen stages.
 - An agent with twelve MCP tools uses Amazon Nova Lite on Bedrock to plan each run, inside strict policies.
-- DynamoDB keeps jobs and the grid log. IoT Core brings in grid data. Cognito signs in reviewers.
+- DynamoDB keeps jobs and the grid log. IoT Core and Lambda bring in grid data.
+- Cognito signs in reviewers. CloudWatch watches the system, and the AWS CDK deploys it all.
 
-## 05 BUILT ON AWS / bench  (00:02:48 - 00:02:58)
+## 06 BUILT ON AWS / bench  (00:03:20 - 00:03:29)
 
 - We measured the same engine on Graviton and on x86.
 - Graviton processed more video per second, at a lower cost per video hour.
 
-## 06 THE VERDICT / close  (00:02:58 - 00:03:14)
+## 07 THE VERDICT / close  (00:03:29 - 00:03:42)
 
 - GRIDPRINT gives exactly three verdicts: VERIFIED, CONTRADICTED, or INCONCLUSIVE.
 - A missing signal never counts against a video.
-- It needs mains lighting, and many phone videos will be INCONCLUSIVE.
 - GRIDPRINT. Evidence from the light in the room.

results/video/duration.json

@@ -1,8 +1,8 @@
 {
  "file": "gridprint_submission.mp4",
- "sha256": "58a9b27aaf043511cd3a3b179575b8263b2956375897c00b4a52a095e98f0e8b",
- "duration_s": 195.433333,
+ "sha256": "5260e9f06fb10c3d252ecb9349b344d76d92910ebf210db9fdea1f8c4989c188",
+ "duration_s": 223.166667,
  "limit_s": 300,
  "ok": true,
- "checked": "2026-10-27T05:24:38.693595+00:00"
+ "checked": "2026-10-27T06:31:05.694934+00:00"
 }

results/video/gridprint_submission.srt

@@ -99,98 +99,122 @@ Then it slides that track along the grid reference log.
 One clear match gives a recording time. A cut or an insertion breaks the match.
 
 26
-00:01:27,494 --> 00:01:31,284
-GRIDPRINT runs live on AWS today. This is the web app.
+00:01:27,494 --> 00:01:29,609
+We filmed a real room with a normal phone.
 
 27
-00:01:31,704 --> 00:01:35,764
-You upload a video. You can add a claimed grid zone and a claimed recording time.
+00:01:30,029 --> 00:01:35,507
+This is an iPhone 13 video, recorded today under indoor mains light in India.
 
 28
-00:01:36,764 --> 00:01:39,953
-Ready-made example cases run through the same live pipeline.
+00:01:36,507 --> 00:01:39,272
+The OpenCV 5 engine measured the full clip.
 
 29
-00:01:40,373 --> 00:01:42,321
-One click starts a full analysis job.
+00:01:39,692 --> 00:01:41,513
+It found no clean flicker in this clip.
 
 30
-00:01:43,321 --> 00:01:47,573
-Here is a finished job. The verdict is VERIFIED, and the report is final.
+00:01:41,933 --> 00:01:45,254
+So GRIDPRINT said INCONCLUSIVE. It did not guess.
 
 31
-00:01:47,993 --> 00:01:51,500
-The page shows the estimated recording time and its uncertainty.
+00:01:46,254 --> 00:01:48,774
+We also ran the same clip live on AWS.
 
 32
-00:01:51,920 --> 00:01:54,722
-Each verdict comes with reason codes in plain words.
+00:01:49,194 --> 00:01:53,437
+Then the live agent asked the user one question: which camera made this video?
 
 33
-00:01:55,722 --> 00:01:57,659
-An agent plans each analysis.
+00:01:53,857 --> 00:01:56,118
+A missing signal never counts against a video.
 
 34
-00:01:58,079 --> 00:02:02,353
-It calls engine tools through an MCP server, and it logs every step.
+00:01:57,118 --> 00:02:00,908
+GRIDPRINT runs live on AWS today. This is the web app.
 
 35
-00:02:03,353 --> 00:02:06,252
-A CONTRADICTED result is never final on its own.
+00:02:01,328 --> 00:02:05,388
+You upload a video. You can add a claimed grid zone and a claimed recording time.
 
 36
-00:02:06,672 --> 00:02:08,781
-A human reviewer must approve it first.
+00:02:06,388 --> 00:02:09,577
+Ready-made example cases run through the same live pipeline.
 
 37
-00:02:09,781 --> 00:02:13,295
-The limitations page tells users what GRIDPRINT cannot do.
+00:02:09,997 --> 00:02:11,945
+One click starts a full analysis job.
 
 38
-00:02:13,715 --> 00:02:15,254
-Clear limits are part of the product.
+00:02:12,945 --> 00:02:17,197
+Here is a finished job. The verdict is VERIFIED, and the report is final.
 
 39
-00:02:16,254 --> 00:02:19,867
-A video arrives through API Gateway and a load balancer.
+00:02:17,617 --> 00:02:21,123
+The page shows the estimated recording time and its uncertainty.
 
 40
-00:02:20,287 --> 00:02:27,030
-An API service on Fargate stores it in S3 with KMS keys, and puts a job on an SQS queue.
+00:02:21,543 --> 00:02:24,345
+Each verdict comes with reason codes in plain words.
 
 41
-00:02:27,450 --> 00:02:32,542
-Workers on Graviton4 run the OpenCV 5 engine in fifteen stages.
+00:02:25,345 --> 00:02:27,283
+An agent plans each analysis.
 
 42
-00:02:32,962 --> 00:02:40,429
-An agent with twelve MCP tools uses Amazon Nova Lite on Bedrock to plan each run, inside strict policies.
+00:02:27,703 --> 00:02:31,977
+It calls engine tools through an MCP server, and it logs every step.
 
 43
-00:02:40,849 --> 00:02:47,867
-DynamoDB keeps jobs and the grid log. IoT Core brings in grid data. Cognito signs in reviewers.
+00:02:32,977 --> 00:02:35,876
+A CONTRADICTED result is never final on its own.
 
 44
-00:02:48,867 --> 00:02:52,443
-We measured the same engine on Graviton and on x86.
+00:02:36,296 --> 00:02:38,404
+A human reviewer must approve it first.
 
 45
-00:02:52,863 --> 00:02:57,478
-Graviton processed more video per second, at a lower cost per video hour.
+00:02:39,404 --> 00:02:43,734
+A video arrives through API Gateway and an Application Load Balancer.
 
 46
-00:02:58,478 --> 00:03:04,149
-GRIDPRINT gives exactly three verdicts: VERIFIED, CONTRADICTED, or INCONCLUSIVE.
+00:02:44,154 --> 00:02:51,621
+An API service on AWS Fargate stores it in S3 with KMS keys, and puts a job on an SQS queue.
 
 47
-00:03:04,569 --> 00:03:06,830
-A missing signal never counts against a video.
+00:02:52,041 --> 00:02:58,525
+Workers on Amazon ECS with Graviton4 run the OpenCV 5 engine in fifteen stages.
 
 48
-00:03:07,250 --> 00:03:10,866
-It needs mains lighting, and many phone videos will be INCONCLUSIVE.
+00:02:58,945 --> 00:03:06,412
+An agent with twelve MCP tools uses Amazon Nova Lite on Bedrock to plan each run, inside strict policies.
 
 49
-00:03:11,286 --> 00:03:13,837
+00:03:06,832 --> 00:03:12,693
+DynamoDB keeps jobs and the grid log. IoT Core and Lambda bring in grid data.
+
+50
+00:03:13,113 --> 00:03:19,648
+Cognito signs in reviewers. CloudWatch watches the system, and the AWS CDK deploys it all.
+
+51
+00:03:20,648 --> 00:03:24,224
+We measured the same engine on Graviton and on x86.
+
+52
+00:03:24,644 --> 00:03:29,259
+Graviton processed more video per second, at a lower cost per video hour.
+
+53
+00:03:30,259 --> 00:03:35,931
+GRIDPRINT gives exactly three verdicts: VERIFIED, CONTRADICTED, or INCONCLUSIVE.
+
+54
+00:03:36,351 --> 00:03:38,612
+A missing signal never counts against a video.
+
+55
+00:03:39,032 --> 00:03:41,582
 GRIDPRINT. Evidence from the light in the room.