Skip to content

Releases: CodingIsCool666/attentrix

ATTENTRIX 0.2.0 Beta 2

Pre-release

Choose a tag to compare

@CodingIsCool666 CodingIsCool666 released this 30 Jul 21:41

ATTENTRIX 0.2.0 Beta 2

ATTENTRIX is a local presenter dashboard for visible audience cues in Zoom
Gallery View. This Windows release preserves the frozen Live v20 model,
gallery_person_relative_v1 presenter policy, and
primary_hand_1440_q82 compute profile.

Beta 2 replaces Beta 1. It restores the narwhals.stable.v2 runtime dependency
required by scikit-learn during Zoom analysis and introduces the official
ATTENTRIX emblem across the application, website, executable, and installer.

Included

  • Live analysis for up to 49 visible Zoom tiles
  • Per-participant highly engaged, following, confused, and attention-drift cues
  • Room-level presenter guidance
  • Background analysis of saved MP4 and WebM meetings
  • Optional local live-session replay, off by default
  • Synchronized video and room-timeline review
  • Installer and portable Windows packages

Privacy

Processing stays on the presenter's computer. ATTENTRIX does not request
microphone access, upload participant frames, or create identity profiles.
Uploaded meetings and enabled live replays are stored only in the current
user's local ATTENTRIX folder.

Important Windows Notice

This research beta is not code-signed. Windows may display an
unrecognized-app warning. Verify the package before running it:

5c34d934125829b6eae5c7fc57cfd87e616d1b63ac7c17a9ba05a80cccc244af  ATTENTRIX-Setup-0.2.0-beta.2-x64.exe
ebf47faf3f68ee3dcc36fd388cfcc40cfb1a6337d3e93a1183e7553493556b69  ATTENTRIX-Windows-x64-0.2.0-beta.2.zip

Verification And Provenance

Beta 2 was built from clean source commit
c8e923a6b9054a057b3d33eb40a1beefe0103fe5. The exact packaged application
passed startup and real saved-meeting analysis checks. The release also passed
793 Python tests, 18 website checks, and security audits of the release Python
environment and website dependencies.

The installer, portable archive, checksum file, manifest, and workflow
verification record are published with this release so the downloadable files
can be checked independently.

Current Evidence

The selected participant policy reached person-level macro-F1 0.751 and
accuracy 0.778 on one frozen three-person Zoom holdout with broad phase
labels. Room-level presenter-action macro-F1 was 0.851.

The compute profile preserved presenter outputs exactly across eight paired
replay trials while reducing backend analysis time by a source-mean 11.41%
with a 95% bootstrap interval of 8.11% to 14.71%.

These are local research results, not a universal classroom accuracy claim.

Limits

ATTENTRIX analyzes only the visible Zoom gallery page. It cannot interpret
cameras that are off, fully hidden faces, or participants on another page. Eye
gaze is not included in this release. The outputs are uncertain behavioral
estimates and must not be used for grading, discipline, attendance, employment
decisions, or covert surveillance.

ATTENTRIX 0.2.0 Beta 1

Pre-release

Choose a tag to compare

@CodingIsCool666 CodingIsCool666 released this 30 Jul 16:53

ATTENTRIX 0.2.0 Beta 1

ATTENTRIX is a local presenter dashboard for visible audience cues in Zoom
Gallery View. This Windows beta preserves the frozen Live v20 model,
gallery_person_relative_v1 presenter policy, and
primary_hand_1440_q82 compute profile.

Included

  • Live analysis for up to 49 visible Zoom tiles
  • Per-participant highly engaged, following, confused, and attention-drift cues
  • Room-level presenter guidance
  • Background analysis of saved MP4 and WebM meetings
  • Optional local live-session replay, off by default
  • Synchronized video and room-timeline review
  • Installer and portable Windows packages

Privacy

Processing stays on the presenter's computer. ATTENTRIX does not request
microphone access, upload participant frames, or create identity profiles.
Uploaded meetings and enabled live replays are stored only in the current
user's local ATTENTRIX folder.

Important Windows Notice

This research beta is not code-signed. Windows may display an
unrecognized-app warning. Verify the package before running it:

150cbef841e24c23cd0955b8a14ee83f9983329c928d5a4ea864b833c5935466  ATTENTRIX-Setup-0.2.0-beta.1-x64.exe
80298ba63bde9a4299edee72c7ce59399b132dce53aa7a2dc0eeb90886166b74  ATTENTRIX-Windows-x64-0.2.0-beta.1.zip

Verification And Provenance

The installer, portable archive, checksum file, manifest, and
workflow verification record
were retrieved anonymously from this release and matched against their local
sizes and SHA-256 values.

The manifest records that this tested beta was built from a development
snapshot with tracked changes. The hashes identify the exact distributed
artifacts, but the source commit alone cannot reproduce them. A future public
build should be produced from a clean source commit.

Current Evidence

The selected participant policy reached person-level macro-F1 0.751 and
accuracy 0.778 on one frozen three-person Zoom holdout with broad phase
labels. Room-level presenter-action macro-F1 was 0.851.

The compute profile preserved presenter outputs exactly across eight paired
replay trials while reducing backend analysis time by a source-mean 11.41%
with a 95% bootstrap interval of 8.11% to 14.71%.

These are local research results, not a universal classroom accuracy claim.

Limits

ATTENTRIX analyzes only the visible Zoom gallery page. It cannot interpret
cameras that are off, fully hidden faces, or participants on another page. Eye
gaze is not included in this release. The outputs are uncertain behavioral
estimates and must not be used for grading, discipline, attendance, employment
decisions, or covert surveillance.

SignalRoom 0.1.0 Beta 2

Pre-release

Choose a tag to compare

@CodingIsCool666 CodingIsCool666 released this 28 Jul 18:25

SignalRoom 0.1.0 Beta 2 keeps the frozen Live v20 engagement model, the selected gallery_person_relative_v1 participant policy, and the primary_hand_1440_q82 compute profile.

What is new

  • Analyze saved MP4 or WebM meetings locally in the background.
  • Optionally save a video-only live replay. Replay is off by default and resets to off for every new session.
  • Review video and participant cues beside a synchronized room timeline.
  • Return home from a live session without leaving the local model locked.
  • Updated Windows package and public documentation.

Privacy

SignalRoom does not request microphone access, upload participant frames, or create identity profiles. Reports, uploaded recordings, and explicitly enabled replay videos remain on the presenter's Windows computer.

Verification

  • 758 Python tests passed.
  • 16 production website checks passed.
  • 20 of 20 release-readiness checks passed.
  • Portable, silent install, installed executable, and silent uninstall checks passed.
  • Python and website dependency audits reported no known vulnerabilities.

This is an unsigned research beta. Windows may show an unrecognized-app warning. Verify the SHA-256 checksums before running it. Beta 1 remains available as the no-video-retention rollback release.

SignalRoom 0.1.0 Beta 1

Pre-release

Choose a tag to compare

@CodingIsCool666 CodingIsCool666 released this 26 Jul 22:46

SignalRoom 0.1.0 Beta 1

SignalRoom gives presenters a second-screen view of visible audience cues in
Zoom Gallery View. It separates highly engaged, following, confused, and
attention-drift signals and turns the room state into a glanceable presenter
response.

Included

  • Local Windows inference using the frozen SignalRoom Live v20 model
  • Validated gallery_person_relative_v1 presenter policy
  • Selected primary_hand_1440_q82 compute profile
  • Up to 49 participants rendered on one visible Zoom gallery page
  • Local session timeline and summary
  • Installer and portable package for 64-bit Windows 10 and Windows 11

Privacy

  • No microphone access
  • No retained meeting video
  • No upload of meeting frames or reports to a SignalRoom cloud service
  • No participant identity profiles
  • Research recording and development controls are disabled

Evidence

On one frozen three-person Zoom holdout with recording-level phase labels, the
selected policy reached person-level macro-F1 0.751 and accuracy 0.778.
Room-level presenter-action macro-F1 was 0.851.

Across eight final paired replay trials, the selected compute profile preserved
presenter outputs exactly and reduced backend analysis time by a source-mean
11.41%, with a 95% bootstrap interval of 8.11% to 14.71%.

These results are local research evidence, not a universal classroom accuracy
claim.

Before Installing

This beta is not code-signed. Windows may display an unrecognized-app warning.
Verify the installer or portable archive against SHA256SUMS.txt.

Eye gaze is not included in this release. It may be added later as an optional,
separately validated versioned update.