Release-claim watchlist · Method 1.0.0

Announced. But available?

AI Release Tracker follows product release claims in our published news archive. Use it to revisit an announcement, inspect its recorded sources, and decide what needs checking next.

Snapshot assembled . Latest matching article: . This is not a source-access or availability-check timestamp.

Release-claim articles
223
Availability unchecked
223
7-day checks due
210
30-day checks due
84
Availability not independently checked. Every entry is an automatically selected claim, including stories that say “available now.” Unknown availability does not mean unavailable. No access tests or completed follow-ups are recorded here.

Where every claim stands

Follow-up status

Suggested check dates fall 7 and 30 days after each article was published. “Due” means the date has passed; no check has been completed for any claim.

7-day suggested checks

223 claims
  • Due; not completed 210
  • Upcoming; not scheduled 13
  • Date unknown 0

30-day suggested checks

223 claims
  • Due; not completed 84
  • Upcoming; not scheduled 139
  • Date unknown 0

Watchlist mentions

Articles mentioning
  • OpenAI50
  • Google / DeepMind43
  • NVIDIA36
  • Anthropic32
  • Meta27
  • Alibaba / Qwen21
  • Microsoft16
  • DeepSeek13

Mentions, not confirmed ownership: one article can mention several organizations.

The watchlist

Find a release claim

“Major” means names on our explicit watchlist, not a verified importance ranking. Organization matches are mentions, not confirmed product ownership. Dates are UTC.

Reset filters

223 matching articles · page 9 of 12. Export includes every matching record, not just this page.

Article published · Innovation

Google and NASA JPL Unveil AI Model Mapping Global Methane Plumes

Availability not independently checked

Google and NASA's Jet Propulsion Laboratory (JPL) introduced MAPL-EMIT, a deep-learning model that detects, quantifies, and localizes methane plumes globally from NASA's EMIT satellite instrument.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “Unveil”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • unite.aihttps://www.unite.ai/google-and-nasa-jpl-unveil-ai-model-mapping-global-methane-plumes/Recorded: Unknown · Article source field
  • unite.aihttps://www.unite.ai/google-and-nasa-jpl-unveil-ai-model-mapping-global-methane-plumes/Recorded: · Source history

Article published · Product

Introducing ChatGPT Images 2.5

Availability not independently checked

OpenAI introduces ChatGPT Images 2.5, a new state-of-the-art image model that brings sharper details, more precise editing, and faster generation to creative workflows.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “Introducing”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • openai.comhttps://openai.com/index/introducing-chatgpt-images-2-5/Recorded: Unknown · Article source field
  • openai.comhttps://openai.com/index/introducing-chatgpt-images-2-5/Recorded: · Source history

Article published · Product

Meta launches AI agent that can access other apps to send emails, make payments

Availability not independently checked

Meta launches AI agent that can access other apps to send emails, make payments Meta has launched an AI agent called Muse that can access other apps to send emails, make payments, and perform other tasks. The agent is designed to be a personal assistant that can automate various tasks and provide users with a more…

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “launches”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (3)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • businesstimes.com.sghttps://www.businesstimes.com.sg/companies-markets/telcos-media-tech/meta-launches-ai-agent-can-access-other-apps-send-emails-make-paymentsRecorded: Unknown · Article source field
  • virginiabusiness.comhttps://virginiabusiness.com/meta-launches-muse-ai-agent-us-app-access/Recorded: · Source history
  • businesstimes.com.sghttps://www.businesstimes.com.sg/companies-markets/telcos-media-tech/meta-launches-ai-agent-can-access-other-apps-send-emails-make-paymentsRecorded: · Source history

Article published · Product

Meta launches personal AI agent Muse for people 18 and over emphasizing safety and privacy

Availability not independently checked

Meta launched a personal artificial intelligence agent Muse for people 18 and over who are looking for help with day-to-day tasks like schedules shopping and turning long-term goals into action plans emphasizing safety and privacy

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “launches”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • wral.comhttps://www.wral.com/news/ap/3a457-meta-launches-personal-ai-agent-muse-emphasizes-safety-and-privacy/Recorded: Unknown · Article source field
  • wral.comhttps://www.wral.com/news/ap/3a457-meta-launches-personal-ai-agent-muse-emphasizes-safety-and-privacy/Recorded: · Source history

Article published · Product

Jensen Huang says AGI has arrived after GPT-6 Astra launch

Availability not independently checked

Firstpost reports that Nvidia CEO Jensen Huang declared AGI had arrived after OpenAI launched GPT-6 Astra, while noting that the claim is not independently confirmed and lacks a universally accepted technical definition.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “launch”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • firstpost.comhttps://www.firstpost.com/tech/nvidia-ceo-jensen-huang-says-agi-has-arrived-after-openai-launches-gpt-6-astra-14043993.htmlRecorded: Unknown · Article source field
  • firstpost.comhttps://www.firstpost.com/tech/nvidia-ceo-jensen-huang-says-agi-has-arrived-after-openai-launches-gpt-6-astra-14043993.htmlRecorded: · Source history

Article published · Product

OpenBMB releases open 2B model with reported Japanese-language strength

Availability not independently checked

GIGAZINE reports that Chinese AI company OpenBMB released MiniCPM5-2B, an open 2-billion-parameter model that scored comparably to the larger Gemma 4 12B on Artificial Analysis’s Intelligence Index.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “releases”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • gigazine.nethttps://gigazine.net/gsc_news/en/20260908-openbmb-minicpm5-2b/Recorded: Unknown · Article source field
  • gigazine.nethttps://gigazine.net/gsc_news/en/20260908-openbmb-minicpm5-2b/Recorded: · Source history

Article published · Security

Proofpoint launches AI SOC Analyst Agent with OpenAI models

Availability not independently checked

SecurityBrief Asia reports that Proofpoint has launched an AI SOC Analyst Agent using OpenAI Daybreak models to investigate threats across its security data. The product is in private preview and is designed to recommend actions while leaving consequential decisions to human analysts.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “launches”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • securitybrief.asiahttps://securitybrief.asia/story/proofpoint-launches-ai-soc-analyst-agent-with-openaiRecorded: Unknown · Article source field
  • securitybrief.asiahttps://securitybrief.asia/story/proofpoint-launches-ai-soc-analyst-agent-with-openaiRecorded: · Source history

Article published · Security

Regulators warn of frontier AI cyber risks as GPT-6 Astra is released

Availability not independently checked

Pinsent Masons reports that GPT-6 Astra’s release coincided with warnings that frontier AI could accelerate vulnerability discovery and cyberattacks.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “released”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • pinsentmasons.comhttps://www.pinsentmasons.com/out-law/news/frontier-ai-cyber-risk-warnings-grow-louder-gpt-6-astraRecorded: Unknown · Article source field
  • pinsentmasons.comhttps://www.pinsentmasons.com/out-law/news/frontier-ai-cyber-risk-warnings-grow-louder-gpt-6-astraRecorded: · Source history

Article published · Product

OpenAI introduces GPT-6 Astra with reported coding and computer-use gains

Availability not independently checked

RTTNews reports that OpenAI has introduced GPT-6 Astra, citing major benchmark gains, limited initial access, planned broader availability and API pricing starting at $10 per million input tokens.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “introduces”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (3)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • rttnews.comhttps://www.rttnews.com/amp/3688689/openai-introduces-gpt-6-astra-with-major-advances-in-ai-and-coding.aspxRecorded: Unknown · Article source field
  • openai.comhttps://openai.com/index/gpt-6-astra/Recorded: · Source history
  • rttnews.comhttps://www.rttnews.com/amp/3688689/openai-introduces-gpt-6-astra-with-major-advances-in-ai-and-coding.aspxRecorded: · Source history

Article published · Product

OpenAI releases GPT-6 Astra to limited customers amid safety concerns

Availability not independently checked

Breitbart reports that OpenAI has released GPT-6 Astra to a limited group of customers, with broader access expected soon, while the company applies stricter safeguards because of the model’s reported cyber capabilities.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “releases”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • breitbart.comhttps://www.breitbart.com/tech/2026/09/04/sam-altmans-openai-releases-astra-ai-model-claiming-the-agi-era-is-here/Recorded: Unknown · Article source field
  • breitbart.comhttps://www.breitbart.com/tech/2026/09/04/sam-altmans-openai-releases-astra-ai-model-claiming-the-agi-era-is-here/Recorded: · Source history

Article published · Product

llama.cpp 0.4.0 adds support for newer AI models and video input

Availability not independently checked

The llama.cpp 0.4.0 release adds initial support for Qwen3.8-Flash-Next and NVIDIA Nemotron-3-Puzzle, video-input options, per-slot server context limits, lazy tensor reading, and ggml 0.23.0 changes.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “adds”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (3)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • github.comhttps://github.com/ggml-org/llama.cpp/releases/tag/v0.4.0Recorded: Unknown · Article source field
  • github.comhttps://github.com/ggml-org/llama.cpp/releases/tag/b10776Recorded: · Source history
  • github.comhttps://github.com/ggml-org/llama.cpp/releases/tag/v0.4.0Recorded: · Source history

Article published · Product

Anthropic publishes open reference blueprint for commerce AI agents

Availability not independently checked

Anthropic has published a public Apache-2.0 repository with shopping and merchant agents built on Claude, including runnable retail, travel, telecom and entertainment examples.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “open”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • github.comhttps://github.com/anthropics/commerce-agentsRecorded: Unknown · Article source field
  • github.comhttps://github.com/anthropics/commerce-agentsRecorded: · Source history

Article published · Product

Claude Fable 5.1 is available in GitHub Copilot

Availability not independently checked

GitHub announced Claude Fable 5.1 for eligible Copilot plans, with gradual rollout and data-retention conditions that administrators should review.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “available”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • GitHubhttps://github.blog/changelog/2026-09-01-claude-fable-5-1-generally-available-in-github-copilot/Recorded: Unknown · Article source field
  • GitHubhttps://github.blog/changelog/2026-09-01-claude-fable-5-1-generally-available-in-github-copilot/Recorded: · Source history

Article published · Product

Meta opens Muse Spark 1.3 reasoning mode and reports lower tool use

Availability not independently checked

Meta releases Muse Spark 1.3, enhancing its capabilities for long tasks and coding. Meta has released Muse Spark 1.3, a new version of its AI model that focuses on long-running tasks and coding. The model has been trained to reduce tool calls and improve efficiency in complex tasks. Muse Spark 1.3 has been tested…

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “opens”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (8)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • eu.36kr.comhttps://eu.36kr.com/en/p/3967796330082819Recorded: Unknown · Article source field
  • techcrunch.comhttps://techcrunch.com/2026/09/01/googles-answer-to-canva-is-an-ai-tool-where-you-prompt-instead-of-design/Recorded: · Source history
  • blog.googlehttps://blog.google/products-and-platforms/products/workspace/google-pics/Recorded: · Source history
  • theverge.comhttps://www.theverge.com/tech/987423/google-pics-ai-image-editor-generatorRecorded: · Source history
  • phoneworld.com.pkhttps://www.phoneworld.com.pk/google-introduces-gemini-3-8-flash-and-3-8-flash-cyber-with-stronger-ai-and-security-skills/Recorded: · Source history
  • digitaltoday.co.krhttps://www.digitaltoday.co.kr/en/view/99891/meta-unveils-muse-spark-1-3-claims-performance-rivals-anthropic-and-openaiRecorded: · Source history
  • eu.36kr.comhttps://eu.36kr.com/en/p/3967349130796296Recorded: · Source history
  • eu.36kr.comhttps://eu.36kr.com/en/p/3967796330082819Recorded: · Source history

Article published · Product

Blockchain.News reports Lyria 3.5 launch in Google’s Gemini app

Availability not independently checked

Blockchain.News reports that Google launched Lyria 3.5 in the Gemini app and API, with tracks of up to three minutes and integrations across several Google products. The report’s claims have not been independently confirmed here.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “launch”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (4)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • blockchain.newshttps://blockchain.news/news/google-lyria-3-5-gemini-ai-musicRecorded: Unknown · Article source field
  • deepmind.googlehttps://deepmind.google/blog/introducing-agentic-video-in-gemini/Recorded: · Source history
  • fortune.comhttps://fortune.com/2026/09/03/google-shipped-four-gemini-flash-models-in-106-days-but-its-flagship-frontier-model-is-still-nowhere-to-be-seen/Recorded: · Source history
  • blockchain.newshttps://blockchain.news/news/google-lyria-3-5-gemini-ai-musicRecorded: · Source history

Article published · Product

GitHub lets Copilot submit pull-request approvals in public preview

Availability not independently checked

GitHub says Copilot code review can now submit approvals that count toward required repository approvals when administrators enable the feature.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “preview”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • github.bloghttps://github.blog/changelog/2026-09-01-copilot-code-review-can-now-approve-pull-requestsRecorded: Unknown · Article source field
  • github.bloghttps://github.blog/changelog/2026-09-01-copilot-code-review-can-now-approve-pull-requestsRecorded: · Source history

Article published · Product

AWS makes Claude Fable 5.1 available on Bedrock with new data controls

Availability not independently checked

AWS says Claude Fable 5.1 is now available through Amazon Bedrock and Claude Platform on AWS, while eligible customers can use a separate safeguards program with zero data retention.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “available”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (4)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • aws.amazon.comhttps://aws.amazon.com/blogs/machine-learning/introducing-claude-fable-5-1-on-aws/Recorded: Unknown · Article source field
  • eu.36kr.comhttps://eu.36kr.com/en/p/3965506148652681Recorded: · Source history
  • anthropic.comhttps://www.anthropic.com/claude-fable-and-mythos-5-1Recorded: · Source history
  • aws.amazon.comhttps://aws.amazon.com/blogs/machine-learning/introducing-claude-fable-5-1-on-aws/Recorded: · Source history

Article published · Product

DigitalToday reports SK Telecom launches A. log AI photo-record service

Availability not independently checked

DigitalToday reports that SK Telecom launched A. log, an AI feature in its A.dot app that organizes and searches photos taken with Meta AI glasses and Android smartphones. The company says photos remain on users’ devices and sensitive information can be automatically masked.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “launches”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • digitaltoday.co.krhttps://www.digitaltoday.co.kr/en/view/97968/skt-launches-a-dot-log-to-organise-and-search-ai-glasses-photosRecorded: Unknown · Article source field
  • digitaltoday.co.krhttps://www.digitaltoday.co.kr/en/view/97968/skt-launches-a-dot-log-to-organise-and-search-ai-glasses-photosRecorded: · Source history

Article published · Product

Anthropic makes Claude in Chrome generally available with autonomous browser actions

Availability not independently checked

BigGo Finance reports that Anthropic has broadly released Claude in Chrome, allowing Claude to navigate webpages, enter text and fill forms with reduced approval prompts. The company says layered defenses cut attack success rates to 0–0.3% in its latest evaluations, though those results have not been independently…

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “available”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • finance.biggo.comhttps://finance.biggo.com/news/04bd1c2f-2660-4a13-8067-472a7c776690Recorded: Unknown · Article source field
  • finance.biggo.comhttps://finance.biggo.com/news/04bd1c2f-2660-4a13-8067-472a7c776690Recorded: · Source history

Article published · Product

The Star reports Harvey built legal model on Chinese Kimi K3 open weights

Availability not independently checked

The Star reports that OpenAI-backed legal technology company Harvey released Harvey Tenet, a legal model post-trained on Moonshot AI’s open-weight Kimi K3. Harvey claims the model improves performance and cost efficiency, but those claims have not been independently confirmed.

Above: AI Understanding’s article text, not a quotation from a linked source. Automatic title match: “open”. Announcement date: unknown. Article updated: .

7-day suggested check
· Due; not completed
30-day suggested check
· Due; not completed
Recorded source links (2)

Source types are unclassified. Links have not been opened by this tracker; source publication and access dates are unknown. A recorded date only indicates when the archive added the link.

  • thestar.com.myhttps://www.thestar.com.my/aseanplus/aseanplus-news/2026/08/29/openai-backed-legal-tech-firm-pivots-to-chinese-kimi-k3-open-weight-modelRecorded: Unknown · Article source field
  • thestar.com.myhttps://www.thestar.com.my/aseanplus/aseanplus-news/2026/08/29/openai-backed-legal-tech-firm-pivots-to-chinese-kimi-k3-open-weight-modelRecorded: · Source history

Open method · 1.0.0

Methodology & limits

01Selection

Title must contain release wording and either product wording or have the Product category. A watchlist name must appear in the title or excerpt. Mentions are not verified product ownership. This is an automatic candidate filter, not independent verification or an importance ranking.

02Unit

One canonical published article matching the watchlist and title rules; not one unique product or launch.

03Follow-ups

Suggested checks at 7 and 30 UTC calendar days after article publication (a proxy, not the announcement date). Due means the suggested date has arrived, not that a check ran. No reminders or checks are scheduled. Updated articles retain their original publication anchor.

04Coverage

Only articles returned by getAllNewsArticles, including its existing publication, suppression and media filters. That reader can silently fall back to bundled articles when the database is unavailable. Feed completeness cannot be established here; this is not a census of AI releases.

05Reproducibility

Download and retain the JSON snapshot: article text and evidence links may change with editorial updates. generatedAt is dataset assembly time, not a source access or verification time. Filters affect records, not the full-watchlist summary. Records sort newest first, undated last, then by canonical URL.

06Data source

Data source: AI Understanding’s published article archive (database and bundled articles). The existing reader does not expose which origin supplied each record or whether fallback occurred. Linked evidence is unclassified, not automatically primary.

  • Keyword selection can include rumors, commentary, comparisons and non-launch announcements, and can miss releases using other wording or names.
  • Several articles can concern one release; one article can concern several products. Mentioned organizations can be competitors rather than the releasing organization.
  • Region, account tier, waitlist, rollout stage, price and actual access remain unknown. This tracker does not measure model quality or run workflow tests.

What would count as confirmed? A separately documented human access check with time, region, account tier, product version, steps and evidence of a successful representative action. It would confirm only those conditions. This version records none; source claims, including first-party announcements, never become confirmation automatically.

The JSON contains the selection patterns, full watchlist, matched wording, dates, source provenance, unchecked status and suggested follow-ups. Missing dates remain null. Report a mismatch or missing source.

Protocol only · No results yet

Can three tools turn a source packet into usable nonprofit work?

Our first repeatable test design compares 3 tools × 10 tasks × 3 repetitions = 90 planned runs. It covers factual extraction, briefings, arithmetic, outreach, planning, privacy and uncertainty using the same synthetic source packet.

  1. Before running: select three accessible tools, freeze their versions and settings, publish the task packet and answer key, verify current prices, and approve a total cash and human-time cap.
  2. Run each task in a fresh session three times with identical inputs and rotated tool order. Preserve failures, refusals, outputs, time and cost. Stop at the approved cap.
  3. A human scores correctness, completeness, instruction compliance and usability from 0–2 each. A pass needs at least 7/8, full correctness and compliance, and no critical error. A second reviewer checks every proposed pass and critical error.
  4. Publish all denominators, disagreements, missing runs and limitations. Do not report results until runs and human review exist. Three repetitions cannot establish broad model superiority.

The full execution protocol, ten tasks, shared fixture, answer key, scoring and budget preflight live in the repository document docs/AIU-ORIGINAL-TEST-PROTOCOL.md. Tool selection, funding and execution are still pending.

Get the weekly AI Understanding briefing