The Threshold Report tracks one question: what crossed the line from AI demo, rumor, or roadmap into something real?
It is not a hype meter and not an "AGI progress" gauge. It is a ledger. Every point on the index traces to a logged event with a primary source you can click. This page is the complete scoring system. The model suggests entries, but the code assigns points from the fixed rubric.
| Lane | Covers | In plain English |
|---|---|---|
| Text | Language models, reasoning, coding, long-context work, research/document analysis, writing, translation | AI is getting better at turning messy instructions into usable work. |
| Audio | Speech-to-text, text-to-speech, voice agents, real-time translation, music, dubbing | AI is getting better at listening, speaking, translating, and sounding human. |
| Video | Text/image-to-video, editing, avatars, scene consistency, motion and physics realism | AI video is moving from weird clips toward usable scenes. |
| Agents | Tool use, computer/browser use, digital task completion, autonomous workflows, cross-modal reasoning that is not primarily physical robotics | AI is moving from answering questions to doing multi-step digital tasks. |
| Robots | AI-enabled physical robots, including humanoids and non-humanoids, drones, warehouse robots, domestic robots, manipulation, navigation, dexterity, and real-world fleet deployment | AI is moving from screens into physical machines. |
| Other | Chips, accelerators, inference efficiency, data centers, cooling, power supply, grid deals, energy infrastructure, space or unconventional data centers, networking, memory, storage, safety/security tooling, standards, and other deployment bottlenecks | The enabling layer that determines whether AI becomes cheaper, faster, more reliable, and easier to deploy. |
Each lane has its own cumulative index, starting at 100 on the launch date. Lanes are not comparable to each other. Video at 187 versus Text at 162 means more logged movement since launch, not "video AI is more capable than text AI."
Demos and announcements score zero until they become real under that lane's rule.
For Text, Audio, Video, and Agents, "real" means the public can use the capability today, including paid access or open weights. Demos, papers, waitlists, private previews, and coming-soon posts score zero.
For Robots, the standard is physical-world availability rather than consumer availability. A robot event can score only when the product or capability is orderable, shipping, commercially deployed, in a real customer or partner pilot, or operating in a real-world field deployment. A keynote demo, lab video, staged prototype reveal, or paper-only result scores zero even if it looks impressive.
For Other, the standard is concrete deployment or capacity. An event can score if infrastructure is live, capacity is online, hardware is shipping, a model-serving efficiency improvement is available, a power or data-center deal is signed, construction has started, or a deployment-relevant tool or standard is released for real use. A concept, feasibility study, unsourced rumor, or general roadmap scores zero.
Items that matter but cannot score yet sit on the public "not yet" shelf at 0 points. If they later become usable, deployed, signed, online, shipping, or otherwise concrete, the later event can score.
Every qualifying event is classified into one tier. Points are fixed. The tier decides the score, never vibes.
| Tier | Points | Definition | Test |
|---|---|---|---|
| Milestone | +10 | A genuinely new capability, deployment, or enabling capacity that is usable, deployed, or concrete now | "Did this create a new kind of thing users, operators, customers, or the AI ecosystem can rely on this week?" |
| Notable | +5 | A clear, verified improvement to an existing capability: major quality/reliability jump, frontier-class model release, materially better robot deployment, meaningful infrastructure/cost/power/capacity progress, or tracked public benchmark movement tied to public access | "Is the same task or bottleneck now meaningfully better for real use?" |
| Minor | +2 | Access expansion, significant price cut, open-weights release of near-frontier ability, existing capability reaching a major new platform/language/customer segment, smaller deployed robot progress, or smaller infrastructure efficiency/access improvement | "Can meaningfully more people or operators use what already existed, or use it more cheaply?" |
| Regression | -3 | A shipped capability, deployment, infrastructure path, or access path is withdrawn, materially degraded, delayed, or shown to fail in ways that change what users can rely on | Requires primary evidence |
Big weeks should look big, but a coordinated PR blitz should not distort the index.
The week's movement per lane is described in plain words:
| Delta | Band |
|---|---|
| 0-2 | Quiet week |
| 3-5 | Incremental |
| 6-11 | Meaningful movement |
| 12-19 | Major leap |
| 20 (cap) | Breakthrough week |
An event only scores if it is backed by a primary source: an official announcement, release notes, model card, pricing page, tracked public leaderboard, customer deployment post, signed-deal announcement, product page, regulatory filing, or similar direct evidence.
Items surfaced only by newsletters, aggregators, or Gmail are logged as unverified at 0 points unless they include a primary link. If a primary source is then found, the entry can be promoted by editing its tier and points. The promotion is visible in git history.
The newsletter includes a Sources by segment list at the end of every issue. Claims without source support should not appear.
Funding rounds, hiring news, executive predictions, hype cycles, vague benchmark claims with no public access, "AGI is near/far" discourse, and general roadmap promises. None of that changes what can be used, deployed, or relied on today.
The ledger is append-only in spirit: if an entry was wrong, a correction entry is added that references the original id and reverses its points, with a note. The git history of data/ledger.json is the permanent audit trail.
A model drafts; a human ships. The automation gathers, classifies against this rubric, drafts the issue, and schedules the Kit broadcast. The operator keeps a veto window before send time and can delete the scheduled broadcast in Kit if anything looks wrong.