AI progress, only when it's real.
Mode:ยทThe Threshold Report - methodology
This page describes how The Threshold Report's instruments are scored (the lane records, the risk and readiness dial, and the 'claims' board).
The whole system in one paragraph
The Threshold Report tracks the progress of AI capability, and the ramifications of that progress. Every lane (category) we cover has a running total that starts at 100, because a lot(!) had happened before we built this, and moves only when something became real: usable, deployed, signed, approved, or released. Fixed rules assign every point (see below). Announcements score zero until they become real, and items that matter but are not real yet sit below "the threshold" as 'Upcoming' until they are.
The rules
What counts
An item scores when it became real - usable, deployed, signed, approved, or released - at the level its lane demands, rather than demoed or promised. An item that is announced but not yet real or available sits in the public Upcoming section with what it is waiting on stated, and enters the record only when that happens.
The lanes
Eight running totals: Text, Audio, Video, Image, Agents, Robots, Science \& Augmentation, and Compute \& Infra (chips and accelerators, data centers and power, memory and storage, networking, and other enabling infrastructure). Each starts at 100 on its start date and is cumulative since day one based on news in the 'New' section. Totals are not comparable to each other: a higher number means more logged movement since its start, not a more capable field.
How much one event moves its lane:
| Event | Lane effect |
|---|---|
| Milestone - a capability crosses a line for the first time | +10 |
| Notable - a clear improvement to an existing capability | +5 |
| Minor - access expands, prices fall meaningfully, or a smaller step becomes real | +2 |
| Regression - something real is withdrawn or degraded | takes back the points it earned (-3 when it predates the record) |
| Restoration - a logged regression is reversed | returns the reversed points |
| Not scored - real news with no capability change: incidents, certifications, partnerships, policy context | 0, reported with the reason stated |
A release that improves more than one lane scores under every lane it improves - one story can move Text, Audio, and Video at once, each under its own event size.
Per issue window a lane can log at most one milestone, two notables, and three minors, and can gain at most 20 points net.
The claims board
The board holds plain-language claims and answers each with one of three statuses.
- No: no public system does it reliably.
- Sort of: demos, narrow cases, or heavy supervision.
- Yes: ordinary users get it reliably from shipped products.
Answers change only on sustained, real availability. A newly added claim gets its first answer on the next daily update, recorded as a starting point rather than a flip.
The risk and readiness dial
The dial runs from 'Under Control' to 'Chaos' and moves based on the 'Security' section news alone: every item reported there is designated Positive or Negative, with the reason stated on the item. Exploits, misuse, and weakened oversight push the dial toward Chaos; real safety wins (collaboration that actually happened, rules in force, foiled attacks, deployed defenses, alignment and interpretability progress) pull it toward Under Control. The dial is a cumulative record, meaning today's reading carries everything that came before it.
Words alone never move it in either direction: an open letter, a pledge, or a call for a pause is reported as real news and moves nothing until someone acts on it - a rule takes effect, a lab verifiably changes behavior, etc. Some action!
How far one item moves the needle:
| Impact | What it looks like | Needle movement |
|---|---|---|
| Minor | contained or single-organization impact: one exploited flaw, one deployed defense | 4 points |
| Notable | sector-level impact: real misuse at scale, a rule in force, real alignment progress | 8 points |
| Major | systemic impact: widespread exploitation in the wild, a landmark rule, a fundamental interpretability result | 15 points |
Chaos adds the points, Under Control subtracts them, and the needle is the running sum, capped at 100 each way. The tally of both sides is shown under the dial.
Independence
No scored item is ever paid for. Paid listings on the board will carry a visible marker and a disclosure. This publication has no positions it is paid to defend.
Edge cases
An item announced first but becomes real later enters the record on the day it became real, not the announcement date. A capability that regresses can lose points under the same bar that granted them.
When a story genuinely falls between two scoring tiers, it receives the lower score. For example, if something could reasonably be classified as a Milestone (+10) or merely Notable (+5), it gets +5. A frontier lab releasing a meaningfully better model is a good example: does it introduce a genuinely new capability, or is it simply a clear improvement to an existing one? If the answer is debatable, the lower score wins. Our restraint is deliberate. Our index is designed with an anti-hype bias: when the evidence does not clearly justify the higher score, we under-score. We want to ensure the progress slope for each lane is trustworthy.
Thank you for going through this, please support us by visiting regularly, subscribing, and following us on social media. We love hearing from the community, and look forward to reading your responses & comments!