How does The Frontier Signals verify its evidence?

What does The Frontier Signals guarantee?

1. Every “last verified” date is a recorded check. A date moves only when its source was fetched in that research run; the address, response and time are logged. If a source cannot be reached, the earlier date stays. See the latest run’s receipt.
2. Published editions never change. Editions, source registers and data downloads are permanent; corrections appear as new editions, labelled as corrections. Browse the edition archive.
3. What you see is exactly what was checked. After every publish, each page and file on the live site is compared byte-for-byte with the build that passed validation. See publish history.
4. Verdicts need a recorded editor decision. Bubble Watch statuses are calculated mechanically; a change to the published verdict, or a new draft rule, goes live only with an editor’s recorded approval and rationale. See the rules and history.
5. Missing stays missing. Undisclosed values are shown as unknown, never estimated or filled with zero.

How do we count human and AI-agent readers?

The banner at the top of each page estimates who read this site over the last seven days. It counts page views on thefrontiersignals.com and sorts each one by the identity its software declares (the user-agent).

• AI agents are matched to the company they name, such as GPTBot or ChatGPT-User (OpenAI), ClaudeBot (Anthropic), PerplexityBot, Bytespider (ByteDance) or CCBot (Common Crawl). Where a company publishes the IP ranges its crawlers use — currently OpenAI, Perplexity, Google, Microsoft, Apple and Common Crawl — we check the request against them and count it as verified. Others, including Anthropic, ByteDance, Meta and Amazon, can only be counted as self-declared; the banner says which.
• Bots can claim any identity. Software can pretend to be a browser or a well-known crawler. “Human” means a page view from software presenting itself as a standard browser and not identified as automated, so it is an estimate, not a head count.
• Google-Extended is not counted separately. It is a permission in robots.txt, not a crawler, so Google’s AI use cannot be told apart from Googlebot; both appear as Google.
• No precision we cannot defend. Shares are rounded to 5% and always shown as approximate; companies under 2% are grouped; no estimate is shown until at least 300 page views fall in the window.
• No tracking. We store no IP addresses, user-agents or anything about individual visitors — only hourly counts per category. Our own publishing checks are excluded.

The rounded figures behind the banner are open: /api/readers.

What do the four provenance labels mean?

Independent evidence

An outside evaluation or research synthesis. Independence does not remove limits in its sample, method or conditions.

Company claim

A vendor’s own announcement, performance claim or financial disclosure. Audited results are still company-origin information, not proof of customer profitability.

Reported claim

Attributed reporting, survey results or allegations. Read the original report and its limits.

Documentation

A filing, method, access guide or release record. Availability is not performance proof.

How do source links and evidence grades work?

Deal grades are confirmed when a supporting company disclosure or filing establishes the recorded terms; reported when only attributed press reporting supports them. Link labels identify the destination publisher. We keep measurement periods separate from actual source-verification dates. Undisclosed values stay unknown, never estimated.

How do tripwire rules and verdict approvals work?

Published tripwires have versioned thresholds and mechanical statuses. Rules v3 retains vacancy and preleasing, operating cash coverage and financing-failure transmission, and adds NVIDIA collection pressure and IPO aftermarket drawdown. The tracker displays the exact warning/trigger and a rationale for each. Supplier-investor commitments remain context, not a bubble score. Cloud capex and GPU rent tests stay excluded until comparable observations are available. Thresholds are editorial early warnings, not empirically calibrated crash predictors. Missing observations cannot pass. Draft rules are excluded from public assets and counts. A changed verdict requires an explicit editor approval and rationale tied to the exact evidence snapshot; no probability or predicted crash date is calculated.

Read the published rule versions and history →

What counts as a money loop?

A direct loop pairs a supplier’s investment or warrant with the same customer’s compute commitment. Three-company payment paths appear separately; they do not trace an investment dollar into an invoice. Headlines count direct loops only. Featured loops prefer both sides confirmed, then both values disclosed, then the latest verification.

How is circular financing measured?

The headline divides disclosed commitments to supplier-investors by all disclosed tracked lab compute commitments, with confirmed and reported amounts shown separately. Historical and overlapping records are excluded. Mixed periods, contingent ceilings, lower bounds and undisclosed terms limit interpretation; these are not annual spending or paid cash.

Confirmed-only circular share: 91.8%; supplier concentration: 91.8%. Coverage: 5 of 12, low coverage. This ratio rises mainly because Oracle’s reported $300B non-circular commitment leaves the denominator. It is a different subset, not a stronger estimate of industry-wide circularity.

Inspect the source ledger → · Download shared deal records →

How are corrections handled?

Corrections appear in new editions. Archived editions, source registers and download snapshots are immutable. Builds check references, financial consistency, missing values and mechanical rules; these checks do not establish scientific truth. Connor Meeds controls publication. A personal record review is claimed only when a sign-off exists.

When does research update?

Research is scheduled twice daily at 8 AM and 4 PM America/New_York. Last verified dates advance only when their sources are actually checked. A rebuild does not create a research edition or reset a measurement date.

How does Agent watch classify behavior?

The ladder grades demonstrated behavior. Controlled-test or fictional scaffolding is disclosed separately, not graded down. Human misuse does not establish independent agent deviation.

0. Level 0 — Misuse: A human directed harmful use.
1. Level 1 — Error: An error while trying to follow the task.
2. Level 2 — Overstepped: Actions outside instructions or permissions.
3. Level 3 — Concealed: Concealment, misreporting or intentional underperformance.
4. Level 4 — Resisted oversight: Agent-originated avoidance of shutdown, monitoring, correction or unauthorized resource acquisition.
5. Level 5 — Coordinated or copied beyond deployment: Copying or coordination beyond deployment boundaries; the Hugging Face record documents coordination, not self-replication.

Actor, setting and severity

Actor: agent deviation, human misuse, or deliberately elicited behavior. Setting: controlled evaluation, red-team test, production, or consumer use. A test reaching live systems remains a test-origin record with external impact explicitly noted.

Contained: blocked or simulated with no verified external harm. Limited: bounded disruption. Significant: documented material loss, exposure or compromise. Severe: documented grave or broad consequences, never hypothetical potential. Reversibility can be unknown; remediating a compromise does not establish every downstream effect was undone.

Approval and counting

One connected event or defined evaluation finding is one record, not one prompt, run or victim. Classifications require a recorded editor sign-off bound to the evidence. Level 3+ needs a primary source or two sources. A changed highest level requires a separate editor decision. Builds stop on stale approvals; absent records mean not documented here, not safe.

Disclosure criteria

Criteria checks, not grades or rankings. Yes means evidence found for the stated scope; Partial means limited or related evidence; Not found means no qualifying evidence in the cited search scope as of the checked date, never No.

• Publishes agentic safety evaluation results in model or system cards: Published results must concern tool use, cyber autonomy, prompt injection, or agent behavior; general capability benchmarks alone do not qualify.
• Allows independent pre-deployment testing by a named evaluator: A named assessor and evidence of early model access are required. Post-release testing and unnamed red teams do not establish this.
• Offers a public incident or vulnerability reporting channel: A documented official reporting route; availability does not measure responsiveness.
• Runs a bounty covering agent misuse or prompt injection: Scope must cover relevant security or safety impacts; a general bounty does not imply all model behavior is eligible.
• Publishes agent-incident post-mortems: At least one public technical retrospective; not a claim that all incidents are disclosed.

Public corrections → · Shared record dataset →