The AI Signals Taxonomy

Definitions, base weights, source criteria, and worked examples for the 14 signal types that feed every Steek trend.

2025-03-29 12 min read

A signal-first index is only as good as the type vocabulary used to label its signals. This page documents the full Steek taxonomy: 14 types organized into three tiers — capability, commercial, and structural — with base weights, source criteria, and worked examples for each.

The taxonomy is intentionally narrow. Forcing every observation into one of 14 buckets is a feature, not a limitation: it is what makes the resulting trend index auditable.

Why three tiers

Capability signals tell you whether AI is getting more powerful. Commercial signals tell you whether anyone is paying for it. Structural signals tell you whether the conditions for both will continue to hold. A trend supported by all three tiers is much more durable than a trend supported by only one.

The velocity model uses tier diversity as a quality check before promoting a candidate to a tracked trend.

Capability signals

TypeBase weightDefinition
Release0.70A model, product, or feature shipped publicly by a credible vendor.
Benchmark0.60A measured score on a recognized public evaluation.
Research0.40A peer-reviewed paper, technical report, or arXiv preprint from a credible source.
  • Release — example: Anthropic releases Claude 3.7 Sonnet on its API and product surfaces.
  • Benchmark — example: New model reaches 85.6% on SWE-bench Verified, +6 points over prior SOTA.
  • Research — example: "Scaling Inference Compute" published with reproducible code on GitHub.

Commercial signals

TypeBase weightDefinition
Pricing0.90A change in published API or subscription pricing.
Enterprise0.85A procurement, partnership, or earnings disclosure indicating production deployment.
Distribution0.70A platform shift that exposes models to a new surface.
Hiring0.50A senior hire or org-design change at a frontier lab or major buyer.
  • Pricing — example: Provider X cuts input-token cost from $5 to $2 per million.
  • Enterprise — example: Fortune 100 names $50M AI line item in Q3 10-Q.
  • Distribution — example: Browser ships native AI runtime accessible from JavaScript.
  • Hiring — example: Major SaaS vendor names new "Head of AI Agents" reporting to CEO.

Structural signals

TypeBase weightDefinition
Policy0.95A regulatory document, executive order, law, or formal voluntary commitment.
Hardware0.80A new accelerator, datacenter, or interconnect with material throughput.
Open-source0.65A new permissively licensed model, dataset, or tool.
Standards0.60A protocol, format, or interoperability spec gaining cross-vendor adoption.
Demo0.45A credible demonstration of a previously unverified capability.
Lawsuit0.70A filing or ruling materially affecting AI rights, data, or liability.
Security0.55A discovered vulnerability, jailbreak class, or material incident.
  • Policy — example: EU GPAI obligations enter application phase.
  • Hardware — example: NVIDIA Blackwell shipments commence; per-system spec disclosed.
  • Open-source — example: Frontier-quality open-weight reasoning model released under MIT-style license.
  • Standards — example: MCP adopted by major IDE; cross-vendor support announced.
  • Demo — example: Computer-use agent completes 30-step office task on independently verified video.
  • Lawsuit — example: Major copyright suit filed against frontier lab over training data.
  • Security — example: New prompt-injection class disclosed for browsing agents; CVE issued.

Source criteria

For every type, Steek requires:

  1. A primary-source URL. A vendor release page, a paper PDF, an SEC filing, a regulator document, a pricing page snapshot.
  2. An observation timestamp. Either explicit in the source or imputed from first observation.
  3. Named entities. At least one organization, model, or product, normalized against the Steek entity table.

An event missing any of the three is rejected at ingestion. The pipeline is documented in the AI Signals Report.

How weights are recalibrated

Quarterly, Steek measures the predictive power of each signal type for trend persistence: given a trend that was tracked for at least 12 months, which signal types appearing in its first 30 days were most correlated with persistence beyond month 12? Weights drift slowly toward the empirical answer. The recalibration log is public.

What the taxonomy does not contain

  • Sentiment. Out of scope. Steek measures whether something happened, not how it felt.
  • Predictions. Out of scope at the signal layer. Predictions live on trend pages and in briefings, not in the typed record.
  • Speculation. Tweets, rumors, and unattributed claims do not pass source criteria.

Reading the taxonomy in context

The taxonomy is most useful read alongside the velocity model (which describes how weights aggregate) and a worked example like The 2025 Landscape (which shows the full pipeline applied to a calendar year). For the live output, visit the trends index.

Related Reading