Methodology

How the engine works

The innovation detection engine runs once daily, searching global news, research and industry sources with neural search to detect new innovations and fresh signals on the ones already tracked. Each candidate article is then analyzed by an AI model to produce a structured assessment. Findings are scored and classified by industry before publication, and tagged when regulation drives them or when they are still a research result.

The three indicators

The five-layer quality gate

1. 1. Novelty and emergence validation

   The model must explicitly determine whether a candidate is genuinely emerging. Established markets, minor product refreshes, seasonal sales and clickbait are rejected before publication. This prevents routine noise from entering the innovation feed.

2. 2. Minimum score threshold

   Findings below the minimum disruption threshold are discarded. The current save threshold is 4, so candidates scoring under 4 are filtered out even if they are novel. This keeps the published set focused on meaningful signal strength.

3. 3. Source quality filtering

   Candidate selection prioritizes higher-authority sources, and lower-authority domains face a stricter save bar. In practice, those sources are only accepted if the assessed disruption score is very high. This mechanism reduces promotional or low-trust content without publishing internal source lists.

4. 4. Duplicate detection

   Candidate URLs already stored as findings are skipped. This prevents duplicate entries and keeps each item in the feed linked to a distinct source article.

5. 5. Strict scoring rules

   Vague, promotional and listicle-style content is penalized, and low-substance articles are capped at a score of 5. Scores of 7 or higher require genuine technological breakthroughs, new market categories, or fundamental shifts backed by concrete evidence. These constraints are designed to keep high scores rare and defensible.

What gets rejected

Most candidate articles do not become published findings. The feed is intentionally curated: items are filtered out when they fail emergence checks, score thresholds, source-quality requirements, or duplicate checks.

Source traceability

Every published finding links directly to its original source URL. This allows readers to verify claims and inspect the underlying material that informed the score.

Source policy and copyright

We report on what sources say; we do not republish what they wrote. Four rules govern how source material is handled, and each one is enforced in the pipeline rather than left to editorial discretion.

• Titles and summaries are ours. Every headline and summary on the site is an original description of the underlying facts, written to state what happened and what was measured. We do not use a publisher's headline as our own, and we do not paraphrase a source sentence by sentence. Facts and items of news are not protected by copyright; a source's particular expression is, and we leave it where we found it.
• Source extracts are bounded by licence. Each source is classified at the point of collection as US federal public domain, open access, press release, press publication, or unclassified. How much source-derived text we may display scales with that classification. Press publications and unclassified sources are held to a very short extract; the classification is stored alongside the finding so any decision can be audited later.
• Mining reservations are honoured. Before a source is analysed we check for a machine-readable reservation of text-and-data-mining rights, using the W3C TDM Reservation Protocol and named AI-agent directives in robots.txt. Sources that reserve those rights are dropped before analysis, not merely excluded from display, and every check is logged.
• Attribution is always present. Each finding names its source and links to the original. We benefit publishers by sending readers to them.

Source classification is an automated, best-effort signal, not a legal determination. If you are a rights holder and want material removed or your domain excluded from the engine, write to info@signalsofinnovation.com and we will act on it — no formal notice is required. The full procedure is in section 9 of our Terms.