How Sygnls produces a signal

Three stores inform the site, and nothing else is crawled, purchased or licensed:

• retailnews.ai/rss.xml — articles, polled on a 30-minute schedule and stored.
• theretailpodcast.podbean.com/feed.xml — episodes, polled on the same schedule and stored.
• The RetailNews.ai article database, read directly at question time by Ask Sygnls. It is not stored. When it cannot be read, Ask says so and answers from the stored index instead.

Both feeds are published within the same owner group as Sygnls, so items from them are not independent of each other. The sources page shows, per store, when a read was last attempted, when one last succeeded, the newest publication held, and the current status.

Flagged for review There is no editorial exclusion list, no paywall handling and no source-quality tiering. Coverage is two feeds plus one direct store — it is not a market census, and absence of coverage is not evidence of absence.

One policy layer governs every surface: signals, themes, outcomes and Ask. Policy version 2026-09-10. Conservative policy default, pending calibration. Not a validated statistical threshold.

• 5 — Previous (comparison) window count required before any percentage change is shown.
• 10 — Qualifying tagged items across the comparison window required before any maturity stage, momentum ranking or coverage share.
• 10 — Items in a single theme's current window required before that theme carries a maturity label.
• 20 — Independently scored outcomes required before an outcome performance rate; scored independently of theme and momentum gates.
• 3 — Cited evidence items required before Sygnls offers review prompts.

A count meets a floor when it is equal to or above it. A count that is missing, negative, fractional, larger than its own denominator or otherwise not a whole number is treated as unusable rather than small, and the figure is suppressed. Below a floor, exact raw counts and evidence links are still shown; ranking, stage labels and percentages are not.

Zero, missing and not-yet-processed are distinct states and are always labelled differently.

• Five Signals: the last 7 days compared with the 7 days before. If the current week holds no tagged items, the page steps back a week at a time, up to 8 weeks, and labels the week it is actually showing.
• Theme trends: a rolling 84-day window, bucketed by week.
• Outcomes: mentions in the 28 days before a dated event, compared with the 28 days before that.

Ingestion, classification and brief generation are separate jobs with separate schedules. A recent successful feed read does not mean classification has caught up, so those timestamps are reported separately rather than merged into one “last updated”.

A fixed list of 15 themes is the only vocabulary a classifier may use: Agentic commerce, AI in operations, Store experience & design, Retail media & store-as-media, Supply chain & fulfilment, Quick commerce & delivery, Loyalty & personalisation, Payments & checkout, Shrink & loss prevention, Resale & circularity, Social & creator commerce, Brand & marketing, Workforce & frontline, Consumer demand & value, Leadership & strategy.

Taxonomy version 1.0, effective 2026-09-10, applying prospectively only. Items classified before that date were tagged under an unversioned taxonomy and have not been reclassified. Sygnls does not backfill a version onto historical tags.

Corrections: only entities have a human review and correction path. Entity merges and review decisions are recorded in an audit trail with the previous and new status, and original mentions and links are preserved rather than deleted. Theme tags have no review queue — a mis-tagged item is corrected only by reclassification.

Maturity stages (Emerging, Peaking, Mainstream, Declining) are rule-based labels derived from volume in the current window against the prior window, the theme’s share of coverage, and the mix of voices. They are descriptive summaries of stored coverage, not forecasts, and they are suppressed entirely when the window or the theme sits below its floor.

Divergence compares how operator voices and trade-press voices talk about a theme; vendor and promotional voices are excluded from it. It is reported as Aligned, Tension or Contradiction, and as “Insufficient data” unless both camps hold at least 2 stanced items.

Outcome lead labels (Quiet, Flat, Lead, Strong lead) describe whether mentions rose before a dated public event. They are descriptive only. Nothing on this site is back-tested or significance-tested, and no label predicts anything.

Flagged for review The cut-offs behind these labels are hand-set and have not been validated against outcome data. Voice credibility weights exist in the code but are not applied in weekly signals, so weekly signals are not voice-weighted.

Ask Sygnls shows a rules-based quality label (High, Medium, Low), never a probability or confidence percentage. The label starts from how many distinct records were actually cited and how many independent origin groups they belong to, and is then only ever lowered by:

• records whose origin cannot be established;
• records that contributed no exact quoted passage;
• records that could not be quoted, where fewer were quoted than went unquoted;
• polling cadence: how often Sygnls reads a store, taken from the source register. Where a cited store is read live rather than polled, the cadence is unknown and the label is capped at Medium.
• publication cadence: how often a source itself publishes. Sygnls has no established figure for any current source, so the age of the newest cited record is reported and not scored, and that missing expectation caps the label at Medium. Where an expectation is established, evidence older than twice that interval lowers the label.

Sygnls does not assess recency “relative to the question” and does not assume a daily or weekly publication rhythm for any source.

Repeats of the same story, and publications inside the same owner group, are collapsed into one group — they cannot inflate independence.

Factual statements in an answer are extractive: the text shown is the exact stored passage from a cited record, with its record identifier, and nothing is presented as a fact unless that passage exists. Paraphrases that cannot be matched to stored text are dropped rather than shown with an evidence badge.

Stored passages are excerpts, not full articles. Where an excerpt is cut short, the answer says so and gives the length held. Follow the citation for the complete text.

Suggested next steps are review prompts drawn from fixed templates and tied to the cited evidence. They contain no model-written factual claims. No recommendation is offered below the evidence floor above.

When the direct RetailNews.ai store cannot be read, the answer states that explicitly and is drawn from the stored Sygnls index; if neither is available, no answer is produced.

Model-assisted, with no human review step: theme classification, stance, speaker type, entity extraction, outcome extraction and Ask answers.

Rule-based and deterministic: freshness, windows, stage assignment, divergence arithmetic, lead scoring, alias resolution and floor suppression.

Human-reviewed: entities routed to the review queue, and editorial featuring. Nothing else here has been read by a person before publication.

• Two polled feeds and one direct store, all within one owner group.
• Classification runs on titles and summaries, not full article or transcript text.
• Item counts are small; most themes sit below the sample floor most weeks.
• No geography or retail-segment field exists, so those filters are not offered.
• Voice weights are defined but not applied in weekly signals.
• Alias resolution is rule-based and hand-maintained; unlisted spellings stay separate until merged.
• No taxonomy history before 2026-09-10; versioning is prospective.