Scoring and Ranking

RSSMonster keeps several scoring concepts separate so each one has a clear meaning. A score may describe one article, a source’s recent history, personal relevance, or the importance of a current story. No single score decides whether an article is allowed to exist or be read.

Concept	Meaning	Use
Overall article quality (`articleQuality`)	Article content quality, independent of source trust and personal interest.	Overall quality badge and article-quality thresholds, including `quality:` expressions.
FeedTrust	Recent source value in the user’s account.	Source contribution to Quality & source trust ranking.
Personal interest (`interestScore`)	Signed relevance to the user.	Personalization contribution to Recommended.
Semantic affinity (`semanticAffinityScore`)	Continuous confidence-adjusted affinity to the strongest positive Interest Island.	Independent stored signal; Recommended weight 0.10 by default.

Quality & source trust ranking combines 70% articleQuality with 30% FeedTrust. Recommended has no direct Quality contribution by default; quality sorting and eligibility remain separate. These ranking combinations do not change Overall quality badges or Article-only quality thresholds. A high-quality article may come from an inconsistent source or cover a subject the user dislikes.

Core Signals

Overall article quality

Overall article quality asks: How good is this particular article?

It is a normalized 0.0–1.0 content score derived from the article’s stored quality, tone, and advertising scores. Feed reputation is not inserted into the article-level value.

articleQuality =
    0.50 × qualityScore
  + 0.25 × toneScore
  + 0.25 × advertisementScore

The persisted inputs use 0–100; the result is clamped to that range and divided by 100 once. A missing component uses the neutral-good value 70. Higher advertisementScore means less promotional content, so all three inputs have the same higher-is-better direction.

The article badge, Settings → Scores → Overall quality, and Smart Folder Overall quality control show this score out of 100. The badge preserves fractional points so that an article shown as 74.75 does not appear to meet a minimum of 75. Badges, accessible labels, breakdown footers, and quality-filter preview chips retain fractional points and omit unnecessary trailing zeros. Badge colors use the unrounded score, including near the 60 and 80 severity boundaries. Smart Folder expressions retain their normalized syntax: quality:>=0.75 means at least 75/100. Settings stores the independent minOverallQualityScore threshold as an integer from 0 to 100, with 0 disabling that gate. Existing writing, tone, and ad-free thresholds remain separate additional filters under Advanced. Existing preferences are preserved by the migration; the new overall threshold starts at 0. Article scoring uses the configured inference service, with local models or a compatible remote provider. OpenAI is not required. Classification supplies the scores; generation supplies summaries and tags in the combined Article analysis. Administrators can inspect availability in Settings → AI / Inference.

New articles start with fallback component scores of 70, subject to Action-owned overrides. These defaults are not an AI assessment. When analysis is skipped, retained scores remain subject to thresholds. Missing/nonfinite components also use 70 when computing article quality. An all-70 component set does not show an Overall quality badge, regardless of analysis status.

When analysis is pending, processing or failed, inferred scores and tags are hidden. These states are exempt from thresholds on inference-owned components. Action-owned Writing quality and Ad-free content overrides remain subject to their respective thresholds; an override of either also removes the Overall quality exemption. Other visibility rules still apply. This analysis-state policy is separate from numeric fallback values.

The existing source-aware ranking remains separate: its menu label is Quality & source trust. Top Stories retains its Quality weight; Recommended has no direct Quality contribution by default. “High overall quality” uses the article-only score, never the Recommended score.

FeedTrust

FeedTrust asks: How consistently valuable has this source been as a source of articles?

It summarizes 30 days of article quality, supporting engagement, deterministic originality, and explicit negative feedback. It is user-specific and remains separate from factual verification, crawl health, and personal interest. See FeedTrust for the complete conceptual model.

Freshness

Freshness is a time-decay value derived from publication time:

freshness = exp(-ageInHours / 48)

A missing publication time produces zero. Intelligent ranking clamps the value to 0–1, which also prevents a future publication timestamp from producing an out-of-range contribution.

Personal interest

Interest is a bounded signed contribution from trusted personal evidence. Island paths separate preference strength, Island confidence and relationship confidence. Direct vectors determine Island relationships; explicit behavioral fallback preserves unrepresented likes/saved/dislikes with recency and intent attenuation. The strongest positive and strongest negative contributions are combined without blindly summing correlated paths. Positive interest promotes relevant material; negative interest penalizes it. FeedTrust is separate.

The internal interest-scoring reference is authoritative for those formulas. Weak relationships and singleton support have less authority; missing evidence means exactly neutral interest. Unread or missing engagement is not automatically a negative signal.

Event evidence

Events contribute coverage, publisher diversity, and corroboration signals. These describe how strongly a current occurrence is supported across articles and sources. Event co-coverage is not FeedTrust duplication evidence.

coverage        = clamp01(log2(max(eventArticleCount, 1)) / 6)
sourceDiversity = clamp01(ln(sourceCount + 1) / 2.56)
sourceSpread    = clamp01(log2(max(sourceCount, 1)) / log2(8))
crossSource     = max(sourceDiversity, sourceSpread)
corroboration   = coverage × crossSource

Counts are normalized to nonnegative finite values; clamp01 maps nonfinite values to zero and bounds finite values to [0,1]. These primitives are shared by Recommended and Top Stories. An article without an Event receives zero for every Event-derived value.

Ranking Modes

Chronological

sort:desc orders newest first and sort:asc orders oldest first. These modes are available when users want a conventional RSS inbox without intelligent ranking. Prioritize trusted feeds in Unread is an optional override that can blend bounded FeedTrust into these chronological modes; when disabled, ordering is strictly by publication time and stable article ID. This preference does not change Recommended, Top Stories, or Quality & source trust ordering. It differs from Briefing’s Use Recommended ordering for summary, which uses Recommended scores to order summary stories and leaves article list ordering unchanged.

Quality & source trust

Quality & source trust (sort:quality) ranking combines the value of the individual article with the source’s recent record:

Quality = 0.70 × articleQuality + 0.30 × feedTrust

The formulas retain the internal symbol Quality for this source-aware score. The two inputs remain independently visible and conceptually distinct. Overall quality names the article-only score and its filters. Writing quality names only the qualityScore component. UI controls and badges use 0–100; search expressions use the normalized 0–1 score. For example, quality:>=0.75 requires Overall quality of at least 75/100, independently of source trust. Query preview chips label the filter Overall quality and the sort Quality & source trust; the quality and sort:quality identifiers remain unchanged.

Semantic affinity

semanticAffinityScore measures the strongest positive Island contribution using the quadratic contribution: q = clamp(positivePreferenceStrength × islandConfidence × max(0, cosineSimilarity)², 0, 1); semanticAffinityScore = round4(max(q1, ..., qN)). It is threshold-free and excludes Article-to-Article fallback. It uses every eligible positive Island. One positive Island can produce affinity. With no supported positive contribution, the evaluated result is zero. The nullable FLOAT starts null and is cleared by duplicate invalidation or semantic reset. All computed values are finite in [0,1]; they are confidence-adjusted signals, not probabilities. Affinity excludes freshness, Feed trust or Article quality.

Normal scoring, expiry refresh and explicit recalculation persist both scores atomically per update with their existing shared evaluation clock. An explicitly requested recommendations:recalculate -- --userId=1 --allHistory refreshes read, unread and saved history in bounded batches without regenerating embeddings or rebuilding Islands. This mode skips unsupported embeddings, leaving their existing scores untouched; a skipped affinity remains null rather than claiming a successful zero. Checkpoints support resumption with the same evidence and evaluation clock. See maintenance commands for scope and restart rules. There is no automatic startup backfill.

Recommended consumes the two stored personalized signals independently: interestScore supplies established signed preference and behavioral evidence, semanticAffinityScore supplies continuous strongest positive affinity. Their calculation and storage are unchanged. Affinity does not change Briefing eligibility, Smart Folder query semantics or learned-preference labels.

Recommended is personalized. Its bounded base score combines:

Recommended =
    0.60 × positiveInterest
  + 0.10 × semanticAffinityScore
  + 0.17 × freshness
  + 0.10 × corroboration
  - 0.30 × negativeInterest
  + ruleMatchBoost
  + hotBoost

First normalize finite interestScore to [-1, 1], or use zero when missing or nonfinite. Then positiveInterest = max(interestScore, 0), negativeInterest = max(-interestScore, 0), and a matching rule contributes 0.08 once regardless of how many rule tags match. The final result retains negative totals and is capped at 1. With these normalized inputs and weights, the range is -0.30–1. Null affinity contributes zero without changing the stored values; no normalization by Island count or total ranking weight occurs. Recommended sorting and API serialization preserve the sign; the presentation label uses the same signed score on a percentage scale, not a probability.

The authoritative implementation is recommendedScore.js. Semantic confidence modifies interest; it is not another final weight.

Every Article eligible under the caller’s ownership, visibility and explicit filters receives a finite Recommended score. No Event, Island, vector or nonzero interest is required. The regression target is 100% Recommended coverage, while personalization coverage may legitimately be sparse.

Input	Absent-value behavior
Authorized identity and view eligibility	Required by caller; never synthesized by ranking.
Interest	Optional; zero when missing or nonfinite.
Event/corroboration	Optional; zero without an Event.
Island, embedding	Optional; not Recommended prerequisites.
Semantic affinity	Optional persisted values; null/missing contributes zero.
Article quality components	No default ranking contribution; Quality metadata/rollback retain the existing fallback of 70.
FeedTrust	Unavailable trust (missing, null, or nonfinite) uses the neutral prior 0.75; finite estimated values, including zero, remain meaningful.
Freshness	Article model uses zero for missing publication time; a plain object without freshness uses 0.5.
Matching rule tags	Optional; no boost without a match.

Recommended is computed at runtime. The separate interest updater processes canonical, unfiltered, unread Articles; that update scope does not restrict runtime Recommended ranking to unread or Island-matched Articles.

Quality has no direct default Recommended weight, and FeedTrust adds no separate raw boost. This gives personalization, freshness and corroboration control of Recommended while retaining Quality & source trust as the dedicated quality sort. Quality metadata, query quality filters and configured minimum-score gates remain unchanged. Quality may still affect eligibility; it does not re-enter through freshness, corroboration, rule/Hot boosts or Feed balancing. Existing user Actions can still set rule/Hot status independently of the ranking formula. Event coverage, cross-source evidence, and event size are not separate Recommended terms; corroboration is its only Event-derived contribution.

Recommended configuration and rollback

Coefficients are read once at process startup from server environment variables. Each accepts a finite number in [0,1], including zero; missing, blank or invalid values fall back independently to the defaults below. Weights are not normalized. Restart the server and workers after changing them.

Variable	Default
`RECOMMENDED_INTEREST_WEIGHT`	0.60
`RECOMMENDED_AFFINITY_WEIGHT`	0.10
`RECOMMENDED_FRESHNESS_WEIGHT`	0.17
`RECOMMENDED_CORROBORATION_WEIGHT`	0.10
`RECOMMENDED_NEGATIVE_INTEREST_WEIGHT`	0.30
`RECOMMENDED_QUALITY_WEIGHT`	0.00

Rollback to the previous formula by setting RECOMMENDED_QUALITY_WEIGHT=0.13, RECOMMENDED_AFFINITY_WEIGHT=0, retaining the other defaults. This restores the existing Quality calculation (including FeedTrust) without changing any persisted scores or requiring a migration. Rule and Hot bonuses remain 0.08 and 0.07; the final upper cap remains 1 and negative totals remain signed. Configurable weights also govern API explanation contributions; zero-weight signals are omitted from promotion reasons.

Supported sorts

The visible toolbar order is Newest, Oldest, Top Stories, Recommended, Quality & source trust. These are the supported sorts for search and Smart Folder expressions.

Recommended, Top Stories, and Quality & source trust are computed across the complete eligible candidate set before a result limit is applied. Equal scores fall back to publishedAt DESC, then id DESC. These computed modes use a stable ordered ID collection for incremental article loading rather than database cursor pages.

Filters Versus Ranking

Ranking changes order. It does not make an otherwise ineligible article eligible, and a low score does not delete or unsubscribe a source. Smart Folder and search filters determine eligibility before runtime ranking is applied.

See the Search Guide for the supported sort expressions and Smart Folders for reusable filtered views.

The forward-only 20261021000000-remove-multi-island-score.mjs migration retires the experimental additional-Island coverage column. Its former 0.05 ranking weight is removed without redistribution. The two remaining stored scores are preserved; no history rescore or embedding regeneration is needed. Previously applied add-column migrations remain in the migration history.