Documentation
¶
Overview ¶
Package scoring owns the article importance model: the rubric dimensions an LLM rates, the deterministic aggregation of those dimensions into a 0-100 score, and the tier thresholds used to bucket articles in digests.
The LLM no longer chooses the final score directly. It rates a handful of narrow, anchored sub-dimensions (0-4 each) and Compute combines them with tunable weights here. Because the raw dimensions are persisted alongside the computed score, the weights can be retuned later and scores recomputed in a batch without re-running the LLM.
Index ¶
Constants ¶
const ( // AggregatorScore is the fixed score forced for aggregator/roundup articles, // preserving the previous prompt's "always set the score to exactly 40" rule. AggregatorScore = 40 // EvergreenCap caps the score of pure-evergreen articles (Specificity == 0), // preserving the previous prompt's "generic/evergreen must score ≤60" rule. EvergreenCap = 60 // PromoCap caps the score of promotional articles (announcements, marketing, // commercials) at the top of the "May Read" tier, so an ad can never be ranked // as "Should Read" or "Must Read" however well it scores on other dimensions. PromoCap = TierShouldRead - 1 )
const ( TierMustRead = 90 TierShouldRead = 75 TierMayRead = 60 )
Read-tier thresholds (inclusive lower bounds) on the 0-100 score.
Variables ¶
var Weights = struct { Specificity float64 Severity float64 Breadth float64 Novelty float64 Actionability float64 Credibility float64 }{ Specificity: 0.20, Severity: 0.25, Breadth: 0.20, Novelty: 0.10, Actionability: 0.15, Credibility: 0.10, }
Per-dimension weights, summing to 1.0. This is the single place to retune the relative influence of each dimension on the final score.
Functions ¶
func Compute ¶
func Compute(d Dimensions) int
Compute aggregates rubric dimensions into a 0-100 importance score using the default global model. See Config.Compute.
func PriorityKey ¶
PriorityKey returns the short manifest bucket key for a 0-100 score using the default global thresholds. See Config.PriorityKey.
Types ¶
type Config ¶ added in v0.4.0
type Config struct {
Weights DimensionWeights
AggregatorScore int
EvergreenCap int
PromoCap int
TierMust int
TierShould int
TierMay int
}
Config is a self-contained, retunable instance of the importance model: the dimension weights, the aggregator/evergreen overrides, and the read-tier thresholds. Profiles each resolve their own Config; callers that want the historical global behaviour use DefaultConfig (and the package-level Compute / ReadTier / PriorityKey wrappers, which delegate to it).
func DefaultConfig ¶ added in v0.4.0
func DefaultConfig() Config
DefaultConfig returns the historical global model, sourced from the package Weights var and the Aggregator/Evergreen/Tier constants so there is a single source of truth for the defaults.
func (Config) Compute ¶ added in v0.4.0
func (c Config) Compute(d Dimensions) int
Compute aggregates rubric dimensions into a 0-100 importance score.
Each dimension is normalised to 0-1, combined via c.Weights into a weighted average, and scaled to 0-100. Overrides are then applied: aggregator articles are forced to c.AggregatorScore, promotional articles are capped at c.PromoCap, and pure-evergreen articles (Specificity == 0) are capped at c.EvergreenCap.
func (Config) PriorityKey ¶ added in v0.4.0
PriorityKey returns the short bucket key for a 0-100 score, used for digest manifest priority tallies. Unlike ReadTier it has no "unscored" bucket; anything below TierMay falls into "opt".
type DimensionWeights ¶ added in v0.4.0
type DimensionWeights struct {
Specificity float64
Severity float64
Breadth float64
Novelty float64
Actionability float64
Credibility float64
}
DimensionWeights holds the per-dimension weights used by Config.Compute. It is the named-type form of the package-level Weights var, so a profile can carry its own weighting without touching package state.
type Dimensions ¶
type Dimensions struct {
// Specificity: generic/evergreen concept (0) → single concrete, recent event (4).
Specificity int `json:"specificity"`
// Severity: informational (0) → active exploitation / critical patch / major breach (4).
Severity int `json:"severity"`
// Breadth: niche product (0) → ubiquitous software or whole sector affected (4).
Breadth int `json:"breadth"`
// Novelty: rehash of known facts (0) → genuinely new disclosure/finding (4).
Novelty int `json:"novelty"`
// Actionability: nothing to do (0) → clear defensive action, patch, IOCs, detection (4).
Actionability int `json:"actionability"`
// Credibility: unsourced blogspam (0) → primary source / vendor advisory / named researcher (4).
Credibility int `json:"credibility"`
// IsAggregator marks roundups / weekly recaps / link digests, which are forced
// to AggregatorScore regardless of the other dimensions.
IsAggregator bool `json:"is_aggregator"`
// IsPromotional marks product announcements, marketing pieces, press releases,
// sponsored content, and vendor commercials, which are capped at PromoCap (the
// top of the "May Read" tier) regardless of the other dimensions.
IsPromotional bool `json:"is_promotional"`
}
Dimensions are the rubric sub-scores rated by the LLM. Each numeric field is on a 0-4 scale; values outside that range are clamped by Compute.
Source Files