Artist and title strings are full of regex metacharacters, so a new candidate could match an unrelated history entry and be silently skipped as already seen.
201 lines
8.8 KiB
Bash
Executable File
201 lines
8.8 KiB
Bash
Executable File
#!/bin/bash
|
|
# ==============================================================================================
|
|
# ================================= Decision Engine ============================================
|
|
# ==============================================================================================
|
|
#
|
|
# PURPOSE
|
|
# ─────────────────────────────────────────────────────────────────────────────
|
|
# Behavior-driven scoring kernel for media automation. Sourced by consumer
|
|
# scripts — not run directly. Evaluates candidates, applies weighted scoring,
|
|
# temporal decay, and deduplication, then returns a verdict.
|
|
#
|
|
# Does NOT download media, search indexers, or manage applications. Has no
|
|
# side effects — no file writes, no API calls, no deletions.
|
|
#
|
|
# Currently paired with: Arrs_Stack/playback_aware_lidarr_discovery.sh
|
|
#
|
|
# ==============================================================================================
|
|
# OPERATIONAL MODEL
|
|
# ==============================================================================================
|
|
#
|
|
# A library, not a program. Consumers source it and drive the pipeline themselves:
|
|
#
|
|
# 1. score_candidate() — sum the four weighted component scores
|
|
# 2. apply_temporal_decay() — reduce by one unit per 30 days of age, floored at 0
|
|
# 3. is_duplicate_candidate() — check the consumer's own history file
|
|
# 4. make_decision() — ACCEPT or REJECT against the consumer's threshold
|
|
#
|
|
# Every input is supplied by the caller and every output is returned to it. The engine holds
|
|
# no state between calls, reads no config, and never acts on its own verdict.
|
|
#
|
|
# ==============================================================================================
|
|
# DESIGN PRINCIPLES
|
|
# ==============================================================================================
|
|
#
|
|
# Domain-Agnostic Core
|
|
# The engine itself has no knowledge of music vs TV vs movies. Consumers
|
|
# define thresholds, weights, and strictness profiles — the engine just
|
|
# scores and decides. This prevents cross-domain bias pollution: music
|
|
# strictness cannot contaminate TV intake logic.
|
|
#
|
|
# Lidarr consumers → highly selective, quality-first discovery
|
|
# Sonarr consumers → balanced, family-aware episodic intake
|
|
# Radarr consumers → broader flexibility with intelligent filtering
|
|
#
|
|
# Temporal Decay
|
|
# Old behavioral signals lose influence over time (one decay unit per 30
|
|
# days). Prevents permanent genre lock-in, historical bias accumulation,
|
|
# and dead-user score dominance.
|
|
#
|
|
# Consumer Owns the Decision
|
|
# The engine returns ACCEPT/REJECT/SCORE. What happens next is entirely
|
|
# the consumer's concern — the engine never acts on its own verdict.
|
|
#
|
|
# ==============================================================================================
|
|
# OPERATIONAL SAFEGUARDS
|
|
# ==============================================================================================
|
|
#
|
|
# No Side Effects — Structural, Not Incidental
|
|
# Writes no files, makes no API calls, deletes nothing, starts nothing. This is the
|
|
# safeguard: a scoring mistake here can only ever produce a wrong number, never a wrong
|
|
# action. Keep it that way — the moment this library acquires a side effect, every consumer
|
|
# inherits it silently.
|
|
#
|
|
# No Root, No Lock, No detect_hosts — Deliberate
|
|
# Correct for a sourced library and should not be "fixed" to match the executable scripts.
|
|
# There is no state for a lock to protect, no privileged operation to gate, and no
|
|
# host-specific config to alias. It runs entirely inside the caller's process.
|
|
#
|
|
# Caller Owns the Verdict
|
|
# ACCEPT/REJECT is a return value, not an instruction. Nothing here can cause a candidate
|
|
# to be added, removed, or downloaded — the consumer decides what a verdict means.
|
|
#
|
|
# Literal Duplicate Matching
|
|
# is_duplicate_candidate() matches with grep -Fx, treating the candidate as data rather
|
|
# than a pattern. Artist and title strings routinely contain regex metacharacters, and a
|
|
# false positive here silently discards a genuinely new candidate.
|
|
#
|
|
# Decay Floors at Zero
|
|
# apply_temporal_decay() clamps at 0, so an old signal can never become a negative score
|
|
# that drags an otherwise-passing candidate below threshold.
|
|
#
|
|
# Integer Arithmetic Throughout
|
|
# All scoring is integer. No floating point means no locale-dependent decimal parsing and
|
|
# no rounding drift between hosts.
|
|
#
|
|
# ==============================================================================================
|
|
# CONFIGURATION
|
|
# ==============================================================================================
|
|
#
|
|
# None, by design. The engine reads no conf file and no environment variable.
|
|
#
|
|
# Thresholds, weights and strictness profiles live with the consumer — see
|
|
# LIDARR_DISCOVERY_* / SONARR_DISCOVERY_* / RADARR_DISCOVERY_* in master.conf. That is what
|
|
# keeps the core domain-agnostic: music strictness cannot leak into TV intake, because the
|
|
# engine never learns which domain it is scoring for.
|
|
#
|
|
# ==============================================================================================
|
|
# RUNTIME MODES
|
|
# ==============================================================================================
|
|
#
|
|
# None — this file is sourced, never executed:
|
|
#
|
|
# source "$SCRIPT_DIR/../Kernel/decision_engine.sh"
|
|
#
|
|
# It has no argument parsing, no --dry-run and no --status, because it takes no action that
|
|
# a dry run could suppress.
|
|
#
|
|
# ==============================================================================================
|
|
# FUNCTIONS
|
|
# ==============================================================================================
|
|
#
|
|
# score_candidate user_score popularity_score recency_score quality_score
|
|
# Returns TOTAL_SCORE (integer sum). Consumer defines actual weight values.
|
|
#
|
|
# evaluate_threshold score minimum
|
|
# Returns 0 (pass) or 1 (fail). Used as: if evaluate_threshold ...
|
|
#
|
|
# apply_temporal_decay score age_days
|
|
# Returns adjusted score. Subtracts (age_days / 30), floor at 0.
|
|
#
|
|
# is_duplicate_candidate candidate history_file
|
|
# Returns 0 (duplicate found in history_file) or 1 (not found).
|
|
#
|
|
# make_decision score threshold
|
|
# Returns "ACCEPT" or "REJECT". Calls evaluate_threshold internally.
|
|
#
|
|
# ==============================================================================================
|
|
|
|
# ── score_candidate ───────────────────────────────────────────────────────────
|
|
score_candidate() {
|
|
|
|
local user_score="${1:-0}"
|
|
local popularity_score="${2:-0}"
|
|
local recency_score="${3:-0}"
|
|
local quality_score="${4:-0}"
|
|
local TOTAL_SCORE
|
|
|
|
TOTAL_SCORE=$(( \
|
|
user_score + \
|
|
popularity_score + \
|
|
recency_score + \
|
|
quality_score \
|
|
))
|
|
|
|
echo "$TOTAL_SCORE"
|
|
}
|
|
|
|
# ── evaluate_threshold ───────────────────────────────────────────────────────
|
|
evaluate_threshold() {
|
|
|
|
local score="$1"
|
|
local minimum="$2"
|
|
|
|
if (( score >= minimum )); then
|
|
return 0
|
|
fi
|
|
|
|
return 1
|
|
}
|
|
|
|
# ── apply_temporal_decay ─────────────────────────────────────────────────────
|
|
apply_temporal_decay() {
|
|
|
|
local score="$1"
|
|
local age_days="$2"
|
|
|
|
local decay=$(( age_days / 30 ))
|
|
|
|
local adjusted=$(( score - decay ))
|
|
|
|
(( adjusted < 0 )) && adjusted=0
|
|
|
|
echo "$adjusted"
|
|
}
|
|
|
|
# ── is_duplicate_candidate ───────────────────────────────────────────────────
|
|
is_duplicate_candidate() {
|
|
|
|
local candidate="$1"
|
|
local history_file="$2"
|
|
|
|
# -F -x, not "^$" anchors: the candidate is data, not a pattern. Interpolating it into a
|
|
# regex makes every metacharacter in an artist or title active — "R.E.M." matches "RxExMy",
|
|
# and an unbalanced bracket makes grep error out entirely. Either way the caller reads the
|
|
# result as "already seen" and silently skips something genuinely new. -F disables regex,
|
|
# -x anchors the whole line, which is exactly what the anchors were reaching for.
|
|
grep -qiFx -- "$candidate" "$history_file" 2>/dev/null
|
|
}
|
|
|
|
# ── make_decision ─────────────────────────────────────────────────────────────
|
|
make_decision() {
|
|
|
|
local score="$1"
|
|
local threshold="$2"
|
|
|
|
if evaluate_threshold "$score" "$threshold"; then
|
|
echo "ACCEPT"
|
|
else
|
|
echo "REJECT"
|
|
fi
|
|
} |