scoring-model

name	scoring-model
description	Scoring model covering ChromaDB collections, fusion formula, and embedding synthesis. Use when implementing or modifying the scoring pipeline: indexer embedding synthesis, scorer collection queries, ranker fusion formula, or export score breakdowns.

Scoring Model — Collections, Formula & Embedding Synthesis

When This Skill Applies

When implementing or modifying the scoring pipeline: indexer embedding synthesis, scorer collection queries, ranker fusion formula, or export score breakdowns.

---

The Five ChromaDB Collections

Collection	Documents	Built By	Score
resume	Resume chunked by ## section	index_resume()	fit_score
role_archetypes	One doc per archetype (synthesized)	index_archetypes()	archetype_score
global_positive_signals	One doc per rubric dimension with signals_positive	index_global_positive_signals()	culture_score
negative_signals	One doc per rubric dimension + archetype negatives	index_negative_signals()	negative_score
decisions	Past JDs labeled yes/no/maybe	decide / review commands	history_score

Two Orthogonal Scoring Axes

• Right kind of role — role_archetypes (archetype_score) + resume (fit_score)
• Right kind of environment — global_positive_signals (culture_score) + negative_signals (negative_score)

---

Embedding Synthesis

Archetype, positive signal, and negative signal embeddings are synthesized from TOML fields, not used raw. See embedding-synthesis.md for code patterns, the what-gets-embedded decision table, and non-semantic score details.

Global Positive Signal Synthesis

One document per rubric dimension that has signals_positive:

# Pseudocode — see indexer.py for actual implementation
for dim in rubric["dimensions"]:
    if signals := dim.get("signals_positive"):
        text = f"{dim_name}: " + ", ".join(signals)
        # → one ChromaDB document with metadata {"source": dim_name}

Key rules:

• Only signals_positive arrays reach the embedding model
• description, minimum_target, weight_* fields are never embedded
• Dimensions without signals_positive produce no document (no error)

---

Score Formula

final_score = (
    archetype_weight  * archetype_score  +   # 0.5 default
    fit_weight        * fit_score        +   # 0.3 default
    culture_weight    * culture_score    +   # 0.2 default
    history_weight    * history_score    +   # 0.2 default
    comp_weight       * comp_score       -   # 0.15 default
    negative_weight   * negative_score       # 0.4 default
) * (0.0 if disqualified else 1.0)

Rules

• Weights are read from settings.toml, never hardcoded
• Weights do NOT need to sum to 1.0 — this is a weighted composite, not a probability
• If final_score <= 0.0, floor it at 0.0 (clean discard)
• A disqualified role always scores 0.0 regardless of component values
• Roles below min_score_threshold (default 0.45) are excluded from output
• negative_score has inverted semantics: high similarity = bad outcome
• culture_score has positive semantics: high similarity = good outcome

---

Score Explanation Format

Archetype: 0.81 · Fit: 0.74 · Culture: 0.65 · History: 0.62 · Comp: 0.70 · Negative: 0.12 · Not disqualified

All six components shown. Separator is · (middle dot), not | (breaks Markdown tables).

---

Missing Collection Behavior

Collection	Missing Behavior
resume	Auto-indexed before scoring via _ensure_indexed()
role_archetypes	Auto-indexed before scoring via _ensure_indexed()
global_positive_signals	Returns 0.0 culture_score — not an error
negative_signals	Returns 0.0 (no penalty) — not an error
decisions	Returns 0.0 history_score — not an error

---

Config Fields

settings.toml — [scoring] section

Field	Default	Range	Notes
archetype_weight	0.5	[0.0, 1.0]	Role-type axis
fit_weight	0.3	[0.0, 1.0]	Role-type axis
culture_weight	0.2	[0.0, 1.0]	Environment axis
history_weight	0.2	[0.0, 1.0]	Learned preferences
comp_weight	0.15	[0.0, 1.0]	Compensation signal
negative_weight	0.4	[0.0, 1.0]	Environment axis (penalty)

CLI Flags

• --archetypes-only rebuilds role_archetypes, global_positive_signals, and negative_signals (all three)