Knowledge Engine · Layer 01

Confidence Scoring

Every query result carries a deterministic confidence score. Base 100, with ten specific deduction rules. No silent aggregation — every penalty appears individually in penalties[].

Design rule

Below confidence 60, suggested_next only recommends human actions — never automated steps. Below 20, the system refuses to proceed entirely.

---

Penalty Matrix

Condition	Deduction	Rationale
LSP unavailable for queried language	−25	Symbol resolution unreliable
Index 1+ commits behind HEAD	−20	Structural data may be stale
Unresolved cross-layer conflict	−20	Two sources disagree on verifiable facts
Recursive CTE hit hub_node cap	−15	Blast radius truncated
Summary-only evidence (no structural backing)	−15	Evidence is lossy, not authoritative
ctags fallback active	−10	Reduced parsing precision
Vector dormant but NEEDS_EXPLORE triggered	−10	Semantic search unavailable
Graph truncated at depth limit	−10	Deeper dependencies unexplored
Dirty working tree	−10	Index reflects last commit, not current edits
Tier 3/4 degradation	−30	Multiple sources materially degraded

score = 100 − sum(deductions)

---

Confidence Labels

80–100

HIGH

May suggest automated next step.

60–79

MEDIUM

Suggest narrower query or reindex.

40–59

LOW

Suggest human verification only.

Score	Label	Behavior
80–100	High	May suggest automated next step
60–79	Medium	Suggest narrower query or reindex
40–59	Low	Human verification only
20–39	Unreliable	Suggest stopping automated work
0–19	Insufficient evidence	Cannot answer. Manual investigation required
N/A	Unknown	Cannot classify this query. Please rephrase

---

Governance Rules

• Every penalty fires individually — zero silent aggregation
• penalties[] array contains each: {rule, deduction, detail}
• New penalties require a 30-day trace review before being added
• Monthly calibration via dev replay --all --since 30d tunes penalty weights