Skip to content

trust-tiers.md

Latest commit

 

History

History
110 lines (79 loc) · 3.73 KB

trust-tiers.md

File metadata and controls

110 lines (79 loc) · 3.73 KB

Trust Tiers

Every resource in qp-vault has a trust tier that governs its search relevance. Trust is not just metadata; it directly affects which results appear first.

The Four Tiers

Tier Search Weight Auto-Behavior Use For
CANONICAL 1.5x Immutable (content updates create new version) Official SOPs, approved policies, signed contracts
WORKING 1.0x (default) Standard Drafts, in-progress documents, proposals
EPHEMERAL 0.7x Auto-archives after TTL (default 90 days) Meeting notes, chat transcripts, temporary memos
ARCHIVED 0.25x Low relevance Superseded versions, retired procedures

How Trust Affects Search

The search ranking formula:

relevance = (0.7 * vector_similarity + 0.3 * text_rank) * trust_weight * freshness_decay

A CANONICAL document with 0.6 cosine similarity scores higher than a WORKING document with 0.8:

  • CANONICAL: 0.6 * 1.5 = 0.90
  • WORKING: 0.8 * 1.0 = 0.80

This means official SOPs always surface above drafts, even when drafts are slightly more textually relevant.

Setting Trust

# On creation
vault.add("SOP content", trust_tier="canonical")
vault.add("Draft notes", trust_tier="working")
vault.add("Standup notes", trust_tier="ephemeral")

# After creation
vault.update(resource_id, trust_tier="canonical")

Trust changes emit a TRUST_CHANGE audit event.

Freshness Decay

Trust tiers also control how quickly a document's relevance decays over time:

Tier Half-Life Meaning
CANONICAL 365 days Stays relevant for a year
WORKING 180 days Decays over 6 months
EPHEMERAL 30 days Decays in a month
ARCHIVED 730 days Slow decay (historical reference)

The decay function: freshness = exp(-age_days / half_life * ln(2))

A 180-day-old WORKING document has 50% of its original freshness score. A CANONICAL document of the same age retains 70%.

Custom Weights

Override trust weights via VaultConfig:

from qp_vault.config import VaultConfig

config = VaultConfig(
    trust_weights={
        "canonical": 2.0,     # Even stronger boost
        "working": 1.0,
        "ephemeral": 0.3,     # Heavier penalty
        "archived": 0.1,
    },
    freshness_half_life={
        "canonical": 730,     # 2-year half-life
        "working": 90,        # 3-month half-life
        "ephemeral": 14,      # 2-week half-life
        "archived": 365,
    },
)

vault = Vault("./knowledge", config=config)

Data Classification

Orthogonal to trust tiers, every resource also has a data classification that controls AI provider routing:

Classification Cloud AI Local AI Encryption Audit
public Allowed Allowed Optional Standard
internal (default) Allowed Allowed Optional Standard
confidential Blocked Allowed Required Enhanced
restricted Blocked Allowed Required Every read 
vault.add("Public announcement", classification="public")
vault.add("Internal memo", classification="internal")
vault.add("Patient records", classification="confidential")
vault.add("Classified intel", classification="restricted")