smart-sourcing

Smart Sourcing

Intelligent sourcing that requires citations only when the cost is justified by the value of verification.

Philosophy

Full sourcing is prohibitively expensive (10-16x token increase). Smart sourcing targets high-value claims where verification materially improves accuracy.

When to Source

REQUIRE Sources

DO NOT Require Sources

Decision Framework

Is this a factual claim?
├── No → No source needed
└── Yes → Would being wrong cause harm?
    ├── No (minor inconvenience) → No source needed
    └── Yes → Is verification cost-effective?
        ├── No (>5k tokens for low-value claim) → State uncertainty
        └── Yes → Perform web search, cite source

Implementation

Inline Citations

When sourcing is warranted:

Python 3.12 introduced `type` statement for type aliases[1].

Sources:
1. [PEP 695](https://peps.python.org/pep-0695/)

Uncertainty Markers

When sourcing isn't cost-effective but uncertainty exists:

Python 3.12 introduced the `type` statement (verify in release notes for your version).

Confidence Levels

Token Cost Analysis

Examples

Good: Source Version-Specific Claim

User: What's new in FastAPI 0.109?

Response: FastAPI 0.109 added Pydantic v2 support as default[1] and
improved OpenAPI schema generation[2].

Sources:
1. [FastAPI Release Notes](https://fastapi.tiangolo.com/release-notes/)
2. [GitHub Release 0.109.0](https://github.com/tiangolo/fastapi/releases/tag/0.109.0)

Good: No Source for General Concept

User: How does async/await work?

Response: Async/await enables non-blocking I/O by allowing coroutines
to yield control while waiting for operations. The event loop manages
execution, resuming coroutines when their awaited operations complete.

Good: Uncertainty Marker

The connection pool default is typically 10 connections (verify in your
database driver documentation as this varies by library and version).

Integration

Works with:

• conserve:response-compression - Keeps sourced responses concise
• conserve:token-conservation - Weighs source cost vs value
• memory-palace:knowledge-intake - Full sourcing for knowledge corpus

When To Use Full Sourcing

When NOT To Use

• Internal project code that doesn't need citations
• Casual conversation without factual claims

Escalate to full sourcing (accept high token cost) for:

• Knowledge corpus entries (permanent documentation)
• Security advisories (safety-critical)
• Compliance/legal claims (audit requirements)
• Research tasks (user expects thorough investigation)

For these cases, use memory-palace:knowledge-intake workflow which is designed for thorough sourcing.

Exit Criteria

• Every claim categorized as requiring a source (version numbers, performance claims, security recommendations, API specs, release dates, pricing/limits, deprecated features) has an inline citation or a verification note
• No web search performed for claims categorized as general concepts, code examples, opinions, common knowledge, logical derivations, or stable syntax
• Claims with medium confidence carry an explicit uncertainty marker (e.g., "verify in release notes for your version") rather than being stated as fact or sourced at high token cost
• Decision tree applied: factual claim → harm if wrong → verification cost-effective → source; each branch followed explicitly for claims exceeding medium confidence