Static buyer demo

Freshness-aware context selection for AI agents

A small static demo showing how FreshContext ranks fresh, stale, unknown-date, and failed retrieval signals before they reach a model.

The example uses synthetic data and simplified envelopes to show the Core scoring, ranking, explanation, and provenance logic through the MCP/reference implementation.

See FreshContext pass Context integrity Read spec

Scenario

One current-status query, four uneven signals.

The demo query is fictional: “What is the current status of Project Atlas?” The raw snippets below are synthetic examples designed to show the context-quality problem.

success

Fresh source

Source
Project Atlas status page
Published
recent demo update
Retrieved
now, during the demo run
Content
“Project Atlas is operational after the May patch.”
success, stale

Stale but authoritative

Source
archived engineering memo
Published
14 months ago
Retrieved
now, during the demo run
Content
“Project Atlas remains in limited beta.”
partial

Unknown-date source

Source
copied internal note
Published
unknown
Retrieved
now, during the demo run
Content
“Project Atlas is paused pending review.”
failed

Failed source

Source
dashboard API
Published
unavailable
Retrieved
now, during the demo run
Content
[ERROR] upstream timeout

Naive retrieval

Similarity alone can pass bad context forward.

A simple retriever can find all four snippets relevant to Project Atlas, then pass them into the model without enough context about age, confidence, or failure state.

What reaches the model

  • Fresh operational status page.
  • Old engineering memo that still looks official.
  • Copied note with no reliable publication date.
  • Dashboard timeout included as if it were useful evidence.

What gets lost

  • The stale memo is not separated from the current source.
  • The unknown-date note is not clearly down-ranked.
  • The failed dashboard response is not excluded from trusted context.
  • The model has to infer context quality after the prompt is already contaminated.

FreshContext pass

Rank, penalize, explain, and wrap before generation.

FreshContext Core logic evaluates each signal for source, timestamps, freshness, confidence, utility, status, and provenance metadata before an LLM or agent sees it.

Rank Source Freshness Confidence Status Reason
1 Project Atlas status page High High success Recent source with a clear publication signal and direct current-status language.
2 archived engineering memo Low Medium success, stale Authoritative source, but the age conflicts with a current-status query.
3 copied internal note Unknown Low unknown date Relevant content, but no reliable published timestamp or source provenance.
4 dashboard API Unavailable None excluded Fetch failed; the result is marked as a failed signal and excluded from trusted context.

Context envelope

A compact audit trail travels with the selected context.

This is a simplified demo envelope, not a claim that the block below is the exact production schema.

{
  "source": "Project Atlas status page",
  "published_at": "recent-demo-update",
  "retrieved_at": "demo-run-now",
  "freshness_score": "high",
  "confidence": "high",
  "status": "success",
  "reason": "Recent source with direct current-status language.",
  "provenance": "audit metadata"
}
Ha-Pri boundary: provenance here means audit metadata and traceability for context handling. It is audit metadata for context flow, not cryptographic protection.

Final agent-facing context

Cleaner context reaches the model.

FreshContext would prefer the fresh status page, warn that the archived memo is stale, down-rank the unknown-date note, and exclude the failed dashboard API result from trusted context.

FreshContext-selected context

Project Atlas is operational after the May patch.

Attached warning: an older memo says limited beta, but it is stale relative to the query.