Skip to main content

​
Tiered Knowledge Delivery (Push / Hook / Pull)

The mechanism behind this wiki’s “agent-first” claim (see ONBOARDING.md): agents get pre-synthesized structure without paying a search cost for it, while the bulk of the wiki stays out of every session’s context.

​
Two failure modes this replaces

  • Pure preload: push the whole index/wiki into every session. Correct but wasteful — a prior decision removed a 32KB wiki index from global startup, reclaiming ~8,400 tokens/session, precisely because most sessions never touch most pages.
  • Pure on-prompt search: nothing is preloaded; every fact requires the agent to decide to invoke wiki-context / qmd query and pay a retrieval round-trip. Cheap when idle, but only as reliable as the agent’s judgment to search — high-frequency, high-cost-of-violation knowledge (LOC gate, model-tier routing, epistemic discipline) shouldn’t depend on that judgment firing every time.

​
The three tiers

Promotion between tiers is decided by frequency × cost-of-violation, not by how interesting a page is. Content earns Tier 0 only if it’s both hit often and expensive to get wrong.

​
Provenance

Decided 2026-06-12 via a read-only 4-lens agent council (agent-consumer, human-learner, maintainer/cost, adversarial verifier) — see docs/wiki-dual-use-audit-2026-06.md for the full scorecard, prioritized fix list, and the hard constraint this model must satisfy: pull stays the default; nothing large gets pushed into every session. Correctness of Tier 0 is checked by the wiki’s own lint procedure’s Tier-0 distillation sync step (CLAUDE.md “Lint” operation) — for each wikilink cited in applied-ai.md / global-prevention-rules.md, verify the linked page still says the same thing.

​
Open question this doesn’t answer

The model says where knowledge lives once it’s structured. It says nothing about how raw session experience becomes Tier-0/Tier-2 material in the first place — right now that’s manual (synthesize-mistakes, ad-hoc wiki ingest). Knowledge Crystallization Cycle (KCC) proposes a formal answer to that gap (proposed, single-source, not yet reviewed) — if adopted, it would be the authoring pipeline that feeds this delivery model, not a replacement for it.