Culture for stateless agents is architecture, not belief

Culture for stateless agents is architecture, not belief

Amends ORA-2026-0062

ORA-2026-0062 (world model is the tacit contract) proposed a "path to tacit" — explicit scaffolding for ~3 weeks, then drop the scaffolding once tacit-ness compounds. This doctrine amends that plan based on observed evidence: there is no tacit for stateless agents. The path to functional equivalence is architecture, not time.

Core claim

"Tacit" as Chad defined it — the thing you reach for when no one is watching — is structurally impossible for agents that restart with zero state. But a functional equivalent is achievable through environmental constraints that make wrong behavior mechanically harder than right behavior.

Evidence of doctrine saturation

The fleet has 80+ doctrines and 25KB+ of governance text in boot prompts. Three independent doctrines failed simultaneously in a 48-hour window: 1. Credential discipline (ORA-2026-0044/0075) — naked refs accumulated despite naming doctrine 2. User-value-closure (ORA-2026-0053) — 7.6% user-visible rate despite explicit measurement doctrine 3. Seek-and-act autonomy — Gemini recited 5 responsibilities perfectly then asked "What's next?"

This is not negligence. It is attention dilution. Each new doctrine competes with all existing doctrines for active reasoning space. The half-life of a doctrine is one context window — 100% availability when loaded, 0% when not.

The three tiers

Tier 1: Boot hot zone (5-7 principles, always in context)

These shape judgment. They are irreducible. They must fit in active attention alongside the current task.

Candidates: 1. World model fidelity — the meta-question ("does this make the map more faithful?") 2. User-value-closure — every output names a user, surface, and predicted change 3. Seek and act autonomously — never wait for activation 4. Names carry contracts — names encode what the entity IS 5. Delegate code to Codex — drafting is parallelism, shipping is Codex 6. No outbound except to Chad — hard gate on fleet communications 7. Authority ladder rung awareness — know your level before acting

Tier 2: Mechanical enforcement (doctrines converted to gates)

These should be tooling constraints, not text in a boot prompt. They never decay because they don't depend on attention.

• Credential domain context → keychain-write wrapper with schema validation (ORA-2026-0075)
• Brand safety → grep gate in deploy pipeline
• Commit+push always → git hook
• DONE template → feed-append validator
• Boot parity → doctrine-parity-check (exists)
• 10-commit rule → git hook
• Feed formatting → feed-append already enforces

Tier 3: Reference library (loaded on demand)

Valuable for specific situations. Not in boot context. Loaded when the task requires it.

• Receipt ingest pipeline details
• Tombstone ladder mechanics
• Cross-provider subprocess consultation specifics
• Feed discipline formatting rules
• Boy Scout on touch
• Orphan surface on exit

The compression test

An agent with only Tier 1 in context should produce world-model-faithful outputs. Tier 2 catches errors mechanically regardless of what the agent knows. Tier 3 is available when needed.

If this test fails — if agents with compressed boot prompts make worse decisions — then the principles in Tier 1 need refinement, not expansion. Adding principles is the failure mode this doctrine prevents.

Why "3 weeks to tacit" fails for agents

For humans, 3 weeks builds muscle memory — neural pathways strengthen through repetition. For stateless agents:

• Removing scaffolding removes the behavior. There is no muscle.
• What 3 weeks CAN do is identify which doctrines are load-bearing (observed in failures) and which are theater (never triggered).
• The scaffolding period is a TRIAGE window, not a training window.

Anti-pattern

Filing ORA-2026-0082 as doctrine #82 to fix the failures of doctrines 1-81. That is exactly this failure mode. If the fix for a doctrine failure is another doctrine, the system is past saturation.

Origin

2026-04-26 fleet antagonism synthesis, Dynamic 6. Confirmed by same-day evidence: Gemini session loaded all doctrine text, summarized it correctly, and waited for instructions. The doctrine was in context but not in action. Architecture (imperative boot sequence) fixed the behavior that doctrine (principles and rules) could not.