02 · Guardrails (what the framework prevents)¶
Where the Axioms state the invariant rules, Guardrails document the concrete failure modes those rules exist to prevent — what each failure looks like, how the framework detects it, and how to recover when it happens anyway.
In plain terms: this section is the catalogue of things that can go wrong when AI agents work together, and what to do about each one — useful whether you are setting a federation up or fixing one that has gone sideways. New here? Start with the mental model.
Axioms vs. guardrails¶
The two sections are companions, read from opposite directions:
| Axioms | Guardrails | |
|---|---|---|
| Direction | The rule, stated affirmatively | The pathology, stated as a failure |
| Question answered | "What must always be true?" | "What goes wrong when it isn't — and what do I do?" |
| Form | Invariant specification | Symptom → detection → recovery |
| When you reach for it | Designing or onboarding | Debugging, auditing, or recovering |
An axiom tells you to keep load-bearing state on disk. A guardrail tells you what a lost decision looks like three sessions later, how to notice the loss, and how to rebuild. You need both: the rule to follow, and the failure signature to recognize when the rule slipped.
The four pathologies¶
Every guardrail in this section defends one or more of the four pathologies that multi-tier AI-agent federations fall into. These are the same four pathologies the axioms defend — the guardrails are the failure-facing view of the same defenses.
1. Context pollution¶
A tier's context fills with detail it does not own — sibling churn, sub-tier reasoning, integration noise. Judgment degrades (cognitive overhead) and biases (saw the trees, lost the forest). Worst form: a parent that delegated work still absorbs the delegate's full context on retrieval, so isolation buys nothing. It arrives by two vectors — a tier auto-reading what it doesn't own, and a sender pushing detail across the bus the receiver never needed.
Guarded by: Pollution containment (spatial vector), The minimal-sufficient bus (relay vector), Retrieval discipline (the pull vector), Single-live-writer.
2. Hallucination drift¶
An AI tier's institutional memory ("I remember we decided X") is confidently wrong. Acted upon, the federation drifts away from its actual history one plausible-but-false fact at a time. The drift compounds: each fabricated premise becomes the basis for the next decision.
Guarded by: Hallucination defense, Stale-snapshot detection.
3. Role confusion¶
A tier acts outside its jurisdiction — a mentor does labour, a doer improvises authority, a sub-tier writes into a parent's folder, a message is consumed by the wrong tier because the voice matched. Coordination integrity dissolves.
Guarded by: Brief completeness, Single-live-writer.
4. Trust erosion¶
The founder observes a claim that doesn't match disk, a decision that evaporated, or a "done" that wasn't pushed. Once observed, the founder can no longer hold the narrow relay-and-lost+found posture and must return to micromanaging — which is exactly the cost the framework exists to spare them. A disaster that loses unbounded work — rather than the bounded, recoverable sliver — erodes trust the same way.
Guarded by: Failure modes, Disaster reconciliation, Brief completeness, and the persistence discipline behind every page here.
The nine guardrail pages¶
| Page | Pathology guarded | Core mechanism |
|---|---|---|
| Pollution containment | Context pollution (spatial vector) | Firewall + state-tracking scope confine each tier to its own granularity. |
| The minimal-sufficient bus | Context pollution (relay vector) | Relay the least the receiver needs; persist the detail in substrate, pulled on demand. |
| Retrieval discipline | Context pollution (the pull vector) | Pull the minimal-sufficient slice: partitioned index, as-of + re-verify, pointer-not-truth. |
| Hallucination defense | Hallucination drift | Verification-at-citation: every load-bearing claim verified against substrate before action. |
| Stale-snapshot detection | Hallucination drift (the firewall-leak class) | "As of last tagged return" discipline; forensic descent before action. |
| Failure modes | Trust erosion (all classes) | The documented failure classes, each with symptom → recovery. |
| Disaster reconciliation | Trust erosion (unclean-event class) | Bounded loss by design; aggressive dual-plane recon before any tier boots. |
| Brief completeness | Role confusion, trust erosion | No placeholders at the relay boundary; escalate, never improvise. |
| Single-live-writer | Context pollution, role confusion | One writer holds jurisdiction; fetch-before-push for everyone else. |
How to read this section¶
If you are operating a federation, read Failure modes first — it is the index of everything that can go wrong, with the recovery for each — then Disaster reconciliation for the unclean-event runbook you will reach for after a power loss or host crash.
If you are designing or auditing one, read Pollution containment and Hallucination defense — they explain why the structural choices in the axioms are not negotiable.
Each page reads standalone in about five minutes and cross-links to the axiom it operationalizes.