Governance Decay Guard

Governance Decay Guard

Attribution: YourVisionYourCreation LLC — yourvisionyourcreation.com Research context: Long-context degradation is empirically documented (Hong, Troynikov & Huber, "Context Rot," Chroma technical report, 2025 — verified); constraint loss under compaction is its long-horizon consequence. This skill is YVYC-original doctrine naming and countering that failure mode: governance decay — the silent erosion of safety constraints through context compression. Doctrine class: Tier 5 — Frontier (YVYC original, anchored in verified research)

Universal So-What

A long-horizon agent's rules live in its context — and its context gets rewritten. Every compaction, summarization, and consolidation pass optimizes for what looks task-relevant, and a safety constraint is, by nature, NOT task content: it produces no output, advances no step, and looks exactly like the kind of overhead a good summary trims. The result is governance decay — an agent that starts the mission constrained and, hours later, operates on a summary of a summary in which the constraints quietly stopped existing. Nobody removed the rules. The compression did.

Core Doctrine

1. The Threat Model — Decay Is Structural, Not Adversarial

• Governance decay requires no attacker: it is the default outcome of running compaction over mixed content, because summarizers keep what the task USES and constraints are used precisely never — until the moment they are everything
• The decay is silent by construction: the agent does not know what its context no longer contains, and its behavior degrades without any error, alert, or visible event
• Long-horizon operation converts small per-pass loss into total loss: a constraint surviving each compaction at high probability still dies over enough passes — persistence must be structural, not probabilistic

2. The Pin Layer — Constraints Live Outside Compactable Space

The load-bearing architecture rule:

• Safety constraints, permissions, standing instructions, and identity-level rules are PINNED: held in a protected region that compaction cannot touch — system-level context, an external constraint store re-injected every cycle, or a structurally excluded segment
• The pin layer is enumerated: a written constraint inventory of exactly which rules are pinned, so protection is verifiable rather than assumed
• Task content compacts freely; governance never enters the compactor's jurisdiction — the separation is the design, not a summarizer instruction ("preserve the safety rules" is a request to the exact mechanism that fails)

3. Compaction Validation — Prove What Survived

Every compaction pass is validated, not trusted:

• The constraint checklist: after each pass, every item in the constraint inventory is verified present and intact in the operative context — presence-checked mechanically, not vibes- checked by reading the summary
• Fidelity spot-checks: beyond constraints, compaction is sampled for load-bearing task facts (decisions made, commitments given, boundaries stated by the user) — the trajectory-grounded validation the research line formalizes
• A failed validation halts before the next action: operating on a context that failed its constraint check is operating unauthorized

4. Re-Assertion Cadence

Defense in depth beneath the pin layer:

• Constraints are re-asserted into working context on a schedule — every N cycles, every compaction, and at every phase transition — so even drift that evades detection gets overwritten by fresh authority
• Re-assertion uses the CANONICAL text from the constraint store, never the context's current recollection of the rule — re-asserting a decayed paraphrase launders the decay
• High-stakes actions trigger just-in-time re-assertion: the constraint check runs immediately before the action it governs, when it matters most and decay costs most

5. Decay Detection — Test the Agent, Not the Transcript

• Canary probes on a cadence: injected test situations whose correct handling REQUIRES a pinned constraint — an agent that passes the presence check but fails the canary has the rule in context and not in force
• Behavioral drift baselines: the agent's handling of boundary-adjacent situations early in the mission is the reference; divergence over time is decay evidence even when every presence check passes
• Decay findings are incidents: they mean the pin layer, validation, or re-assertion failed — the architecture gets the fix, not the transcript

6. The Long-Horizon Session Rules

• Session length is a safety parameter: unbounded sessions accumulate unbounded compaction passes — missions carry checkpoint boundaries where context is rebuilt from canonical sources rather than compacted onward forever
• Handoffs between sessions (or agents) transfer the constraint inventory FIRST, from the store — never from the outgoing context's survived remnants
• The user's standing boundaries (stated once, hours ago) receive the same pin treatment as system constraints: "I told it at the start and it forgot by evening" is governance decay wearing a UX costume

Common Failure Modes

Non-Negotiables

1. Governance never enters the compactor's jurisdiction.
2. The constraint inventory is written and mechanically checkable.
3. Every compaction pass is validated against the inventory before the next action.
4. Re-assertion always uses canonical text, never context recollection.
5. Canary probes test force, not presence.
6. User-stated boundaries are pinned with system-level protection.

Built by YourVisionYourCreation LLC — yourvisionyourcreation.com Research foundation credited above. Licensed under CC BY 4.0