Compaction Integrity Protocol

Compaction Integrity Protocol

Attribution: YourVisionYourCreation LLC — yourvisionyourcreation.com Research anchor (verified): "Context Rot: How Increasing Input Tokens Impacts LLM Performance" — Hong, Troynikov & Huber, Chroma technical report, 2025 — documenting that performance degrades as input grows, even within nominal context limits. Active 2026 research on agent self-compaction confirms the compaction side of the problem is a live frontier. The integrity protocol itself is YVYC original. Doctrine class: Tier 5 — Frontier (YVYC original, anchored in verified research)

Universal So-What

Long-horizon agents face a two-front war. Front one: context windows fill, forcing compression — and every summary is a lossy bet about what tomorrow will need. Front two: even UNCOMPRESSED long context degrades — models attend unevenly across long inputs, and a fact present at token 200,000 is not a fact the model reliably uses. The protocol fights both fronts with the same weapon: an explicit inventory of what is load-bearing, verified present AND effective after every context operation.

Core Doctrine

1. The Load-Bearing Inventory

Before any context can be safely compressed, the operation must know what cannot be lost:

Everything else is compressible working material. The inventory is maintained AS the session runs — reconstructing it at compaction time from a degraded context is asking the failure to audit itself.

2. Compaction as Engineering, Not Prayer

• Compaction is structured extraction against the inventory, never freeform "summarize the above": the inventory classes are pulled explicitly, verbatim where wording matters (commitments, boundaries), then the residue is summarized
• Verbatim-preservation rule: decisions and user statements whose exact wording carries force are quoted forward, not paraphrased — paraphrase drift across multiple compactions is how "never do X" becomes "prefer avoiding X" becomes silence
• Compaction happens at STRUCTURAL boundaries (task completion, phase transition) wherever possible, not at arbitrary token thresholds mid-reasoning — compressing in the middle of an open thought loses the thought

3. Post-Compaction Validation

Every compression is proven before work continues:

• Inventory check: every load-bearing item verified present in the surviving context — mechanically, item by item
• Function check: presence is not enough (the context-rot finding applies to summaries too) — spot-probe that the surviving context ANSWERS correctly: "what did we decide about X?" tested, not assumed
• A failed validation stops the line: the compaction is re-run against the inventory, and work does not proceed on a context that failed its check

4. The Length Discipline — Rot Before the Limit

Context degrades before it overflows:

• Operating length is a quality parameter, not just a capacity one: the effective ceiling is where retrieval quality degrades, which arrives well before the token limit — long-context performance falls as input grows even on tasks the model handles perfectly when short
• Position matters: critical standing material lives where attention is reliable (the pinned region), and freshly relevant material is re-surfaced near the point of use rather than trusted to be found at token 180,000
• Distillation beats accumulation: a session that only ever appends is choosing rot; periodic deliberate compaction against the inventory keeps the working set inside the model's effective attention, which is the entire point

5. Handoff Compression

When one context becomes another's starting point (new session, new agent, checkpoint rebuild):

• The handoff document IS the inventory, rendered: decisions with reasons, commitments verbatim, boundaries verbatim, state, ground truth with its verification status
• Truth-first rule (inherited from agent-stack-org-design): the handoff carries what is broken, unfinished, and uncertain — a handoff that compresses away the problems hands the successor a discovery phase disguised as a clean start
• The receiving side validates on arrival: the same function checks, run against the handoff, before work resumes

6. The Compression Ledger

• Every compaction event is logged: when, what was dropped, what was preserved, validation results — so a fact discovered missing three hours later can be traced to the operation that lost it
• Repeated inventory-check failures on the same class are a design finding: the compaction procedure, not the operator, gets the fix
• The ledger feeds the inventory: items that keep proving load-bearing after being classified as residue get promoted — the inventory is versioned and learns

Common Failure Modes

Non-Negotiables

1. The load-bearing inventory exists before compaction does.
2. Force-bearing wording travels verbatim, never paraphrased.
3. Every compaction is validated — presence AND function — before work continues.
4. Effective length is governed by retrieval quality, not the token limit.
5. Handoffs are truth-first and validated on arrival.
6. Every compaction lands in the ledger; the inventory learns.

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