CAIN-42 MCPGateAGENTIC AI 75 · #15
Retained reasoning and context compaction
OpenAI reported substantial ARC-AGI-3 improvements from preserving useful reasoning and compacting history.
Last reviewed 2026-11-11
Agentic AI 75 / #15
2026 status: Benchmark-specific research result · Category: Context management
The 2026 development
OpenAI reported substantial ARC-AGI-3 improvements from preserving useful reasoning and compacting history.
The governance gap
Compaction is lossy by design. If the summary drops the authority, the constraint or the failed attempt, the next step can repeat a prohibited action with a clean-looking context.
Where CAIN-42 fits: Moat 4 — Execution Provenance & Attestation
Core question: Can we prove what actually happened? Hash-chained, signed records binding decision to execution to observable effect.
Treat the compacted context as a derived artifact with a link to what was dropped. The record keeps the original trajectory; the model gets a summary; a reviewer can always walk back.
How the wiring works
- Hash the original context window before compaction.
- Mark the summary as derived and list the retained constraints and dropped spans.
- Keep authority and prohibition statements outside the lossy summary.
- Rehydrate the original when a decision depends on a dropped span.
What is real today
LIBRARY Production code and tests exist in this repository; there is no hosted route for this exact use.
Context fingerprinting and continuity code exist. No ARC-AGI-3 benchmark or compaction-quality claim is made here.
| Component in this repository | What the module says it is |
|---|---|
cain/runtime/fingerprint.py | CAIN Sovereign — Agent Behavioural Fingerprint (Prompt 7, Phase 10). |
cain/continuum/memory_continuity.py | CAIN CONTINUUM — Authority-Preserving Memory + continuity proof + consolidation attack lab |
cain/argus/memory_firewall.py | CAIN ARGUS — memory-to-authority firewall (Prompt 14, section 12). |
Descriptions are the first line of each module's own docstring, read without importing it. A module path is a pointer into this repository, not a hosted endpoint. Its existence is checked by the test suite; its status for this use is the label above.
Check it yourself
- Run the live memory-sanitisation demo
/memguard-keystone - Fire a live decision through the CAIN-42 pipeline
/fabric/try— the same enforcement path every product sits behind.
Commercial package
| Offer | Context Integrity Review |
|---|---|
| Buyer | Model providers and agent-platform teams |
| Price (proposal) | $3,500–$7,500 per review |
| Bundle | Memory & Context — Memory that cannot launder authority. |
| Operator stream | Stream 2 of the 10-stream operator playbook |
Agent-platform teams, enterprise AI governance, regulated data owners is the bundle buyer. The entry offer for this bundle is Memory authority audit; hosted memory-sanitisation pilot; the range is $2,500 per audit; metered from $0.0002 per inspection. A price on this page is a proposal, not a validated rate, and no customer is claimed.
Buy this as a product or service
CAIN-A75-015 — Context Integrity Review DELIVERABLE
A bounded engine run on the buyer's sample: findings, a written report, and a SHA-256 artifact manifest. The manifest is unsigned: it is not a signature and not an attestation.
Kind: Assessment · Price (proposal): $3,500–$7,500 per review · Lead time: 5 business days.
Open the product page · all 75 products
What this page does not claim
What this page is not. CAIN-42 is not affiliated with, endorsed by, integrated into, or shipped with the organisation named in the 2026 item. The item is public context; the page describes how CAIN-42 would govern that class of capability. No page here claims a customer, a partnership, a certification, a patent, or a production integration. Prices are proposals. Built by a three-person team.