Key Takeaway
Make what the agent sees and where it sits programmable at compile time: an in-memory registry can cut context cost, recover most task performance, and enforce tool access by removing schemas before the model ever sees them.
ON THIS PAGE
Key Findings
An in-memory, addressable registry of instruction blocks lets you assemble an agent's context by priority and absolute position at each compile step, rather than concatenating a static file. Controlling block position and grouping produces measurable lifts in task success, while a relevance filter applied before placement cuts context size dramatically. Removing a tool schema from the registry stops the model from calling that tool entirely, and registries that reference one another enable inter-agent coordination with zero prompt-token overhead. Blackboard Pattern
Not sure where to start?Get personalized recommendations
Data Highlights
1Relevance gating (filter before placement) reduced prompt token cost by 67.8% while recovering 83% of the promoted-condition success rate.
2Grouping a critical block with its content-adjacent neighbors raised a model's per-position mean pass rate roughly fivefold at the hardest registry size for Mistral-7B-Instruct-v0.3.
3Structural schema eviction produced a 0% tool-invocation rate after eviction versus 100% invocation when the schema was present.
What This Means
Engineers building multi-agent systems and product leads responsible for agent reliability — use these primitives to reduce prompt cost, control which instructions actually influence behavior, and implement hard access controls without relying on model compliance. Researchers and evaluators focused on agent governance and multi-agent trust can use the registry pattern for reproducible provenance and immediate revocation of capabilities. Evaluation-Driven Development (EDDOps)
Key Figures
Fig 1: Figure 1 : Compile-time placement and grouping of the critical block in a thirty-one-block firmware registry. Task success rate is plotted against the absolute position of the critical block. Shaded bands show one standard deviation across three independent task seeds where available. (a) Task query precedes the registry. The near-ceiling region extends to approximately position eleven before the cliff. A residual region of elevated signal appears around positions twenty-two through twenty-six, the same content-adjacency effect as in panel (c) shifted by one slot. The dashed black trace shows a control sweep at registry size thirty-one in which the two output-format blocks have been removed from the seed library and replaced with neutral firmware-domain filler, with the residual signal reduced but not eliminated. Curves are shown for registry sizes one through thirty-one. (b) Cluster sweep with the task query preceding the registry, showing the three-block group placed at front, centre, and back of the registry. (c) Task query follows the registry. The near-ceiling region extends to approximately position six before the cliff, with a residual bump at position twenty-three compatible with a content-adjacency interpretation involving the two output-format blocks in the seed library. (d) Cluster sweep at registry size thirty-one with the task query following the registry, comparing single-block promotion against three-block group promotion in front, centre, and back orderings. The back ordering, in which the critical block follows its two output-format neighbours, sustains success across positions where single-block placement fails.
Fig 2: Figure 2 : Cross-model position-sensitivity profiles, single-mode task-following sweeps. Mean task pass rate is plotted against the absolute position of the critical block. All panels use the same seed file, task pool, and prompt scaffold, with model identity the only variable across panels. Each viridis-coloured line is one registry size from one through thirty-one. Panels are ordered to span the weak-to-strong axis of model performance on this benchmark. (a) Qwen2.5-7B-Instruct, single run. Wide priming bump at absolute positions twenty-two through twenty-seven peaking at ninety percent at position twenty-four for size twenty-six, with front-slot accuracy decaying as registry size grows. (b) Llama-3.1-8B-Instruct, single run. Steep front cliff after a short plateau, compressed bump at positions twenty-three through twenty-five. (c) Mistral-7B-Instruct-v0.3, single run. Catastrophic collapse for sizes ten and above. The dashed black overlay shows cluster-back placement at registry size thirty-one with the three-block group of critical block plus the two priming neighbours moved through the registry as a unit, lifting back-half pass rates from near zero to between thirty and eighty percent and producing a smaller broader lift across the front positions. (d) Qwen3-14B, single run. Near-ceiling across nearly all positions including the priming bump region, with the primacy plateau flat at one across all sixteen registry sizes. The dashed black overlay shows Qwen3-8B at size thirty-one in the cluster-back arrangement under the task-preceding scaffold, plotted at its true absolute position, representing the paper’s entire methodological stack applied to the smaller Qwen3 sibling. The overlay outperforms Qwen3-14B in the mid-registry zone and is outperformed at the back and front. n = 20 n=20 tasks per cell across all conditions.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
The experiments use controlled firmware-style tasks and several instruction-tuned models; results may vary across domains and conversational workloads. The design trades persistent database-backed retrieval for low-latency, in-memory operation, so state is session-local and requires different operational choices for long-term archives. Hardware, quantization, and model family affected magnitude of effects; run diagnostic sweeps on a representative weaker model before deploying cluster arrangements to production models. Continuous Monitoring
Full Analysis
RAMPART replaces static instruction files with an in-memory ordered registry of addressable blocks. Each block contains content, a runtime UUID, provenance tags, and a lightweight embedding computed at cold start. At every compile step the system filters blocks by relevance, orders them by explicit priority, and assembles the final string the model sees. Because the registry is session-local and stays in RAM, there is no disk I/O after startup and compile-time assembly is deterministic and inspectable before the model's first token.
Controlled experiments swept the absolute position of a critical instruction block across registry sizes and models. Position and adjacency matter: promoting a critical block to the front or grouping it with content-adjacent neighbors can recover large amounts of lost performance. Pre-promotion relevance gating cut prompt token cost by about two-thirds while retaining most of the performance gain. Evicting a tool schema block before compilation eliminated tool calls entirely, demonstrating a structural access control stronger than instruction-based bans. When registries hold references to one another, agents can exchange constraints and proposals by writing blocks — a method call with zero prompt tokens — enabling low-overhead multi-agent coordination with full provenance and instant revocation of capabilities. Market-Based Coordination Pattern Zero-shot learning
Explore evaluation patternsSee how to apply these findings
Credibility Assessment:
Single author with low h-index (6), no listed institutional affiliation, arXiv preprint with zero citations — limited reputation signals. Fits emerging/limited-info category.