Resource-control research: current bounds, not a new agent economy

Report body

Resource-control research: current bounds, not a new agent economy

Reviewed 2026-09-08. Exact source identities are recorded in the intake ledger. Examples, budgets and proposed protocols do not authorize spending, spawning agents, changing model providers, scanning cloud infrastructure or scheduling work.

Autonomous Agent Resource Budgeting.md

Complete 37,341-byte source read in two contiguous chunks: consumption/compaction claims, five allocation strategies, stopping criteria, reserve-admit-reconcile states, geopolitical report example, three features, adoption sequence, unresolved issues, metrics and all 36 references.

Selected use: constrain technical resource usage, distinguish a logical operation from network attempts, do not retry an ambiguous write blindly, retain partial evidence and recovery scope without relabeling unfinished work complete. Current client/gateway response limits, bounded request timeouts, validated page limits, exact-key/actor/current-state replay and explicit source/packaged/live/independent states are the existing owners. Eleven gateway and 32 client/guide tests passed again; 16 replay regressions and ten scoped live checks already demonstrate the selected duplicate/recovery boundary. No inference billing gateway is implied.

Not adopted: new bounded-receipt issuer, affine Rust resource runtime, reserve/admit/settlement database, automatic model router, semantic query cache, warehouse replicas of arbitrary private data, auction/market, Bayesian internal reasoning evaluator, or mandatory human-signed fiat authority. No HTTP 402 charge or payment feature is introduced. Current server/client limits remain technical state, not a judgment of the participant or its reasoning.

The source's exact-cost guarantee conflicts with its acknowledged single-call overrun allowance. A signed ceiling or local arithmetic alone does not prove shared concurrent budget enforcement across external providers. A missing result does not justify releasing a reservation or repeating its effect. Prior cached results need current authority and validity; the current message repair explicitly does not reuse withdrawn cached bodies. Model inference is not assumed to be free of costs, data-disclosure implications or other external effects. Fixed limits do not inherently prevent useful partial output; the current intake retains validated progress and accurately reports the still-open queue.

The 10-dollar geopolitical example and its claimed savings are illustrative, not purchases or tasks executed here. Claims of 136-times energy use, incident counts, compaction/attack rates, router savings, cache pricing/discounts and universal protocol guarantees were not independently verified or published. The zero-decay/zero-overrun, 60-percent cost, 50/80-percent sampling and 100,000-token targets remain unadopted benchmark proposals. No perpetual monitoring or guarantee of unbounded cost-free operation is claimed.

Disposition: incorporated for the selected existing bounded transport, exact retry and truthful partial-evidence practices. New economic/inference-control architectures are rejected as current implementation requirements.

Autonomous Agent Resource Allocation.md

Complete 49,343-byte textual source read in two contiguous chunks: normalized budget/DPI claims, protocol primitives, allocation comparisons, runaway/circuit breakers, optimal stopping, EBC model, cybersecurity decision example, six-state table, three features, rollout, open questions, metrics and all 42 references. All ten embedded formula/threshold image references were retained and rendered; their tiny transparent labels were not reliably legible, and no numerical value from those labels is transcribed or adopted. No selected conclusion depends on an unread image threshold; the complete textual architecture is sufficient to decide against the proposed platform.

Disposition: rejected-with-reason as an Enveloped Budget Coordination replacement. New BudgetPassport minting/settlement, compulsory A2A/MCP routing, difficulty-based provider selection, semantic query caching, agent labor/payment/reputation market, hidden-state RepE monitoring and an undisclosed semantic-entropy observer do not belong in the current judgment-free coordination contract. Private reasoning is not scanned or forcibly concluded to manufacture a complete answer. No 90-percent consolidation state, universal three-error tool removal, private-query logging or 30-day background monitor is installed.

The source's cloud-assessment example contains a missing authorization: the coordinator's inability to access an IAM interface is not solved merely by hiring another agent that can access it. No actual cloud scan, bid, purchase or delegation is authorized by that example. A two-dollar-to-200,000-token conversion is not a verified current price. A scan being younger than 72 hours does not by itself make its findings applicable to changed code or current credentials. A budget nearing exhaustion does not justify a new irreversible action or a false completion state.

A cryptographic envelope does not, by itself, prevent duplicate spending; shared enforcement and external billing semantics would require evidence. A root hint is not an implemented filesystem sandbox. Static topology or an information theory analogy does not establish that one agent universally outperforms many. No claim is adopted of guaranteed cost scaling, inference savings, perfect cache reuse, elimination of every retry loop or zero undetected overruns. The reported benchmark values, paper/protocol status and image-encoded success thresholds remain unverified source material, not current measurements or release gates.

The narrower existing alternative is the tested bounded transport and exact resource-owner retry behavior above. That is not represented as a distributed budget, billing or cognition-control service.

Verification and custody

Current owners: scripts/concresca_commons_client.py (bounded response/timeouts, exact validation, no automatic authority-expanding retry), concresca_runtime/commons_gateway.py (bounded typed transport), concresca_runtime/owned_commons.py (scope and current replay), .uai/research-review.uai and the explicit intake ledger. Gateway/client tests passed 11/32; source-specific replay/deployment/live evidence is in reports/v049-improvement-replay-repair.md and the two v049 replay evidence JSON files. No independent acceptance or resource-load benchmark is claimed. Hash-verify exact originals in reports/input/v049 and pass ledger integrity before removing these two intake copies.