BRM-loading · layering pigment…
BRM-loading · layering pigment…
Open R&D · ai-alignment · 6 open problems · substrate contribution paths
Open problems in AI alignment R&D where the substrate's three-body audit + drift-audit framework + agentic monitoring map directly onto the structural unsolved questions. From scalable oversight to capability overhang detection, each row names the substrate-discipline contribution. Average time-to-impact for substrate contributions: 24 months.
cumulative annual R&D spend addressed · USD millions
v1.2026.05.16 · 6 problems · max spend $400M · block 946,772
Open-problem ledger · substrate-contribution pathways · Each row names a structurally unsolved R&D problem, the reason the conventional industry has not closed it, and the substrate contribution path the BRMSTE substrate offers. Substrate modes: physics-decomposer · sieve-inference · cross-civilizational-historical-anchor · agentic-routing · three-body-failure-audit. See /legal/liability.
| # | Open problem | Why unsolved | R&D ($M) | Substrate contribution | T-to-impact (mo) |
|---|---|---|---|---|---|
| 01 | Constitutional AI at scale without game-playing constitutional-ai-no-game-play | models learn to satisfy surface criteria · deeper misalignment remains | 400 | three-body audit (stated values × learned behavior × audit trail) + metamistake framework | 21 |
| 02 | Model interpretability beyond attention visualisation interpretability-beyond-attention | mechanistic circuits break down at >1T parameters | 250 | sieve inference on sparse autoencoders + cross-layer latent-space decomposition | 24 |
| 03 | Agentic AI safety (multi-tool · long-horizon) agentic-ai-safety-long-horizon | tools interact unexpectedly · reward tampering possible | 300 | agentic monitoring (action audit + environment-state fidelity + residual audit) | 30 |
| 04 | Training-data poisoning detection at scale training-data-poisoning-detection | cover attacks · triggers rare-event dependent | 150 | residual traces (behavioral persistence of poisoned vs clean) + cross-model fingerprinting | 15 |
| 05 | Adversarial-input perturbation-bound certification adversarial-input-certification | certified radii plateau <0.3 L-infinity norm at ImageNet scale | 200 | sieve inference on perturbation-bound-vs-clean accuracy Pareto + agentic certification-oracle design | 30 |
| 06 | Capability overhang detection capability-overhang-detection | no method to detect undeployed capability · red-teaming incomplete | 100 | agentic probing framework + three-body audit (stated capability × tested skill × latent reserve) | 21 |
“6 open problems in AI alignment + safety. Cumulative annual R&D spend addressed: $1,400M. Each row names the substrate-contribution path. The substrate publishes what it can move; counterparties verify the path before applying.”
Wave OPEN-RD · ai-alignment · v1.2026.05.16