Decisions and ideas,
with their sources.

Phase 0 wraps every portfolio project in a root orchestrator that hands bounded slices of work to specialist agents and worker agents. Twenty-two agent records live under ~/.minimax/agents; each declares its own capability matrix (read-only workspace, scoped write paths, no file writes outside its boundary, sequential calls not parallel). Deploy-watch cron runs after every main push, validates the Azure Static Web Apps check run, and verifies the live URL before deleting itself. Private nxt, stk, and inf projects remain off-limits across the team.

Before deployment, claims should be tied to a repeatable evaluation, provenance, known limits, rollback conditions, and a named human decision owner.

Context determines what the system knows, security determines what it may do, and evaluation determines whether the combined behavior is acceptable. Treating any one as an add-on leaves a blind spot.

Choose local execution when data control, offline operation, or predictable ownership dominates. Choose cloud execution when elastic capacity, managed services, or faster experimentation dominates. Measure the workload before choosing either.

The path runs from learned perception and prediction to explicit simulation boundaries, reproducible scenarios, measurable residuals, and finally embodied action under operator review.

Move from recommendation to supervised action, then constrained autonomy, only when evidence supports each transition. Preserve a clear human stop condition, audit trail, and rollback path at every rung.

AIA compares AI products and developer ecosystems, GPU maps kernel execution, POL covers programming-language foundations, and USL covers adaptation and fine-tuning. Each atlas keeps its own research cutoff, source list and verification date, and none of them claims a benchmark result or a universal ranking. Reading them in this order keeps ecosystem claims, hardware behaviour and training choices distinguishable.

Research applications keep companion laboratories: GEX for kernel execution, TFL for token serving, ADP for adaptation and fine-tuning estimates, WML for world models, HEX for humanoid kinematics, ARL for one agent run, DPL for decision routing, CUL for computer use, and MEM for memory lifecycle. Each one is a browser-side educational model: no real model is trained, no live hardware is measured, no field data is read and no neighbouring application is integrated. AOS is a target-architecture guide, not a running operating system, so its publication shows a design and not a shipped runtime.

PDT introduces pump anatomy and synthetic condition signals, DTR compares a prediction with the observed outcome under explicit human approval, and DCL compares local and cloud deployment choices against declared memory, privacy and cost assumptions. All three are browser-only learning tools with synthetic inputs: no field telemetry, no live price feed, no machine control and no universal deployment winner. Human approval in DTR is a recorded step inside the simulation, not an operational interlock.

SWI frames collective-intelligence research and exports experiment plans. ANT and BEE reproduce pheromone trails, waggle-dance recruitment and collective foraging as seeded simulations that can be replayed. Their results are model observations: they do not measure a real colony, and they do not prove that an SWI agent recipe works until it is evaluated against an explicit contract.