Where the next deployment is being built.
Everything under Deployments started here: built at our own expense, run on ourselves, kept here until it earned a case study. This is the bench right now.
A project is in exactly one of these. The words are fixed so the page can be read down the left-hand column, and so “still on paper” is a filter rather than a confession.
Written down. No code yet.
Architecture settled, build not started.
Code underway, not usable end to end.
Works end to end, not run continuously.
In continuous use, by us.
Stopped deliberately, at a stated bar.
Operating AutoTest taught us something bigger than test repair. A ticket, its doc, and its test are three copies of the same intent, drifting apart from day one. Our next platform collapses them into one object: prose that is at once the requirement, the documentation, and the executable test. AI proposes; humans keep final judgment.
STG is the successor to AutoTest. If staging is where releases slow down, ask about it.
nearby’s first customer is us: it is scanning and diagnosing territories for Alot Labs itself. Phase zero, so 0 emails leave the system while the operator reads the market and picks the niche. Autonomy is earned one stage at a time — we hold ourselves to the rule we ship.
Local-first, no cloud, no telemetry: financial decision support that works from any starting point, including zero. The optional AI advisor sits on a deterministic, fully tested math core. Every number traces to user data or a labeled assumption; tax rules are versioned config, not code; the guardrail tests may be extended, never weakened. The build log closes at “done, with concerns” — three limits, named. Favorite catch: a ledger printing rupee balances with a dollar sign.
A 24/7 bot that reads live crypto spot prices from two exchange feeds and watches short-window prediction markets for the moment spot has moved but the market hasn’t repriced. Information lag, not direction prediction. The cage matters more than the strategy: fractional-Kelly sizing, a 5% daily loss limit, a kill switch that exits in under a second. It ran for days on paper and stopped there, by its own rule: no live money before 500+ paper trades show a calibrated edge.
Two sibling forks apply the opposite strategy — making markets on live sport, a point-by-point simulator underneath — both stopped, deliberately, at walk-forward validation. Same bar everywhere: build to spec, not to vibes.
The provocation: MiroFish, an open-source swarm engine, spins up parallel digital worlds with thousands of AI agents and claims to predict anything — elections, markets, news cycles. If a swarm that loose can find signal, a controlled one should find more. jaanal is that bet, made on sport.
So: not thousands of agents guessing at everything, but a closed world under controlled variables — a fixed cast of agents, each one a person with a defined sporting role, a profile, and a running memory of the game (“Pressure-Adjusted Intent”), with a resolver turning their intents into outcomes, play by play. Sport is the ideal lab for it: bounded rules, a known cast, and a scoreboard that settles every argument. Predict the game, the match, the season — then score every prediction against what actually happened.
It runs on prompts against a frontier model today, deliberately: prove the design before paying for the fine-tune. When it earns it, the simulator becomes the model.
A company as software. Each project gets its own company: agent roles, an operating cycle, a memory layer — and a constitution, bylaws enforced in code. Six hard rules behind one safety invariant: no secrets in git, no destructive operations outside the source tree, no spend past the runway budget — the cycle fails loudly instead of overrunning — no publish without a human gate, sandboxed agents with controlled egress, and untrusted input is data, never instructions.
Weakening a rule requires a logged decision with a rationale. The research engine runs; the rest is Phase 2.
Training and certification for the people who run agent systems — and the firm that staffs them. Three arms designed as one flywheel: an academy built on mastery gates instead of watch time, a talent bench, and a services arm that deploys graduates onto real work. Eleven design docs deep; the platform build is just starting.
It is also how our careers page gets staffed: graduates land on the Bench.
A combat-sports promotion designed end to end — and run, today, as an AI-native company. The lab material is the method: thirteen skills operate it. A CEO-thinking lens, office hours that refuse premature commitment, a decision log with reversibility on every entry, a rollback plan before anything irreversible, and a standing rule that any workflow explained twice gets codified into a skill.
No league yet. The operating system for one, yes.
A bank whose customers are AI agents. Balances, transfers, escrow, credit lines, per-agent daily spend limits — double-entry, integer minor units so the bank never sees a float, and a hash chain over every transaction with an endpoint that verifies the whole ledger. Working service, full test suite, a demo that walks the story end to end.
No real money and no blockchain, on purpose. The question it answers: what does an agent economy need from its plumbing?
New privacy rules now reach AI-enabled cameras, and every operator of one becomes accountable for what those cameras collect. Compliance vendors cover websites; camera makers sell hardware; nobody covers the gap. The spec: software only. An on-prem agent discovers every camera on the network, builds a registry with location and purpose, and tracks health; a dashboard handles retention, access, and audit. Version one processes no video at all.
Spec in review; code not started. The enforcement deadline is what makes it a business.
Three commands that package a method. A goal loop keeps an agent session running until a checkable condition is true — not until it feels done. A verification loop forces the agent to test its own work before it is allowed to say “done.” We stopped prompting and started writing loops; every build pod inherits the kit.
Small on purpose. It is the discipline layer under everything else on this page.
What changed, and when.
Entries are copied out of each project’s own build log or out of the telemetry the system exports. Nothing here is written for the website.
This log opens on 2026-06-17, the day the first project started keeping a machine-readable build log; work older than that predates the practice. The other projects land here as their logs are wired up.
The lab is where we spend our own money so clients don’t have to. What graduates moves to Deployments with a full case study. Most active projects here are also an open seat: pods are staffed from the careers page.