Rahul Krishna

AI you can leave alone: every decision it makes is provable, and every irreversible step waits for your yes.

That is the thesis behind everything on this account. Agents are useful when you can hand them real work, close the laptop, and trust what you find in the morning: not because the agent promised, but because it left evidence.

Proofward: the seven-tool family

Proofward is the home of a family of seven tools. Each one owns a single question an unattended agent must answer with proof: which capability may run, under what budget, delegated to whom, designed how, published how, verified how.

Repo	What it does	Status
ChoiceGate	Picks the single best capability for a task (skill, tool, MCP server, integration), installed or not: materially better options surface as owner-approved choices, browser or manual work is the last resort. Deterministic receipts, refusals with named reasons, never installs or runs anything itself.
RunSteward	Hand an agent work too big for one sitting and walk away: queues with spending caps, checkpoints, resume after limit resets, and a reviewable morning report. Bundles Claude Carry, the overnight runner.
Cairnspan	One local agent hands a bounded task to another (Claude Code and Codex through their own logged-in clients, no shared keys), with fail-closed launchers and a receipt for every run.
StormWorthy	Research that argues with itself: an adversarial refuter, typed claims verified against their cited sources, and honest abstention when the evidence is thin.
Rigwright	Authors and tests the small skill files agents run on: one measurable outcome per skill, its own evals, and packaging generated for both Claude Code and Codex.
PixelHelm	Agent UI design with hard gates: generate, render in a real browser, judge against accessibility and an honesty floor (unknown data shows as unknown), repair, repeat.
ReleaseBench	Private repo to public without a public mistake: audit, secret scan, governance files, README polish, release prep, every step stopping at a reviewable result.

How they compose

Rigwright authors the small skill files agents run on. ChoiceGate picks which capability runs for a given task; it binds to an owner-accepted capability registry rather than managing capabilities itself. RunSteward governs long unattended runs and ingests ChoiceGate's receipts. Cairnspan is the transport when a bounded task has to cross from one agent to another. StormWorthy verifies claims against cited sources, and PixelHelm vendors its engine to judge design briefs. ReleaseBench is the exit door: the checked path from private work to a public repo.

flowchart LR
    RW[Rigwright<br>authors skills] --> CG[ChoiceGate<br>picks what runs]
    REG[owner-accepted<br>capability registry] -. binds to .-> CG
    CG -- receipts --> RS[RunSteward<br>governs long runs]
    RS -- tasks that cross agents --> CS[Cairnspan<br>cross-agent transport]
    SW[StormWorthy<br>verifies claims] -. verification engine .-> PH[PixelHelm<br>judges design briefs]
    RS --> RB[ReleaseBench<br>exit door to public]

Start here: StormWorthy and Cairnspan.

All seven are MIT licensed. Each README states exactly what is verified and what is not.

Independent products

Two products stand on their own, outside the Proofward family, built with the same rule: show the evidence or say nothing.

• CohortWatch (with its web app): open, cohort-relative momentum tracking for accelerator startups, scored from public evidence with confidence and limits shown beside every reading. Live at cohortwatch.dev.
• Laptop Deal Watcher: an accuracy-first laptop price watcher that prefers an honest "no price" over a plausible guess; a deal becomes a BUY only after model, seller, condition, plausibility, and corroboration checks all pass.