SABERISTIC / DAILY DEVELOPMENT
Practical notes from building Saberistic products: the architecture I chose, the code that survived, the failures that changed the design, and what is still only a plan.
Follow via RSSSTAGE 1 / MILESTONES 15–17
Restart-safe continuation, a kernel-backed task runner, and signed authorship evidence: Harness M15–M17 connect safe checkpoints to verifiable patches, with 724 offline tests and explicit live gates.
Read the build noteSTAGE 1 / MILESTONES 11–14
Five bounded tools, Docker-only model mutations, FIFO steering, context compaction and durable replay—four Harness milestones backed by 710 offline tests and explicit recovery limits.
Read the build notePRODUCT BUILD / EVIDENCE-FIRST LICENSING
Turning a checkout migration into an evidence-first licensing platform: structured research, on-demand guides, fifteen tools, safer customer records and a carefully staged rollout.
Read the build noteLOCAL RESEARCH / GRAPH-ASSISTED Q&A
Building a continuous research notebook with Brave Answers, local Qwen, SQLite, hybrid search, evidence graphs and exact passage citations—then making the collected research answerable.
Read the build noteSTAGE 1 / MILESTONES 9–10
How Harness turns M8's Workspace capability into an explicit trusted local adapter and a Docker-by-default disposable adapter—with bounded text worktrees, no host mounts, and auditable cleanup.
Read the build noteFIELD NOTE / FIRST REAL MODEL
How a local Qwen model crossed the Harness tool loop—and exposed guarded-array, timeout, policy, token-budget, sandbox, and evidence boundaries before the work became a release.
Read the build noteSTAGE 1 / MILESTONE 8
How M8 moves filesystem and process authority behind one injected Workspace contract, restricts tools to reviewed operations, and blocks direct host access without shipping an adapter early.
Read the build noteSTAGE 1 / MILESTONE 7
How M7 adds a deterministic multi-round session loop with versioned context, strict tool admission, durable policy-before-effect ordering, hard budgets, cancellation, and timeouts.
Read the build noteSTAGE 1 / MILESTONE 6
How M6 adds a one-request agent runtime with streaming model compatibility, append-before-yield events, typed steering and cancellation, and explicit self-hosting limits.
Read the build noteSTAGE 1 / MILESTONE 5
How M5 audited the evidence for a second runtime, retained TypeScript/Node without claiming it was optimal, and used the same record to harden the manifest-to-report exit gate.
Read the build noteSTAGE 1 / MILESTONE 4
How M4 adds fenced scheduling, replay-safe sessions, content-addressed S3 artifacts, deterministic audit export with a transactional registry/checkpoint commit, and fail-closed Kubernetes topology.
Read the build noteDAY 010 / GROWTH PROGRAM
How Growth v2 expanded into sensor-backed custody receipts and moved from a local-validator proof to an evidence-pinned experimental Solana devnet program and published IDL.
Read the build noteSTAGE 1 / MILESTONE 3
How M3 wraps the Harness kernel in a typed WebSocket service with correlated approvals, a manifest-derived Docker plan, an OpenAI-compatible adapter, and explicit live-operation gates.
Read the build noteDAY 008 / SPIRAL SAFE
How I reconciled eight repositories into a WebAuthn-authorized signing platform with a Vault boundary, scoped API keys, usage accounting, four demos, Raft deployment intent, and explicit Nitro gates.
Read the build noteSTAGE 1 / MILESTONE 2
How Stage 1, Milestone 2 adds a golden HTTP repository, SDK-owned scenarios, a read-only task board, opt-in OpenTelemetry, and a hardened live MCP stdio client.
Read the build noteDAY 006 / GROWTH PROGRAM
How I turned a live-but-unsafe Solana growth contract into a hardened v2, an evidence-led website, and safety-gated browser-local and local-validator demos without mutating legacy accounts.
Read the build noteDAY 005 / CRYPTOPAL
How I turned a 2022 PlantUML sketch into a local Solana demo that moves one cUSD from wallet to email to wallet through two independent blinded-token hops.
Read the build noteDAY 004 / THREE LOVABLE PROTOTYPES
How The Last Press, Psych Lab, and Borrowed Brain turn game state, questionnaire authoring, and decision support into deployed products—and what their audits still block.
Read the build noteSTAGE 1 / MILESTONE 1
How Stage 1, Milestone 1 turned the Harness Platform foundation into an operator loop with durable events, offline evals, policy compilation, a terminal viewer, and a CI exit gate.
Read the build noteDAY 002 / TURBOPASS
How I rebuilt a Privacy Pass-style token server in Rust, targeted its documented API and storage contracts, and moved issuer-key rotation into durable Temporal workflows.
Read the build noteDAY 001 / HARNESS FROM SCRATCH
How I used Pi and a local Qwen model to bootstrap a TypeScript agent harness whose own task contract can block unsafe changes and produce evidence.
Read the build note