BUILD / DEFER — THE LEARNING LAYER
Sep 16, 2026
Given small team, limited capital:
ComponentVerdictNoteCommitment ledgerBUILD — v1, tinyA table plus a discipline. This is the experiment.ContextBundle as content-addressed artifactBUILD — v1Already in Round 1's v1. Now load-bearing for replay, not just audit.Calibration storeBUILD — v1Derived views over the commitment ledger. Nearly free.Resolution schedulerBUILD — v2, smallDurable timers surviving model and schema change.Policy as versioned artifactBUILD — v2Cheap now, unbuildable retroactively once behavior is implicit everywhere. Do it before you need it.Value / expected-value accountingDEFER — v3Meaningless until calibration data exists.Counterfactual environmentsWRAP — v2podman/microVM snapshot-fork-reset. Justified as an instrument, not a fabric.ConsolidationDO NOT BUILDOpen research. Integrate whoever solves it. Building this with a small team is how the company dies.Weight-level continual learningIGNORENot your problem, and likely solved above you.
The pattern: build the measurement apparatus, defer the learning mechanism. The ledger and calibration store are cheap, retroactively impossible, and generate the data that tells you whether the expensive part is worth attempting. The consolidation step is the frontier and you should be a consumer of it.
D. OPEN QUESTIONS I NEED ANSWERED
Three things gate a Round 3 that would be worth more than Round 2:
Is "Terra" my seven primitives, or a separate system? If separate, §5 is unearned and I'll redo it against the real design.
Hermes/OpenClaw specifics — persistence model, memory write path, whether anything reads the log. I answered for the archetype; the specifics may already contain a resolution mechanism I dismissed.
Which product are you building this quarter? Round 1's unresolved fork — batch personal computer on a 16GB i5, versus cloud planner with local sovereignty, versus developer-workstation orchestration — still isn't decided, and it determines whether the resolvability experiment above should be run on your work or on code, which is the one domain where verification is cheap and my §6 obsolescence argument is strongest.
That last point is the honest closing note. Coding is the domain where resolution procedures are abundant and nearly free — tests, compilers, type checkers, benchmarks. If you run the experiment there, resolvability will look excellent and you will conclude the thesis is validated. It won't be. It will be validated for the one domain that least needs it, because brute force plus a verifier already works there. Run it on email, documents, and household logistics — the work an actual personal computer does — and take the low number seriously if you get one.
