Research and infrastructure from Blocpod
We build and study real agentic execution systems.
Controlled execution is our subject, not our slogan. The products, standards and open-source work below are public so that the claims on this site can be inspected. Each entry says what it proves about how Blocpod systems are built.
- 01
Live product
Agentiquette
That Blocpod studies what makes agents finish the job, not just start it.
The index and intelligence layer for agentic execution systems: skills, instruction files, loops, memory architectures, evaluation and governance. Entries are scored against a public, version-locked methodology. Metric fields are API-sourced and stamped with a fetch date; anything older than 30 days is hidden at render time rather than shown stale. Editorial fields are human-owned and never touched by automation.
- 46 entries indexed, 45 fully scored (as of the live site)
- Framework: prompts tell agents what to do, skills teach them how, loops force them to finish, memory lets them improve, evaluation proves they worked, governance keeps them safe
- Zero-dependency Node, no build step
- 02
Open source / Apache 2.0
Bitcoin AI Provenance
That "the system records what it did" can be a verifiable property instead of a promise.
A reference implementation for Bitcoin-anchored AI provenance receipts and bounded autonomous-agent actions. It answers four questions independently: what was asserted, who signed it, what it depended on, and whether the batch root was committed in a confirmed transaction. The threat model states plainly that Bitcoin proves commitment and ordering, not truth.
- Ed25519 identities, domain-separated signed envelopes, canonical JSON
- Signed agent policies with model, tool, target, daily-spend and action-count limits
- Merkle batching with inclusion proofs; 37-byte commitment payloads
- No third-party runtime dependencies; unit, integration, adversarial, determinism and release gauntlets
- 03
Open source / MIT
ORIGIN
That Blocpod builds verification systems that refuse to appoint an authority, including itself.
A Bitcoin protocol observatory. Evidence over authority. A local-first system for examining Bitcoin-family implementations that keeps three questions separate: what the founding document claimed, what historical software demonstrably did, and what a specific implementation does now. Publishes inputs, methods, outputs, uncertainty, content hashes, signed checkpoints and inclusion proofs.
- v1 release candidate; bundled implementation results are labeled demonstration fixtures, not live audits
- Append-only log verification, Merkle commitments, Ed25519-signed operator checkpoints
- Attestation-mode validation that rejects mutable or unsafe build inputs
- 04
Blocpod system
Meridian
That verified context is an architecture, not a retrieval trick.
A claim-based context operating system. History becomes atomic, provenance-linked claims; claims render into living digests; digests compile into compact, task-specific context packs with hard budgets. Contradictions are quarantined for a human. Approval replaces labor. An edge compiler turns the memory into instruction files for Claude Code, Cursor and Codex, so engineering agents work from the same verified context as the operator.
- Falsifiable thesis stated in the repository: a compiled pack of a few thousand tokens should beat raw history on real tasks, blind-judged; the comparative harness is open work and the README says so
- Status honestly published: research prototype and local runtime, not production personal-data infrastructure
- 05
Engineering practice
Skillmark and the Blocpod skill library
That the agents building Blocpod systems are held to the same standards as the systems.
Blocpod runs its own engineering through governed agent runtimes with published instruction files, versioned skills and evaluation. Skillmark is the scoring engine behind Agentiquette. The rule set that governs Blocpod’s own executive agent, the FounderOS authority matrix, is the same green, yellow, red model we build into client systems.
- Authority classes enforced in code, not prompts
- Untrusted content is data, not instruction; extractors have no action tools
- Agent-generated text is never treated as independent factual evidence
More public repositories at github.com/Blocpod. Systems with verified metrics are documented on the Work page.
Put the research to work
Bring us the bottleneck.
The control model in these projects is the one we build into client systems. If you have a workflow where accountability matters, bring it to a diagnostic.