Ingested from ActiveGraph (Nakajima, "The Log is the Agent",
arxiv 2605.21997, May 2026 — v1.0 May 2026, MIT/Apache). The
focused web search across 5 themes (event-sourced agent runtimes,
deterministic replay + content-addressed caching, multimodal data
primitives, context compression, durable workflows) confirmed
ActiveGraph is the uncontested 2026 SOTA at the event-sourced-
substrate layer. The other four themes have alternatives but
none with the same level of articulation; this wave ingests the
typed primitive JouleClaw's receipts + cascade had been pointing
at but didn't name.
1. jouleclaw-event-log (new crate) — typed substrate.
- EventKind { Goal | Object | Relation { from, to, edge_kind } |
LlmReq { request_hash, prompt_summary } | LlmResp {
request_hash, response_summary, tokens } | ToolReq { tool_id,
args_hash } | ToolResp { tool_id, args_hash, response_summary }
| Patch | Receipt }. The Relation variant uses `edge_kind`
(not `kind`) to avoid the silent serde-tag collision bug.
- Event { prev_hash, ts_ns, kind, event_hash } sealed at
append. EventLog::append computes prev_hash from the
previous event's hash (genesis = BLAKE3 of zero bytes for
the first event).
- EventLog::verify_chain catches tampering at the first
break with the exact index.
- replay() — deterministic fold producing a ProjectedGraph
(goal + typed objects + typed relations + receipts), each
element annotated with the event_hash that produced it.
- replay_strict() — compares a fresh fold to an expected
projection byte-for-byte under canonical JSON; returns
ReplayError::Divergence with the diverging content hashes.
A green strict replay is a proof the run is reproducible
(the published ActiveGraph claim).
- ContentAddressedCache — populates from the log's LlmResp +
ToolResp events, indexed by the documented 2026 SOTA
convention: LLM key = BLAKE3 over (system, user_messages,
model_id, tool_defs, output_schema); tool key = (tool_id,
args_hash). On replay, a behavior firing with a recorded
hash hits cache; no new model call.
- lineage_of(log, object_id) — given an artifact, return every
event in its causal chain (goal that motivated it, LlmReq/
LlmResp pair that produced it, ToolReq/ToolResp if any,
Object event itself, Relations referencing it after creation).
- 13 tests.
2. jouleclaw-jcr1::event_log — JCR-1 event-log/1 schema.
- EventLogReceipt { jc_receipt, id, closed_at, steward, log_id,
head_event_hash, chain_depth, counts: EventLogCounts,
total_tokens }.
- 2 conformance vectors with embedded receipt_hash.
- 3 jcr1 event_log tests.
3. Cross-language conformance — Rust + Python + Go verifiers all
shipped, byte-equal across the three implementations. The
three-impl pattern (matching JCR-1 v1) is preserved.
4. Demo binary — examples/wave54_event_log_replay.rs in
jouleclaw-stack. Builds a 9-event agent run (goal → LLM
exchange → tool exchange → object → relation → inline receipt),
demonstrates: (1) replay determinism; (2) strict-mode rejection
of a tampered projection with divergence pinned to a content
hash; (3) cache-served replay of the recorded LLM request (no
new model call); (4) lineage trace recovering 7 events in the
causal chain back to the goal; (5) event-log/1 JCR-1 receipt
sealed over the log snapshot.
The triangle (conservation + cone + oracle + elicitor) stays the
type system; wave-54 adds the event-sourced substrate that any
JouleClaw agent runtime can be built ON. The log IS the agent.
Workspace check clean.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
49175df95e