research-tree

Standing goal: keep one current research answer, built bottom-up as a tree of findings, and re-synthesize only the branch whose source actually moved.

One-line scenario: a research agent built its answer as a tree (raw sources → findings → per-sub-question syntheses → one root answer); revise a single finding three levels down and only its ancestor path re-synthesizes, leaving the rest of the tree dark.

This is the example for structural recursion / propagation UP a recursive tree with per-branch memoization. The marquee frame: revise Finding B2 and only B2 → Synthesis B → Root lights; the seven sibling findings and Synthesis A & C stay skipped moved[] fresh 0. The lit path is bounded by tree depth, never tree size.

DAG sketch

                         Sources Gateway            (entry · external-driven)
                         one facet PER leaf
        leaf:A1 …             leaf:B2 …            leaf:C2
           │                    │                    │
   Finding A1  A2  A3    Finding B1  B2  B3    Finding C1  C2    (8 leaves)
        \    |    /          \    |    /          \    |  /
       Synthesis: A         Synthesis: B         Synthesis: C    (3 sub-syntheses)
            \                     │                    /
             \____________  Root Synthesis  __________/          (the apex)

Edges point UP: leaf → sub-synthesis → root. The gateway exposes one independent facet per leaf, so revising one leaf's source moves exactly one leaf:<id> facet and wakes exactly one finding, whose change propagates up only its branch.

• src/sources-gateway.prose.md: the entry gateway; ### Continuity: external-driven; projects the corpus into one leaf:<id> facet per finding.
• src/finding.prose.md: a leaf; subscribes to ONLY its own leaf:<id> facet; a corrupt excerpt fails the leaf and propagates nothing.
• src/sub-synthesis.prose.md: an interior node; fans in from its own sub-question's findings only (convergent fan-in wakes it once per drain).
• src/root-synthesis.prose.md: the apex; fans in from the three sub-syntheses; the heaviest node and the dominant fresh tick.

Conformance expectations

A conforming harness proves quiet zero-fresh replay, one-leaf changes that wake only the leaf's ancestor path, dark sibling branches, contained corrupt-leaf failure, chain verification, and byte-identical regeneration.

A run leaves a state on disk that any conforming harness can replay keyless:

the bottom-up cold boot, the quiet skips, then the hero: revise one finding
and watch only its ancestor path re-synthesize.

What to try

• Revise one leaf's source rev: only that finding → its sub-synthesis → the root re-render. The other branches skip at fresh 0.
• Revise a leaf under a different sub-question: a different path lights, the same root re-synthesizes.
• Corrupt a leaf's excerpt: the finding fails, carries zero fresh, and wakes no ancestor; the prior answer stands.

The state-dir a run produces

src/ ships the 4 authored contracts and mounts 4 nodes on its own; the 13-node / 20-edge topology is what the reference harness's per-entity expansion produces from them (see the Harnesses section of the repository README). Below is one harness's replay layout; the skill's native layout is state/world-model/{node}/ with a per-node receipts.jsonl (state/filesystem.md).

<state-dir>/
  compile/topology.json   # the TopologyWorldModel (13 nodes, 20 edges, single
                          # entry gateway, acyclic): MANDATORY for replay
  compile/labels.json     # nodeId -> friendly label
  receipts.json           # the flat, chain-verifiable ledger trail
  world-models/<hexNodeId>/published.json + versions/sha256_*.bin
  beats.json              # the scripted beat timeline (cold -> quiet -> surprise)