01
Assign research content to model roles
Ocean moves major tree operators from Python-created placeholder content to specialized model roles. Field splitting, kernel mapping, anchor transfer, mechanism construction, experiment design and refutation return research material; the runtime requests, wraps, stores and advances it. The central change is responsibility for content, beyond adding role names.
02
Hold an unresolved envelope at its stage
When field resolution is missing, the system saves a draft and pending_subagent and remains in phase_0 for later fulfillment. A formal envelope includes contribution, baselines, boundaries and budget. However, a low-confidence response can set paused_envelope_undecidable while advancing the phase pointer to the frontier stage, a discrepancy later addressed by Stars.
03
Model-output contract for operators
apply_operator maps operations to roles, consumes nodes, children or a single claim-bearing response, and wraps nodes with ancestry and audit fields. With no bridge or an unresolved call it returns an empty list instead of placeholder children. Missing model fields remain empty for validation, separating research content from its structural shell.
04
Staged mathematics–physics requests
Mathematical search, physical interpretation and mapping construction receive separate role responses under a conditional bridge route. The design links mathematical objects, physical meaning and a research mechanism while retaining criticisms and responses. It makes cross-domain mappings discussable, but a plausible mapping without real retrieval and adequate refutation is not a scientific discovery.
05
A remaining literature fallback
When literature cartography is unresolved, this snapshot still constructs a local frontier, marks it unresolved and proceeds. That differs from the operator path’s refusal to create placeholder children. Ocean changes how real content enters the tree, while failure handling across phases remains inconsistent and missing literature still affects downstream novelty judgments.
06
Inputs, state and outputs
Inputs require an envelope, budget and a bridge backend capable of role responses. Nodes retain content, evidence, ancestry and operation provenance; run state retains pending roles. Tree, ledger and review artifacts remain the handoff. Producing no nodes may mean an unfulfilled call or validation failure rather than an exhausted research space.
07
A content path beyond Hill
Hill introduces the bridge and semantic gates; Ocean connects the bridge to expansion and mathematics–physics work. Python shifts from constructing research answers to wrapping model answers. Phase advancement and literature fallback still need later repairs, so the history should not imply monotonic accumulation of every capability.
08
Read the responsibilities in an expansion
An expansion has three responsibilities: policy selects where to continue, a role answers the requested question, and the runtime stores that answer as an object. Ocean separates the second responsibility from placeholder construction. Comparing role, parent input and response reveals whether a mechanism actually addresses the question and whether missing children reflect absent responses or failed validation.
The same reading applies to the bridge: a mathematical object supplies analyzable structure, physical interpretation relates it to the target system, and a mapping mechanism needs a falsifiable prediction. Having all three names does not establish the chain; responses and criticism must be read. Ocean creates interfaces for that scrutiny without yet enforcing consistent fulfillment across the evidence chain.
09
Progress and unfinished work in source
This account follows operator and orchestrator source, beyond changelog headings, to establish the model-content path. Interfaces do not establish completed long-running research, and ledger structure does not prove comprehensive retrieval. No model or experiment was run here; low-confidence envelope and placeholder-literature behavior explain the later Stars phase-gate changes.
Reading example · Explaining the workflow
Reading a model-driven mechanism branch
Suppose a locked research direction needs two testable reasoning mechanisms.
- 01
The mechanism role receives the parent, envelope, operator and requested branch count.
- 02
The role returns mechanism fields; the runtime wraps node structure and leaves missing evidence visible to validation.
- 03
Retain the stall reason when a call is unresolved and separately inspect any literature fallback.
A candidate branch traceable to model responses and runtime handling, still requiring literature and experimental validation.
A conceptual reading example. No generation, retrieval, review or experiment was run, and no performance result is reported.
Changes in this version
Content generation
- Earlier approach
Major operators still rely on local placeholders.
- This version’s approach
Major operators request roles and wrap their responses.
Separate research-content production from state management.