01
Make feedback alter exploration
7.1 begins with counterexamples in 7.0: high and low observations can lead to the same idea; unrelated text and an unrelated valid observation can open a review gate; zero-cost reading can execute with zero action quota. Repairs target reproducible behaviour, explicitly connecting observations, subsequent actions and review rather than substituting longer proposals for feedback.
02
Conditional rival updates
The new epistemic module compares prespecified predictions with observed labels. Incompatible rivals are conditionally eliminated only under a host-certified deterministic categorical mapping and a compatible response. Unknown labels, noise, missing predictions and uncertified semantics remain unresolved. A response matching no rival signals possible model-class or measurement mismatch rather than deleting every explanation.
03
Revisions cannot silently resurrect rivals
QuestionRevision explains introduced rivals, reintroduced refuted rivals and removed live rivals separately. Resurrection requires independent new evidence rather than the same receipt that excluded the rival. Reusing a question ID for changed scientific semantics is rejected. The contract permits reframing while preserving traceable reasons for reversals and changes of direction.
04
From a chain to a persistent frontier
FrontierGraph stores problem, idea, obligation and action nodes with branches, parents, dependencies, sources and outcomes. Macro control selects affordable dependency-ready obligations and rotates stale untested branches. OPEN, BLOCKED, REFUTED and COMMITTED are workflow states; COMMITTED expresses attention or commitment, rather than scientific confirmation.
05
Macro focus enters the micro loop
The selected branch supplies its question, idea, observations and pending action to the next context. A proposal must extend that branch and preserve the pending specification before semantic and budget checks. The micro loop proposes, attests, observes and updates; its outcome returns to the frontier. Deferred actions can execute later, while actions invalidated by changed rivals retain a blocking reason.
06
Newly acquired text must enter the next round
Literature actions acquire identified spans that enter later bounded reading contexts. Reading is evidence acquisition, not a completed scientific experiment. Unrelated observations cannot open mechanism confirmation; literal overlap with a negative paper creates prior-work risk that still needs semantic review. Discriminating a question and supporting an idea mechanism remain distinct review conclusions.
07
Costs, quotas and resume identity
Declared, attested and actual costs are checked separately. Reservations precede calls; overruns remain charged and stop execution. Zero action quota blocks reading even if its price is zero. Resume locks the proposal provider rather than permitting a replacement generator on an old journal. Oversized or unreadable source fingerprints stay explicit, and complete rounds save frontier, candidate and budget checkpoints.
08
Inputs and artefacts
Inputs remain a direction, source manifest, budgets, provider and host adapter, now with persistent branch context. Outputs add frontier events, compatibility analyses, transition contracts, deferred actions and rejections alongside readable reports. The graph lets later rounds inherit state without giving policy hidden truth or accepting a model-reported novelty score.
09
Counterexample repairs and reported checks
Tests specify high/low responses, unknown labels, resurrection rules, parent changes under macro focus and later execution of deferred actions. The historical report records 2,096 passes, 41 skips and no failures, with Ruff, Mypy and version checks passing. Source availability is indexed, but span counts do not certify reading quality or correctness. This review inspects contracts and assertions without executing archived tests.
10
The frontier capability boundary
The 7.1 CLI host remains a finite synthetic feature world. Programmatic domain adapters are possible, but an independently verified real scientific-domain implementation is absent. Providers still return one idea per call; multi-policy generation, cross-branch Fusion and long-horizon graph gains are unestablished. Representation expansion, epistemic-debt routing and meta-policy revision remain hypotheses requiring semantic validation, full-cost baselines and unseen task families.
Reading example · Explaining the workflow
Illustration: test one branch, then return to a deferred action
Two open branches, an attested response mapping and two costed actions.
- 01
Macro control selects a dependency-ready obligation.
- 02
A host observation conditionally updates compatible rivals.
- 03
The unexecuted action remains in the frontier and can regain focus.
- 04
Recheck the old action after rival changes, then execute or block it.
Branch state retaining sources and budgets, with traceable continuation, rollback or stopping.
An architecture example, not evidence of long-horizon scheduling gains.
Changes in this version
Feedback semantics
- Earlier approach
High and low observations can lead to the same follow-up idea.
- This version’s approach
Conditional updates record compatible, excluded and unmodelled rivals.
Make observed responses affect research state.
Branch state
- Earlier approach
A proposal chain can lose deferred actions.
- This version’s approach
A persistent frontier returns pending obligations to later contexts.
Preserve research routes that remain untested.
Review and resume
- Earlier approach
Unrelated observations and provider replacement expose admission gaps.
- This version’s approach
Unrelated observations cannot confirm mechanisms; resume locks provider identity.
Give confirmation and record continuity separate contracts.