01
Give innovation a type and lineage
Poincaré makes ResearchDelta the research unit: a candidate specifies changes to the question, representation, assumptions and mechanism, discriminating consequences, falsifiers and parentage. V6.7 moves evidence from report fields into append-only event chains and replaces a one-shot pool with a multi-generation ecology, retaining contrarian budgets, multiple niches and stepping stones.
02
Commit before selection and reveal
Candidates should not rewrite predictions after seeing a challenge. Commit–select–reveal–execute persists the candidate and prediction before challenge selection, reveal and execution. External root anchors, signatures and an independent verifier expose coordinated edits to ledger and report hashes. Search-fertility credit is separated from world-settled value; admitting descendants does not give ancestors discovery credit.
03
Frontier, compiler, court and archive
Frontier snapshots organize strongest priors and agendas. Isolated islands produce raw ideas, which compile into ResearchDelta. Prior absorption and adversarial court assign repair, duplicate, experiment-ready or missing-adapter states before archiving and event commitment. Later generations select elite and tail parents for mutation. After closure, the verifier rechecks the event chain from its root.
04
Research deltas and claim binding
ResearchDelta carries problem, representation and assumption changes, mechanism and claim graphs, consequences, falsifiers, world bindings, executable artifacts, uncertainty and lineage. ClaimBindingCertificate links claims through mechanisms, observables, metrics, predictions, artifact hashes and adjudication. Parent events must already be persisted and match lineage parent_candidate_ids exactly.
05
Lineage and unsettled states
Inputs include topic, frontier priors, agenda allocation, generation and pool budgets, world adapters and explicit fixture configuration. Outputs include typed candidates, absorption verdicts, adversarial court results, archive cells, lineage, commitment events, signed reports and independent verification. Candidates missing adapters retain BLOCKED_ON_MISSING_WORLD_ADAPTER rather than presenting a specification as an executed experiment.
06
Conceptual walkthrough: archive a representation change
A proposed representation change must state why the old question is insufficient, information loss between representations, a mechanism graph, consequences differing from priors and falsifiers. Even if compilation and court succeed, an unregistered mathematical checker prevents settlement. Descendants can link to committed parent events; this demonstrates organization, not real-world validation of their mechanism.
07
Trace the entire event chain
Compared with Noether, V6.7 extends recomputation from report digests to causal events, artifacts, predictions, adjudication, lineage and derived views. Strict canonical JSON rejects NaN, default=str and process-salted hashes. Generation gains multi-generation archives, but islands and descendants remain local compositional generation without external LLM calls; V6.8 adds real provider and observation-feedback paths.
08
Actual world-adapter coverage
Source includes append-only EventStore, external immutable receipts, a signed root, detached report signatures and an independent verifier. WorldFactory marks only algorithm as production_executable; mathematical, simulation, data/ML and prospective worlds lack full production execution, with math explicitly raising NotImplementedError. Default production candidates therefore remain unsettled; fixtures keep zero credit.
09
Record integrity is not discovery validity
Signatures and anchors can expose edits within their protected storage scope, while trust still depends on anchor control, tasks, adapters and checking logic. Well-typed objects, valid lineage and passing event verification do not establish valuable questions or scientifically confirmed mechanisms. This version honestly retains non-executable candidates and leaves clear extension points for an observation-driven loop.
Reading example · Explaining the workflow
A candidate without a mathematical checker
A specification proposing a representation change with discriminating rival predictions.
- 01
Complete ResearchDelta question, representation, mechanism and falsifier fields.
- 02
Apply prior absorption and adversarial court.
- 03
Retain the missing-adapter state without a registered math checker.
- 04
Archive and commit events; descendants cite persisted parents.
- 05
Verify from the root after closure; unsettled candidates earn no discovery credit.
A typed, lineage-bearing candidate still awaiting world execution.
Conceptual contract example; no mathematical experiment or external model call was run.
Changes in this version
Candidate
- Earlier approach
Claim metadata and audit reports
- This version’s approach
ResearchDelta and full-chain binding
Specify what actually changes
Evidence
- Earlier approach
Digest recomputation
- This version’s approach
Append-only events, external roots and independent projection
Check causality and coordinated tampering
Search
- Earlier approach
One candidate pool
- This version’s approach
Protected budgets, lineage and niches
Preserve tails and stepping stones