IJCCRL • Deepalienist • R&D Roadmap 2026
Deepalienist Experience Optimisation Roadmap — A Safer, Measurable Plan for experience.exp
We are opening a focused development track around one of the most interesting practical layers in a modern Stockfish-derived engine: the experience file. The goal is simple to describe and difficult to do properly: improve how Deepalienist reads, evaluates and reuses accumulated experience without breaking source stability, compatibility or default playing behaviour. This is not a marketing shortcut or an instant Elo promise. It is a disciplined roadmap built around measurement first, engine safety first, and transparent experimentation only after the data is in hand.
What this roadmap is about
Why this matters
- PracticalAn experience file can become a real competitive asset only if the engine uses it intelligently, consistently and safely under real search conditions.
- ScientificBefore claiming improvement, we need to know what is actually being read, what is being filtered, and how often stored moves meaningfully influence root decisions.
- EngineeringThe goal is not to bolt on risky heuristics. The goal is to measure the active subsystem, preserve compatibility and only then test targeted optimisations.
- CommunityFor serious engine followers and subscribers, this is exactly the kind of behind-the-scenes work that separates cosmetic versioning from real technical development.
What has already been done
- AuditInstrumentation has been added to the active experience subsystem path so the engine can report load/save timing, probe activity, filtered hits, root-move reordering events and entry-writing behaviour without changing default strength.
- SafetyThe new audit layer is disabled by default, which means the current competitive behaviour remains unchanged unless explicitly activated for testing.
- ToolingA separate offline audit tool has been created to inspect and compact textual experience.exp files outside engine runtime, allowing research without contaminating match conditions.
- DisciplineThe inactive legacy learning path was deliberately left untouched in this first phase. The work has stayed on the active implementation only.
What comes next
- Phase 1.5Run the audit tooling against real Deepalienist experience files to measure distributions, duplicates, multi-move positions and practical retention under different thresholds.
- Phase 2Collect engine-side reports from real tests to determine actual probe hit-rate, filtered hit-rate and the practical frequency of root influence.
- Phase 3Only after the numbers are available, open a separate opt-in experimental branch for ranking refinement, confidence filters and softer root-move promotion.
- RuleNo default Elo claim, no miracle claim and no public “strength gain” statement will be made before controlled comparison work is completed.
| Work block | Objective | Current status | State |
|---|---|---|---|
| Subsystem mapping | Confirm the active experience path and avoid touching inactive learning code during the first pass | Completed as the foundation for safe work | Done |
| Engine-side instrumentation | Measure load/save timing, probe activity, filtered hits, reorders and write counts | Implemented behind opt-in controls | Done |
| Offline experience audit | Inspect and compact textual v1 files without modifying runtime behaviour | Implemented as a separate external tool | Done |
| Real-file measurement pass | Analyse actual Deepalienist experience data and quantify useful signal vs noise | Next operational step | Planned |
| Experimental ranking branch | Test safer move-ranking refinements and confidence-based filters | Deferred until after the audit data is reviewed | Planned |
| Public performance claim | Only publish strength conclusions after controlled comparison work | Explicitly not claimed at this stage | Pending |