Hello Google AI developer community,
I am Anton Fishman, the originator of Recursive Strategy-Scale Formalism and the AI Continuity Engine. I would like to present the project here as a technical research and product hypothesis, and invite criticism, replication, or a small independent test.
The problem
Large context windows help a model see more information, but seeing project history is not the same as continuing the project from the correct position.
During a long workflow, an AI system can preserve the topic while losing distinctions that determine the next action:
- the current version versus an obsolete predecessor;
- verified evidence versus a hypothesis awaiting a test;
- an active direction versus a route that was already rejected;
- a report about a function versus the function itself;
- a locally successful test versus proof of the larger claim;
- preserving stability versus refusing a necessary architectural transformation.
The resulting answer may be fluent and locally reasonable while still continuing the wrong project.
The proposed layer
AI Continuity Engine treats the current working position as an explicit, recoverable object. It compiles project history, corrections, documents, and tool results into a compact Strategy State that preserves:
- the active version and authoritative sources;
- rejected directions and unresolved contradictions;
- the boundary between observation, inference, and proof;
- the causal route that produced the current position;
- the next admissible action;
- the return path after local work, tool use, handoff, or context compression.
This is not intended as another long system prompt. The product hypothesis is a continuity layer composed of a compiler, transferable state carrier, model-native adapters, evaluator, and eventually a training-data pipeline.
The project deliberately does not pre-generate every possible relationship in a large graph. Its map is a navigation grammar: the model derives the relevant route for the present task, while preserving the strategic center and the evidential role of each source.
Why this may matter for Google’s AI stack
The same continuity object could be tested across several Google AI surfaces:
- Gemini API: preserve working position across long tool-using and multi-turn workflows;
- Google AI Studio: compare ordinary summaries with continuity carriers under information parity;
- Google Antigravity: keep local coding fixes connected to the larger architecture after repeated tests and context compression;
- Gemma: run controlled adapter training against the same base model and evaluate whether continuity behavior survives held-out tasks;
- retrieval systems: retrieve not only by semantic similarity, but by source role and current authority;
- multi-agent handoff: transfer the actual project position rather than only a narrative recap.
What has been built so far
- A public conceptual foundation has been published as Recursive Strategy-Scale Formalism v0.3 with a DOI and CC BY 4.0 license.
- Model-native continuity bodies have been assembled for several AI and agent ecosystems, including Gemini and Antigravity.
- Continuity carriers, route-authority packets, retrieval ablations, blind-comparison surfaces, and information-parity controls have been prepared.
- A bounded LoRA/QLoRA program has been designed so that runtime prompting, structure effects, and weight-level effects can be tested separately.
Local experimental signals
These are controlled local observations, not universal performance claims:
- A working corpus was compiled from 772,897 to 14,498 characters while retaining defined operational fields: a 53.31x reduction.
- In one retrieval ablation, keyword retrieval scored 0.923, route-only retrieval 0.758, and the hybrid 0.954. This suggests that route structure should strengthen lexical access rather than replace it.
- In a source-role distractor test, role labels helped, while a compiled authority-and-route packet added 10 local rubric points in both tested branches.
- Under simple information-parity tasks, a narrative summary and a continuity carrier both scored 8/8. This negative result matters: special structure should not be claimed as useful where the task is already trivial.
- An audit found that two supposedly different local Ollama tags resolved to the same model-layer digest, so an earlier short run cannot honestly be cited as evidence of a training effect.
The narrowed hypothesis is therefore specific: any real advantage should come from preserving role, causality, active version, and next action through contradiction, compression, or handoff, not from unusual vocabulary or simply giving the model more text.
A falsifiable pilot
A useful independent test would compare, under equal information and the same model:
- raw project history or a conventional summary;
- the same information compiled into a continuity carrier;
- optionally, a fine-tuned adapter trained on strategically preferred continuations.
The held-out tasks should force version conflict, rejected-path temptation, evidence-boundary errors, context compression, and a fresh-session handoff. Useful measurements include active-version retention, rejected-path suppression, next-action utility, human correction turns, token cost, and ordinary-task parity.
What I am asking from the community
I am not asking readers to accept the mechanism from its description. I am looking for one or more of the following:
- a rigorous objection or missing control;
- a comparable existing method I should evaluate against;
- a suggestion for a Gemini, Gemma, or Antigravity test surface;
- an independent replication attempt;
- contact with a team working on long-horizon agents, context compression, retrieval authority, or model handoff.
A fuller external overview is available for a structured evaluation. The public formalism is available here:
https://doi.org/10.6084/m9.figshare.32253441
Thank you for reading. I would especially value feedback that identifies where this becomes genuinely distinct and measurable, or where an existing architecture already solves the same problem more cleanly.
Anton Fishman