Technology
Future
What is left to build in Genome 2.0, and ideas for a later version. Nothing on this page exists yet.
Plans, not promises
Every item here is either written down but not built, or an idea that may not work. There is no date for any of them, and this page gives no predicted number for any of them. What is built today is on the technology page.
How to read this page
Specified, not built means a design exists and the code does not. Research direction means the idea is not a planned task. Some research items need new evidence before anyone should build them. The reference numbers (G41, R03 and so on) are rows of the status matrix in the Agent Genome book.
Finishing Genome 2.0
The platform behind this technology has a task list for Genome 2.0. These are the parts that were still open when
the status tags were written, at commit 0575676. The IntelliNix repository has had
commits since then that touch several of these items, so this table may be out of date. It will be re-checked at the next review.
| Item | What it is | Status |
|---|---|---|
| Confirm or dissolve gene families | Let a person accept or dissolve a proposed group of related genes. | Specified, not built G41 |
| Inheritance between agents | Start a new agent from another agent’s genome, with learned genes arriving as proposals, and offer it in the Concierge. | Specified, not built G42, G43 |
| Governance alerts, panel and lineage download | Warn an administrator about unusual genome activity, and give them a screen for governance and for downloading lineage. | Specified, not built G48, G74 |
| More capability arms in the benchmark | Test more parts of the genome (for example the layers and the task-aware selection) in GenomeBench, and extend the CI gate to cover them, not just the kernel and learned strata. | Specified, not built G81, G83 |
| Other model providers in the benchmark | The benchmark can call only the Gemini client today. Other providers are needed before any non-Gemini result exists. | Specified, not built G82 |
| Dynamic compilation, drift and regression alerts | Pick what to compile based on the task, and warn when quality drifts or regresses. | Specified, not built G56 to G58 |
| Attested scorecard and quality screens | A scorecard whose figures are attested, and the screens that show it. | Specified, not built G59, G75 |
| Life-stage screens | Screens in the web app for threads, roles, families and the effect of a genome on an agent. | Specified, not built G73 |
| Rollout runbook and per-phase sign-off | A written rollout plan, and a live benchmark sign-off before each phase reaches customers. | Specified, not built G87, G88 |
Genome 3.0 proposals
Genome 3.0 is a name for ideas, not a product. Most of them answer a limit that the Research page states: the benchmark is a simulation, the baseline is lexical, and no one has separated which part of the genome does the work.
| Idea | What it would do | Status |
|---|---|---|
| Stronger baselines | Compare against embedding retrieval, a larger number of retrieved items, reranking and published memory systems, not only lexical top-3 retrieval. | Research direction R03 |
| Ablations | Separate what consolidation, strength and the fixed prompt scaffold each contribute, with a scaffold-only control. | Research direction R04 |
| Honest statistics | Intervals that treat the benchmark’s units as related, such as resampling whole personas or a mixed model, and a steady-state metric. | Research direction R08 |
| Semantic clustering | Group observations by meaning, not word overlap, while keeping the deterministic checks as the gate. | Research direction R09 |
| Cross-model study | Compile one genome in all four dialects and measure how each model family follows it. | Research direction R10 |
| Learned value functions | Replace the hand-set values used to choose genes with values trained offline and served as a frozen record. | Research direction R11 |
| Transfer between agents under consent | Carry a small evidence summary with a transferred gene. The gene would still arrive as a proposal. | Research direction R12 |
| On-device compilation | Compile from an AGF file on a device or with a local model, and check that the result matches the server. | Research direction R13 |
| Log compaction that keeps exact erasure | Smaller stored logs and snapshots without losing the ability to erase one person’s data precisely. | Research direction R06 |
| Per-tenant keys and data residency | Separate encryption keys for each workspace and hosting in a chosen country. | Research direction R07 |
| Consented pilots | Try the genome with real users who agreed to it, and measure with the live scorecards. | Research direction R05 |
| Cross-workspace learning | Learn shared defaults across workspaces. It is out of scope today because it would cross the line between customers. | Research direction R02 |
Distillation and fine-tuning
Research direction R01 One hypothesis is that a model trained on outputs a customer accepted might need fewer genome characters for the same result. Nobody has tested it, and it conflicts with the protections that the genome gives today. Erasure by provenance, lineage and tenant isolation work on text. They do not reach model weights. For that reason the platform behind this technology does not train on customer data, and this design is a hypothesis with an experiment sketch only.
A fair test would use synthetic personas only, hold out some personas, compare against retrieval and a plain list of past feedback, and include an erasure test: delete a person’s data, then rebuild the model or adapter and check that the effect is gone. It would be pre-registered and have stopping rules. This site gives no predicted result for it.
Open questions
- Do the rule-checker results say anything about real customers?
- How much of the measured gain comes from the fixed prompt scaffold and not from learning?
- Does the report block work on other model families?
- How do learned genes drift over months?
- What does the prompt cost against what it saves?
These are listed on Evaluation and Models as well. Honest answers need measurements that do not exist yet.