LIFE Roadmap Report

Six-Month Roadmap

Work plan for LIFE from April 15, 2026 to October 15, 2026, centered on three parallel streams: app delivery, benchmarking, and research loops.

Window: April 15, 2026 to October 15, 2026
Status: Draft v4
Primary Sequence: Internal testers → Alpha testers → Systematic improvement
Overview

Priority Order

The roadmap is organized around a simple delivery sequence, while still allowing three streams to move in parallel.

6
Month Window
3
Parallel Streams
3
Core Delivery Steps
4
Hard Gates
1
Bring the app to internal tester level

Stabilize product infrastructure, release internal builds, and validate core voice, content, and workflow quality with the org first.

2
Bring the app to alpha tester level

Use internal feedback to reach a broader alpha standard across voice, memory, session quality, and testing operations.

3
Build systems that let LIFE improve in a disciplined way

Use the benchmark stack, knowledge base, and research loops to make prompt, constitution, and model decisions based on evidence.

Workstreams

Three Parallel Streams

Each stream has a different job. The app ships the product, benchmarking measures quality, and research loops turn benchmark results into improvement.

Stream 1

App

Get the app from current state to internal tester level, then to alpha tester level.

Production hardening Voice Video content Updated memory Dream system Deep research Council mode Slack alpha channel
Stream 2

Benchmarking

Build the benchmark stack that measures whether LIFE is actually improving.

Golden Questions BodhisattvaBench v0.5 SOME Tradition sub-benchmarks Combined benchmark Benchmark KB
Stream 3

Research Loops

Use the benchmark stack to run systematic improvement loops on prompts, constitutions, and model behavior.

Prompt variants Constitution variants Language experiments Interpretability Weekly runs Daily micro-runs
Outcomes

North-Star Outcomes By October 15, 2026

These are the concrete states the roadmap is trying to reach by the end of the six-month window.

App

Research-Ready App

  • iOS and Android builds distributed to a research / alpha cohort
  • Production auth, RLS, WebSocket, session, and voice flows verified end-to-end
  • All Stage 1 and Stage 2 content uploaded, cleaned, and usable
  • Stage 2.4 no longer blocked by placeholder content
  • Safety, moderation, and analytics sufficient for live cohort operations
Voice

Voice Experience Finished

  • Voice feels reliable, low-friction, and stable across real devices
  • Latency is masked well enough that the app feels coherent, not stitched together
  • Yellow-zone personalization is live for selected practices
  • Audio stitching / session continuity is implemented for high-value cases
Memory

Memory + Dream System v1

  • Wisdom has a production-grade user memory layer
  • A scheduled dream / reflection layer periodically reviews user history and updates derived profile knowledge
  • User imports from Anthropic and ChatGPT exist for memory/profile bootstrapping
  • First-party profile inference is useful, inspectable internally, and safety-bounded
Knowledge

Knowledge Base + Deep Research

  • Curated contemplative knowledge base exists as an md-based wiki KB with provenance, quality review, and versioning
  • The markdown wiki KB is accessible to agents as shared infrastructure
  • Separate KBs exist for contemplative knowledge, benchmark data, and user data
  • Deep research runs on top of vetted internal knowledge, not random raw files
Agents

Agent Access Layer

  • Golden Bot evolves from proof of concept into a usable internal agent for benchmark data
  • People can chat with agents that access knowledge layers
  • Agents can be called from Slack
  • A research agent can access the benchmark KB and user-data KB where permissions allow
Benchmarks

Benchmarking + Research Loop Operational

  • One production-grade AI benchmark is running repeatedly and generating decision-useful reports
  • One SOME exists at alpha/beta maturity
  • Prompt, constitution, and provider experiments run on a fixed weekly cadence, with optional daily cadence if rate limits allow
  • Results are visible in a report artifact, not trapped in chats
R&D

Long-Term Model Bets De-Risked

  • Benchmark targets are defined
  • Curated datasets exist
  • Reward / rubric design exists
  • An experimental harness exists before any serious fine-tuning or RL bet
Infrastructure

Shared Knowledge Architecture

The knowledge layer should be inspectable and reusable across product agents, benchmark agents, and research agents.

Markdown Wiki KB

The long-term plan is to move the broader knowledge base into a markdown-based wiki KB that approved agents can access.

  • Markdown is the canonical source of truth
  • Vectors are a derived retrieval layer
  • Review and provenance stay attached to the markdown source
  • The same KB can be reused across product, benchmark, and research agents

Separate Knowledge Layers

These layers should remain distinct even if they share tooling.

  • Contemplative knowledge KB
  • Benchmark data KB
  • User data KB
Cleaner access Less accidental mixing Better research hygiene

Important direction: the roadmap treats vectors as a retrieval layer, not the primary knowledge substrate. The markdown wiki is the thing to curate, review, diff, and share across agents.

Agent Layer

Agent Surfaces

The roadmap includes multiple agent entry points, but they do not all start with the same scope.

Golden Bot

Golden Bot is the current proof of concept.

  • Initial scope: access Golden Questions benchmark data
  • Let people chat with that benchmark data directly
  • Be callable from Slack
Focused first scope Slack callable

Research Agent

A separate research agent should support deep research and internal analysis work.

  • Access the contemplative KB
  • Access the benchmark KB
  • Access the user-data KB where permissions allow
Deep research Internal analysis

Alpha Testing Slack Flow

There should be an internal alpha-testing Slack channel for the app where people in the org can test, report bugs, suggest UI changes, and call an agent such as Claude or Codex to investigate and attempt a fix in the cloud.

Surface Initial Use Policy
Slack alpha channel Internal app testing, bug reports, UI feedback Broad org participation
Cloud-fix agents Investigate issues and attempt fixes Loose internal permissions for now
Change governance Review + revertability Tighten later if the system expands
Dependencies

Hard Gates

These are the ordering constraints that keep the three streams from getting ahead of their foundations.

Gate 1

Knowledge Base Foundation

Needed before deep research, SOME depth, tradition sub-benchmarks, and high-quality expert routing can really work.

Gate 2

Golden Questions Benchmark Baseline

Needed before serious system-prompt iteration, language testing at scale, and reliable prompt-comparison loops.

Gate 3

BodhisattvaBench v0.5 Baseline

Needed before constitution experiments, alignment-method comparisons, and any early fine-tuning or RL design that needs a target.

Gate 4

Internal Tester App Release

Needed before larger alpha expansion and before trusting app-generated data on memory, dream behavior, and user feedback at scale.