Roadmap & Project Status

Orrin is an experimental research prototype under active single-developer development. This page is the honest status: what works today, what's experimental, and what's out of scope. It mirrors the README's claims-and-evidence framing; the authoritative to-do list is docs/POST_RUN13_MASTER_PLAN_2026-10-07.md, and the evidence is the run reports.

What works today

Capability	State
Symbolic-only runtime (no API key)	Working
Continuous cognitive loop	Working
Persistent goals (durable daemon, WAL/snapshots)	Working
Memory: working + long-term, consolidation, retrieval	Working
Control signals + regulation	Working
Host coupling (reflex / signals / cadence)	Working
Global workspace + ignition	Working
Effect ledger + grounded production reward	Working
Quality standard (human-ratified bar)	Working
Runtime telemetry + Face & Brain UI	Working
Native desktop app (unsigned builds)	Working
LLM as a gated, fail-closed, multi-provider tool	Working

Experimental / in progress

Area	Status
Learning from outcomes at scale	Experimental — mechanisms exist; being validated on staging runs
Native from-scratch language model	Experimental — learns from reading; gated speech handoff
Self-extension (self-written cognitive functions)	Experimental / high-risk — sandboxed and reviewed, see Self-Code and Extension
Long-run behavioral stability	Experimental — long runs can drift into states not yet fully characterized
Benchmarks & evidence ledger	Ongoing — see Benchmarks and Verification

Out of scope

• Orrin being "human-like" or sentient — not claimed. Cognitive terms name engineering mechanisms.
• Production hardening / security guarantees — this is a research prototype (Security Model).

How progress is measured

Development proceeds against staging runs: a fresh instance lives for a while, then its behavior is audited against an acceptance gate (docs/NEXT_RUN_TESTS.md) and sealed into a life capsule (Existence and Lifecycle). Each run gets a dated report under docs/Behavioral Evaluation & Runtime Diagnostics/demo_runs/. That evidence — not a feature checklist — is what moves a capability from Experimental to Working.

Where it stands (Runs 1–13)

Thirteen staging lives have been scored; the acceptance gate has not passed yet, but the binding constraint has climbed a ladder:

Runs	What the gate was stuck on
1–4	basic mechanics: does work get produced and credited honestly
5–8	economics: one goal monopolizing attention (broken in Run 8: 90.9 % → 42.6 %)
9	honesty: failures that weren't really failures
10–12	the research feed going silent, and "growth" that was hollow
13	growth became real for the first time — questions answered with new knowledge, and a prediction about his own behavior confirmed on data recorded afterwards. Still failing: research silences, goal churn, and a qualitative finding: he mostly thinks about himself (74 % of long-term memory was summaries of his own activity)

Next (Phase B, the Run 14 build): stop the goal churn, unblock goal generation, keep his own logs out of his memory, and make most of his memories about his world. docs/PHASE_B_BUILD_BRIEF_2026-10-08.md has the details.

Known limitations

• Internal APIs still change quickly.
• Desktop builds are unsigned (expect OS trust prompts).
• No first-class low-resource install profile yet.
• Some capabilities are evidence targets, not settled claims.

Where to follow along

• Releases — tagged checkpoints
• docs/POST_RUN13_MASTER_PLAN_2026-10-07.md — everything still to build, in order
• Run reports — the behavioral evidence trail (docs/NEXT_RUN_TESTS.md has the gate history)
• ricmassey.com/orrin.html — plain-English progress notes after each run or build, and whether he's running right now