Prismatic Agents

A long-term engineering proposition

Persistent cognition. Engineering knowledge that carries forward.

The technology remains the company. Its own operating experience should help improve the technology.

Prismatic should learn from improving Prismatic.

Machines work across shifts, restarts and releases. Their cognitive infrastructure needs persistent identity, memory, governance and evidence. Prismatic's source studio is the first place to test whether those same mechanisms can help maintain the software itself.

A failed model or tool call should leave evidence, a reproducible test, an accountable repair and a lesson for the next agent session. A model swap should not erase that engineering record.

Meta · The Patient Venture opportunity

A mind that describes itself—and learns how to improve.

Prismatic describes what it is for, how it is organized, what is running, what failed and what previous repairs taught it. Its North Star and blueprint express intent. Dated runtime observations and repair receipts show what actually happened. The Evidence Keeper compares these accounts and makes the unknowns visible; the Founder's Navigator brings what it finds to the founder.

Each studio that Genesis creates assigns these roles to its own agents by responsibility. Its main companion becomes the Founder's Navigator. The agent whose work is closest to systems, verification and reliability becomes the Evidence Keeper.

The implemented meta layer keeps a versioned self-description and a local vault metadata change catalog. It tracks coverage without uploading private file bodies wholesale. Approved Brain adapters and repair history provide operational evidence; other records remain inventory-only until investigated. A changed file is not proof that its contents were understood.

Describe → observe → compare → investigate → approve → test → measure → remember. A later agent can use the retained evidence to explain a decision or avoid repeating a failed approach. This is accumulated engineering knowledge, not a claim of consciousness or automatic model retraining.

The investment thesis: give a small team time to turn operating experience into tested engineering knowledge, so every investigation can inform the next. The potential value is fewer repeated failures, stronger reliability and lessons that survive model changes—not simply more generated code.

For Patient Venture, the question is testable: does retained knowledge reduce repeat failures, improve reliability and survive a change of model? Daily review is active; those production benefits still need measured proof.

Activation status · September 29, 2026

Daily review is active. Production improvement still needs proof.

The Evidence Keeper's daily evidence review is scheduled for 07:00 America/Toronto. Its first review completed on the configured frontier model and was delivered through the Founder's Navigator for founder review.

The founder-approved, isolated coding path is installed. Qualification checks distinguish a working coding fixture from a real Prismatic repair. No autonomous production fix, verified local-model improvement or hardware efficiency gain is claimed.

From observation to an approved repair

  1. Observe: read the North Star, blueprint, Brain Guide, operational telemetry and prior outcomes.
  2. Understand: the Evidence Keeper distinguishes evidence, hypotheses and gaps, then proposes a bounded task.
  3. Review: the Founder's Navigator brings the founder the evidence, scope, tests and coding prompt.
  4. Code and test: an approved task runs in an isolated checkout. A successful tool exit alone is not completion.
  5. Release within authority: production deployment is a separate decision with its own validation and rollback.
  6. Measure and remember: compare suitable workloads, retain failures and corrections, and use the evidence in later work.

Frontier assistance, separately qualified local intelligence

A frontier model can help investigate and write code. The intended local model and physical host must pass their own workflows. Qwen3.8-27B is a candidate until its artifact, quantization, runtime and hosts are pinned and qualified. Passing a cloud test does not establish local tool reliability, prefill, decode, memory use or watts.

Keep complete relevant evidence. Measure queueing, inference, tool execution, review and delivery separately. A faster response with missing evidence is a regression.

The record should teach the founder too.

Ask what failed, which task is approved, which patch was tested, what is actually running and why it advances the North Star. The Evidence Keeper explains reliability, the Founder's Navigator explains alignment, and the domain specialists guide readers from dated evidence.

The reusable company or personal knowledge guide can explain an owner's approved records. It does not carry this source-code worker, its credentials or its improvement history into customer studios or Genesis.

Why Patient Venture fits

Reliable persistent cognition needs time: reproduce failures, qualify local models and hosts, test recovery, observe recurrence and demonstrate that useful lessons survive model changes. A fair Patient Venture partner can give the founder breathing room and fund a small engineering team to build that evidence.

Approved research can improve guides, comparisons and calculators, attracting engineers and potential design partners. Revenue may fund further research. This complements the technology; it is not a promise of income, investment returns or a proven commercial moat.

What to ask us to demonstrate

Useful measures include accepted outcomes, repeat incidents, false completions, review effort and resource cost on specified workloads. Code volume and tool-call counts are not measures of intelligence.