Tempershift Labs

PROJECTFORGE · ACTIVE DEVELOPMENT

AI engineering that separates reasoning from authority.

ProjectForge is a local-first engineering partner designed to inspect real project state, gather evidence, explain what it sees, propose bounded work, and keep execution behind explicit approval.

Observed evidence before interpretation. The working build records page identity, captured structure, screenshot evidence, SHA-256, and completeness state before grounded analysis.

THE PROBLEM

AI can be useful without being silently authoritative.

Many engineering agents collapse several different things into one action: observing the project, deciding what is true, proposing a change, and making the change. ProjectForge deliberately keeps those stages separate.

THE MODEL

Evidence → proposal → approval → execution → verification.

The AI reasons and proposes. Deterministic controls validate evidence and scope. The user explicitly approves. Controlled execution performs only authorized work. Independent verification checks the resulting state.

WHAT EXISTS TODAY

01

Grounded observation

Project and webpage evidence is captured before interpretation. Missing evidence is reported instead of guessed.

02

Evidence identities

Captured state can be tied to deterministic identities such as SHA-256 so reasoning is anchored to a specific observation.

03

Frozen scope

Proposed work is bounded before approval so authorization applies to an exact scope rather than a vague instruction.

04

Controlled execution

Write authority is separate from reasoning authority and is revalidated immediately before execution.

05

Independent verification

An executed action is not treated as proof of success. Resulting state is reread and compared with what was authorized.

06

Persistent engineering state

Scope, execution, and verification history can survive restarts without resurrecting consumed authority.

A FEATURE, NOT A FAILURE

ProjectForge can say “insufficient verified evidence.”

In testing, the Web Observatory refused to produce a semantic ranking when the captured page did not contain evidence strong enough to support one. That refusal is part of the product philosophy: uncertainty should stay visible.

INSUFFICIENT VERIFIED EVIDENCE Missing verified semantic evidence prevents a reliable answer. Generic page text and bare numbers are not treated as ranking signals.
Technical development notes

Web Observatory

Embedded Electron webview, explicit Analyze Page capture boundary, rendered text/structure capture, screenshot evidence, bounded context budgets, and local llama-server interpretation.

Engineering control pipeline

Investigate → Understand → Assess → Propose → Freeze Scope → Explicit Approval → Controlled Execution → Independent Verification → Close or Refuse Closure.

Persistent state

Frozen scopes, execution authorizations/results, and verification records are persisted with integrity validation so restarts remember history without creating new authority.

STATUS

Early software with a deliberately strict foundation.

ProjectForge is not presented as production-ready. Its current value is the engineering model being tested and the working control boundaries already demonstrated.

Discuss an AI engineering project