← Mark Feldman
Active build · v0.1

AI can do the work. Foundry makes the work dependable.

Foundry is a provider-neutral operating layer for repeatable AI work: explicit authority, durable workflows, traceable context, human approval, and evidence that survives the chat window.

The concept in one line

Deterministic software keeps the books. AI handles judgment. Humans retain consequential authority.

Why it exists

A capable agent is not yet a reliable way of working.

General agents are excellent at adaptable judgment. Repeated workplace tasks need something else around them: a defined job, approved knowledge, bounded tools, recoverable state, measurable quality, and a person who still owns the decision.

General agentFoundry adds the operating layer
A model can attempt a sequenceFoundry checkpoints it and resumes without replaying completed work
A model can call a toolFoundry decides whether that tool is allowed here, now, for this job
A model knows the conversationFoundry assembles approved, source-traceable context
A model produces an answerFoundry records evidence, cost, decisions, quality checks, and outcome
A provider supplies judgmentFoundry keeps the work portable when the provider changes

The governed work loop

Judgment moves. Authority does not.

01

Define the job

Mission, scope, evidence, tools, budgets, and success are versioned before the run.

02

Assemble context

Only approved, bounded, source-traceable material crosses into model judgment.

03

Run and checkpoint

Deterministic steps and AI decisions are recorded. Interrupted work resumes safely.

04

Human gate

Consequential effects stay blocked until an authorized person approves the exact outcome.

Every step leaves durable state, evidence, cost, decisions, and an observable outcome.

The architecture

Reuse the procedure. Keep trust local.

Foundry separates portable machinery from private context and from the model providing judgment. A reusable Skill carries instructions—not repository access, credentials, approval power, or a hidden data destination.

1

Portable core

The operating machinery

Contracts, workflow execution, checkpoints, adapters, evaluations, budgets, and event history.

2

Isolated workspace

The trust boundary

Agents, approved knowledge, installed Skills, resource grants, private state, and local runtime policy.

3

Replaceable runtime

The judgment engine

Hosted or local models sit behind one validated boundary. The provider does not own the workflow.

Default deny
No secrets in state
No model self-approval

Where it stands

Five chapters built. The sixth is being proven.

Foundry is not a finished platform. It is an executable research-and-build program with complete foundations through durable workflows, and active work turning one proven procedure into a capability that transfers cleanly.

  1. 00

    Foundation

    Validated state, append-only events, contracts, and a tested command surface.

    Complete
  2. 01

    One real cycle

    A bounded, human-approved build cycle with decisions, artifacts, and a recorded outcome.

    Complete
  3. 02

    Hardened harness

    Provider-neutral model calls, tool authorization, recoverable transitions, and exact provenance.

    Complete
  4. 03

    Context that earns trust

    Source-hashed retrieval, bounded cited context, and measured grounding instead of prompt stuffing.

    Complete
  5. 04

    Durable work

    Checkpointed workflows, retries, budgets, parallel joins, authorized effects, and human gates.

    Complete
  6. 05

    Reusable capabilities

    Promoting proven procedures into portable Skills without smuggling data or authority with them.

    Active
  7. 06

    Generalization

    Prove the system transfers to a second, unlike product and remove assumptions that do not travel.

    Next

Evidence, not theater

The system is allowed to find itself wrong.

151 + 149

Tests and subtests passing

Verified against the current working tree on August 31, 2026.

8 / 8

Frozen Skill-routing cases

The first run scored 7/8. The false positive became a tighter boundary; the identical rerun passed all eight.

0

Model-owned human approvals

A model result cannot approve itself, even when its structure is valid.

A recent correction

Safe execution was not the same as a good answer.

A read-only Steward run passed structural checks but produced an answer that was not grounded or coherent enough to trust. Human approval exposed that the evaluation had measured execution safety without adequately measuring answer quality.

Foundry now separates those scores, blocks approval unless both pass, rejects several deterministic failure patterns before the human gate, and accounts for failed or rejected model calls instead of making their cost disappear.

The new gate is implemented and tested. A fresh end-to-end model result that earns the stricter quality proof is still outstanding.

What remains

Prove transfer, then broaden.

The immediate work is to re-earn the Steward quality proof under the stronger gate, generalize output evaluation beyond one repository, and run the complete capability against an unlike codebase. After that, a second private product will test whether Foundry's procedures travel beyond software-project stewardship.

01Re-earn end-to-end answer quality
02Transfer to an unlike repository
03Generalize through a second product
04Add standards only where they earn their keep

The assessment

Foundry is already more than a concept: its core boundaries, retrieval, workflow durability, workspace isolation, evaluation machinery, and first reusable Skill are executable and tested. It is not yet a generally proven platform. Its strongest quality is that the remaining distance is visible in contracts, tests, recorded decisions, and gates that are allowed to stay red.

Back to current projects