Sebastián Ojeda
ES

DICS + Project Integrity Kernel

Agentic Engineering Governance

Two complementary systems for authority, evidence, stale detection, handoffs and context recovery in agent-executed projects.

Status
Experimental prototypes under validation
Role
Applied research, architecture, implementation and evidence-contract design
Year
2026
Back to projects

Governing agentic work without depending on conversational memory

DICS and Project Integrity Kernel are separate repositories because they solve different problems, but together they express one capability: Agentic Engineering Governance / Decision & Evidence Systems.

Their shared thesis is simple: a project executed with agents cannot treat chat as the canonical source of state, evidence or authority.

Conversation is cache; durable state is authority.

The implementation repositories are private. This case study publishes the architecture and evidence boundaries without pretending there is a public product release.

DICS: observing decisions and sources without taking control of the product

DICS is a Decision Intelligence & Control System. Its current M1 projection is deliberately read-only.

It models:

  • sources and claims;
  • authority and configuration precedence;
  • zero-write adapter contracts;
  • local bootstrap and session primitives;
  • replay fixtures;
  • static no-write audit.

The important decision is negative: DICS does not execute product work and does not write to governed GitHub, Drive or ledger sources in M1. Canonical truth remains in the original sources and any DICS cache is disposable.

The M1.5 contract also prevents a false conclusion: green CI does not prove usefulness against real sources. Live validation remains a separate gate.

Project Integrity Kernel: integrity, evidence and recovery

PIK is a runtime-agnostic kernel implemented in Go + SQLite for reconstructing the state of agent-executed projects.

The EK-EXP-001 experiment already proves properties stronger than merely “saving state”:

  • multi-source ingestion with explicit authority;
  • stale handoff detection;
  • authoritative artifact resolution;
  • REWORK lineage preservation;
  • immutable evaluation snapshots;
  • exact evidence-to-artifact binding;
  • no automatic PASS inheritance;
  • kernel-computed Critic independence;
  • ResumePack freshness and tamper detection;
  • optimistic STALE_WRITE rejection;
  • precise Human Gate recovery;
  • a genuine Context Amnesia Test from a fresh runtime.

PIK remains an experimental prototype ready for local evaluation. It is not presented as a SaaS platform, PM suite or agent runtime.

Why they belong together

DICS answers: what is happening, which sources claim what, and which decisions need attention?

PIK answers: which state and evidence can be considered intact after handoffs, rework or total context loss?

Together they form an engineering story about authority, provenance, staleness, evidence and recovery rather than “agents doing more tasks.”

What this story demonstrates

  • agentic system design with authority outside chat;
  • evidence-bound reviews;
  • stale detection and optimistic concurrency;
  • recoverable handoffs between runtimes;
  • explicit Human Gates;
  • separation between governance/observation and execution;
  • local-first, runtime-agnostic architecture;
  • discipline not to market experimental capabilities as production.