Erasmus–Protomentat

A personal-first, persistent cognitive system for a human–AI partnership.

The system combines bounded conversational continuity, deterministic-first capabilities, an epistemic ledger, sleep consolidation, mission execution, and a 10th-Man cognitive immune system.

Core separations

• RAG preserves explicit memory and evidence.
• Ledger preserves current propositions, confidence, contradictions, and tangible wrongness.
• Sleep integrates session experience.
• Skills and LoRA preserve validated adaptive intelligence.
• 10th-Man immunity detects divergence and prevents shared hallucination.
• Mission engine converts cognition into bounded execution.
• OKF Knowledge Foundry converts PDF source folders into provenance-bearing draft OKF v0.2 candidate bundles without promoting them into canonical knowledge.

OpenCode Erasmus layer

The repository includes a provider-neutral OpenCode agent, explicit slash commands, and small composable skills. These files are a thin interaction layer over the real Erasmus CLI, MCP server, SQLite stores, and governed runtime.

From the repository root:

python scripts\validate_opencode_layer.py
opencode --agent erasmus

Inside OpenCode, use:

• /erasmus — route a request to the smallest applicable workflow;
• /erasmus-setup — configure a repository without overwriting local instructions;
• /erasmus-spec — write a bounded specification;
• /erasmus-implement — execute an approved plan;
• /erasmus-review — independently review an exact diff;
• /erasmus-research — produce a bounded cited research note;
• /erasmus-handoff — capture objective continuation state;
• /erasmus-doctor — inspect the local interaction and runtime surfaces without mutation.

To install the versioned agent, commands, and skills into the current user's global OpenCode configuration:

pwsh -NoProfile -File install\Install-ErasmusOpenCode.ps1 -Action Install -WhatIf
pwsh -NoProfile -File install\Install-ErasmusOpenCode.ps1 -Action Install

The installer does not copy project opencode.json, choose a provider/model, or write credentials. Use -Action Repair, Rollback, or Uninstall for the corresponding bounded operation.

Commands

Core runtime

• erasmus init — apply schema migrations and initialise the database
• erasmus status — table row counts and applied schema versions
• erasmus integrity — run PRAGMA integrity_check
• erasmus checkpoint — JSON-dump the latest committed checkpoint
• erasmus backup <dest> — hot-backup the database to a file
• erasmus restore <src> — restore from a backup file

Runtime and sleep

• erasmus runtime-validate <config> — validate a local endpoint configuration
• erasmus runtime-discover <config> — list models and advertised capabilities
• erasmus runtime-smoke <config> --prompt "hello" — run one bounded local session
• erasmus runtime-embed <config> "text" — request embeddings when advertised
• erasmus sleep — consolidate events into experience candidates (idempotent)
• erasmus sleep-report <run-id> — inspect classifications, reasons, and stage history
• erasmus sleep-decide <candidate-id> <decision> <target> <evidence-id> --actor ... --authority ... --reason ...

Missions

• erasmus mission-create --contract ... --title ... --objective ...
• erasmus mission-inspect <mission-id>
• erasmus mission-authorize <mission-id> --actor ... --evidence ... --approval-id ... --deny
• erasmus mission-run-one <mission-id>
• erasmus mission-pause|mission-resume|mission-cancel|mission-rollback <mission-id>

Governance, evidence, and cognition

• erasmus review --proposition "..." — run the 10th-Man review pathway for a proposition
• erasmus immune-process <event.json> --authority ...
• erasmus immune-inspect <incident-id> — inspect incidents and dormant state
• erasmus immune-agents — list inactive immune agents
• erasmus immune-false-positive <incident-id> <detector> --reason ... --actor ... --authority ...
• erasmus immune-retire <agent-id> --reason ... --actor ... --authority ...
• erasmus divergence-calibrate <baseline> <detector> <kind> <threshold> --actor ... --reason ...
• erasmus divergence-evaluate <fixtures> [--calibration ...] --authority ...
• erasmus divergence-downweight <calibration-id> --recommendation ... --actor ... --reason ... --authority ...
• erasmus skill-observe <candidate-id> <source-event-id> <evidence-id> ...
• erasmus skill-promote <candidate-id> <target> ...
• erasmus skill-draft <candidate-id> <document> ...
• erasmus skill-evaluate <candidate-id> <fixtures> ...
• erasmus skill-inspect <candidate-id>
• erasmus skill-export --actor ... --authority ...
• erasmus ledger-evidence-add ...
• erasmus ledger-propose <statement> <evidence-id> ...
• erasmus ledger-transition <proposition-id> <operation> <evidence-id> ...
• erasmus ledger-confidence <proposition-id> <confidence> <evidence-id> ...
• erasmus ledger-supersede <proposition-id> <replacement-id> <evidence-id> ...
• erasmus ledger-inspect <proposition-id> / ledger-query <proposition-id>

Graph and toolchain

• erasmus graph-validate <manifest> / graph-import <manifest> / graph-list
• erasmus graph-inspect <capability> / graph-plan <goal> [--authority ...] / graph-export <dest>
• erasmus bootstrap-validate <fixture> / bootstrap-resolve <fixture>
• erasmus toolchain-validate [document] [--manifests tools/manifests]
• erasmus tool-publisher-register <publishers>
• erasmus tool-register <manifest>
• erasmus tool-verify <manifest> <artifact>
• erasmus tool-install <manifest> <artifact>
• erasmus tool-list
• erasmus tool-inspect <tool-id> <version> <target>
• erasmus tool-activate|tool-deactivate|tool-quarantine|tool-revoke|tool-uninstall <tool-id> <version> <target>
• erasmus tool-execute <capability-id> <capability-version> <target> [args...]
• erasmus tool-health <tool-id> <version> <target> --authority ...
• erasmus tool-export <dest>

OKF foundry

• erasmus-foundry build <pdf-folder> <candidate-bundle> <runtime-config>
• erasmus-foundry validate <candidate-bundle> --write-report

The knowledge-system command namespace shown in docs/architecture/knowledge-system/OPERATOR_API_AND_RUNBOOK.md is design-only and not available as runtime CLI today.

The foundry deliberately stops at draft candidates: it does not write ledger propositions, verify concepts, reconcile contradictions, build indexes, or silently grant authority. See docs/architecture/okf-knowledge-foundry.md.

See docs/runbook-windows.md for PowerShell verification commands.

Status

This repository is an implementable experimental kernel. It is personal-first but contract-shaped so it can later evolve into isolated dyadic deployments.

See docs/DEVELOPMENT_TRACK.md for the locked phased architecture and scope boundaries.

Consolidated guarded repository missions

Guarded repository missions, navigation, and code integrations are now additive to the current kernel. Migration 19 owns their tables; published migrations 17 and 18 retain knowledge-policy and source-registry ownership. The explicit legacy migration converter recognizes the older branch-only mission migration 17, preserves its data and timestamp, and applies the missing mainline schema. The current bounded advisory worker runner remains authoritative: worker output does not grant execution authority. Run python -m pytest tests -q for migration, authority, provenance, and repository-mission regression coverage.

Consolidated guarded missions

Guarded repository missions are additive to the current bounded, advisory worker interface; worker responses retain authorization: none. Mainline migrations 17 and 18 keep their identity; mission tables use migration 19. Databases from the older mission branch are recognized and upgraded without losing their mission rows or recorded timestamps. Offline rollback removes only mission tables and migration 19. The mission schema is included in installed wheels.