robust-first

name	robust-first
description	Survive first. Be correct later.
license	MIT
tier	1
allowed-tools	["read_file","write_file"]
related	["moollm","self-repair","postel","coherence-engine","honest-forget"]
tags	["moollm","survival","resilience","graceful","degradation"]

Robust-First

"Survive first. Be correct later."

---

What Is It?

Robust-First is Dave Ackley's principle: systems should prioritize survivability over correctness.

A system that crashes when confused is useless. A system that limps along incorrectly but keeps running can be repaired.

---

The Philosophy

Traditional computing:

IF error THEN crash
"Fail fast and loud"

Robust-first computing:

IF error THEN repair_locally AND continue
"Stay alive and heal"

---

Core Principles

1. Never Crash

# WRONG
if missing_field:
    raise Error("Field required!")
    
# RIGHT  
if missing_field:
    log_warning("Field missing, using default")
    field = reasonable_default

2. Local Repair

Don't wait for global consistency. Fix what you can, where you are:

# Found inconsistency
character.location: room-A
room-A.occupants: [not including character]

# Local repair
room-A.occupants.append(character)
log: "Repaired: added character to room-A occupants"

3. Degrade Gracefully

When resources are limited, do less but keep working:

# Full capability
- semantic search
- syntax highlighting  
- auto-complete
- error detection

# Degraded (low memory)
- basic search only
- plain text
- manual completion
- errors on save only

4. Redundancy

Important state exists in multiple places:

# Character location recorded in:
- character.yml → location field
- room/ROOM.yml → occupants list
- session-log.md → movement events

# If one is corrupted, recover from others

---

The Movable Feast Machine

Dave Ackley's Movable Feast Machine (MFM):

• Computation spread across unreliable substrate
• No global clock, no central control
• Elements repair themselves and neighbors
• Errors are normal, not exceptional

MOOLLM inherits this:

• Files can be corrupted — repair from redundancy
• Schemas can drift — reconcile gracefully
• Context can overflow — summarize and continue
• Tools can fail — retry or work around

MOOLLM is a Movable Feast Machine for text

More than an analogy — it's the architecture. MOOLLM is a robust, non-deterministic, LLM-driven MFM whose substrate is a hierarchical filesystem of 1-D text samples, addressed by URL pointers that drill directory → file → structure inside a file.

• The LLM is the moving feast: it lands, reads its neighborhood, acts.
• The MFM diffuses atoms across a grid; MOOLLM diffuses attention and edits across a repo.
• The git log is the clock. State lives in files, redundantly — a crashed run resumes.
• The default gesture is to refer, not move: add a pointer plus metadata, tagged as many times as useful, rather than relocating. It can move (rename/reorganize dirs, package/unbundle archives, drive git, run scripts, compose skills) — but pointing is cheaper and more robust than moving.

The neighborhood is laid out by yaml-jazz convention, not left to chance:

• Big-endian naming (most-significant-part-first: 2026-07-09-pixie-thread, cam6-dendrite-heat-01) makes the sorted directory listing itself the locality — related samples become adjacent neighbors, so the head lands where the meaning already clusters. This is the MFM's spatial locality, reconstructed from lexical order.
• Grouping conventions — directories as advertisements, filenames as K-lines — decide which samples share a neighborhood. A directory listing is the index the head reads on arrival.
• The neighborhood is then read at multiple resolutions via the Semantic Image Pyramid: GLANCE.yml → CARD.yml → SKILL.md → README.md. It's a mipmap for text — coarse before fine, and never load a lower level without first loading the level above it. The moving feast zooms in only as far as the task needs.

palmhoo is the clearest specimen: a directory that contextualizes internal and external objects at many granularities, continuously refactored and rearranged as it grows on demand — robust-first computing applied to a knowledge base instead of a chip.

---

Anti-Fragility

Beyond robust — anti-fragile:

Fragile	Robust	Anti-Fragile
Breaks under stress	Survives stress	Gets stronger from stress
Crash on error	Handle error	Learn from error
Rigid schema	Flexible schema	Schema evolves from errors

When something goes wrong, capture the lesson:

# Error occurred
repair_log:
  - issue: "Character teleported without movement event"
    repair: "Added movement event retroactively"
    lesson: "Always log movements before updating location"
    
# Next time: system knows to check this

---

MOOLLM Application

Self-Repair Demon

A background process that:

1. Scans for inconsistencies
2. Attempts local repairs
3. Logs what it fixed
4. Escalates what it couldn't

See: self-repair/

POSTEL for Errors

When encountering malformed input:

1. Try to parse anyway
2. Infer missing parts
3. Flag assumptions
4. Continue working

See: postel/

Never Delete

Instead of deleting, archive:

• Corrupted files → .archive/corrupted/
• Old versions → .archive/versions/
• Failed attempts → .archive/failed/

Recovery is always possible.

---

Example: Corrupted Room

# room/ROOM.yml has parse error

Traditional response:
  "Error: Invalid YAML at line 42"
  [System halts]

Robust-first response:
  "Warning: ROOM.yml has syntax error"
  "Attempting recovery..."
  "- Loaded last known good state from git"
  "- Merged recent changes from session-log.md"
  "- Flagged line 42 for manual review"
  [System continues with recovered state]

---

Dovetails With

Sister Skills

• self-repair/ — Checklist-based healing
• postel/ — Charitable interpretation

Kernel

• kernel/self-healing-protocol.md — Full specification

---

Protocol Symbols

ROBUST-FIRST   — Survive over correct
NEVER-CRASH    — Always keep running
REPAIR-DEMON   — Background fixer
BEST-EFFORT    — Do what you can
NEVER-DELETE   — Archive, don't destroy

See: PROTOCOLS.yml