protocol

# PROTOCOL > **"Protocol names ARE K-lines."** MOOLLM protocol symbols are Minsky's K-lines made concrete and greppable. --- ## Protocol Names = K-Lines A **K-line** (Minsky, "K-Lines: A Theory of Memory", Cognitive Science 1980; *Society of Mind*, 1986) is a mental structure that, when activated, reactivates an entire constellation of knowledge, skills, and associations. **Protocol names are exactly this:** ``` YAML-JAZZ ← Activates: semantic comments, jazz interpretation, LLM-as-musician POSTEL ← Activates: charitable interpretation, robustness, collaboration PLAY-LEARN-LIFT ← Activates: exploration, pattern-finding, crystallization, sharing COHERENCE-ENGINE ← Activates: consistency, constraints, orchestration, referee, simulation, DM ``` Type the name → activate the tradition → invoke the behavior. The UPPER-CASE-HYPHENATED format makes these K-lines **findable**. --- ## Activation: scene and stance, not only a pointer A K-line is **not** only a compact reference to a concept. In Minsky’s account, activating a K-line tends to **re-instantiate a configuration** that worked before: which sub-skills, tone, and moves to bring online. Henry Minsky put it in terms of a **whole state of mind** (or “scene”) bound to a name—closer to a **re-enterable closure** than to a bare pointer. That matters in MOOLLM because: - **Shared index** — `PROTOCOLS.yml`, `skills/INDEX.yml`, and prior turns give the same symbol the same load-bearing role. Without that, a name is noise. - **Embodied triggers** — A photo, a warning sign, a vocal cue can act like a K-line if the culture around it is shared: they snap attention to a **lived** situation, not a map coordinate. **Linguistic analogy (not Minsky’s wording, but useful):** In some languages, *am Bahnhof Zoo* vs bare *Bahnhof Zoo* shifts from bare reference toward **situated, insider scene**—the way *the 405* in Los Angeles is not “a highway” but **that** corridor of life. Same mechanism in spirit: **name → thick situation**, not only **name → fact**. **Optional cultural example:** Nina Hagen’s *Berlin Zoo* (1982) is often read as **performative** activation—noise, animal theatre, West Berlin grit—not as zoology. The track works as a **charged name** for a subculture-tinted “zoo” of people and noise. A zoo sign about a lion spraying urine through bars is thematically on that wavelength: blunt, bodily, public—**scene-first**, not taxonomy-first. **Contrast:** **Frames** (Minsky) describe the **shape** of a situation (slots, defaults). **Drescher schemas** (see `schema-mechanism`) describe **learned causality** (Context → Action → Result). **Schemapedia** (see `schema`, `schemas/registry.yml`) indexes interchange, causal, frames, activation, relational SQL/SQLite, and SoM—each family with its own delegate skills. **Protocol symbols** here describe **which competence bundle to load**. See `knowledge-frames` for the full map. --- ## The 1979 memo mechanics The memo (AI Memo 516; full OCR text in [sources/aim-516-k-lines-1979-ocr.txt](sources/aim-516-k-lines-1979-ocr.txt)) gives the machinery, and every piece maps onto context engineering: - **K-node assignment + K-line attachment.** A memorable event mints a K-node; its K-line makes an *excitatory attachment to every currently active P-agent*. **A K-line is a stored attention mask over a P-pyramid**: which nodes to activate and how much. Reactivation makes P "virtually hallucinate" the old partial state. - **The level-band principle** (Minsky's "most important idea"). Attach only to an intermediate band: too low *"would impose false perceptions and conceal the real details of the present problem"*; too high *"would make us hallucinate the present problem as already solved."* Restore the middle of the pyramid — structure — leaving the top free for the current goal and the bottom free for current facts. Context-loading policy, 1979. - **Weak fringes = frame defaults.** Connections at the band's edges are weak, so they lose cross-exclusion competition to anything the present situation asserts. Defaults that yield to evidence, implemented as low attention weights. - **Cross-exclusion.** Agents cluster in mutually inhibiting groups — radio buttons. Force one on and it suppresses its rivals: cheap state reset, built-in short-term memory (persistence), and *dispositions* — a P-pyramid's repertory of pre-activated subsets, i.e. ways of seeing (the Necker cube flip). Conflict inside a group makes the whole group drop out and default upward: don't average contradictions, zoom out. - **K-recursion.** New K-lines attach to currently active K-nodes, not raw P-agents: *"new memories are composed mainly of ingredients from earlier memories."* Vocabulary compounds — each named view becomes a component of the next. - **Crossbar via sparse coding.** K-lines don't need full wiring; random small subsets of a shared line bundle suffice (Mooers zatocoding 1956, Willshaw 1969) — hashing and Bloom filters as neuroanatomy. Plus the Society answer: most agents never need to talk. The P-pyramid — the agent-scoped hierarchical view a K-line masks over — has its own design doc: [designs/P-PYRAMID.md](../../designs/P-PYRAMID.md). --- ## Naming Convention | Rule | Example | Why | |------|---------|-----| | ALL CAPS | `YAML-JAZZ` not `yaml-jazz` | Stands out in prose | | Hyphen-separated | `PLAY-LEARN-LIFT` not `PLAY_LEARN_LIFT` | Shell-friendly, readable | | No colons | `POSTEL` not `WP:POSTEL` | Cleaner than Wikipedia | | Semantic | `ROBUST-FIRST` not `RF` | Self-documenting | --- ## Why This Convention? ### 1. Grep-Friendly ```bash # Find all uses of a protocol grep -r "YAML-JAZZ" . # Find all protocol definitions grep -r "^[A-Z-]*:" PROTOCOLS.yml # Find all mentions in docs grep -rn "POSTEL\|PLAY-LEARN-LIFT" skills/ ``` ### 2. Vector-Search Friendly The UPPER-CASE format creates **distinct embeddings**. When you search for "YAML-JAZZ", you get: - The protocol definition - Files that reference it - Examples that demonstrate it Not general YAML documentation. ### 3. Human-Scannable In prose, protocols pop: > When parsing user input, apply POSTEL (charitable interpretation). > If the command is ambiguous, use YAML-JAZZ to infer intent. You can't miss them. --- ## Protocol Layers ### Kernel Protocols (Low-Level) Defined in `kernel/`, fundamental to operation: | Protocol | Purpose | Location | |----------|---------|----------| | `FILES-AS-STATE` | Everything is files | kernel/constitution-core.md | | `WHY-REQUIRED` | Tool calls need reasons | kernel/tool-calling-protocol.md | | `APPEND-ONLY` | Never modify logs | kernel/constitution-template.md | | `MINIMAL-DIFF` | Change only what's needed | kernel/constitution-core.md | ### Skill Protocols (Mid-Level) Defined in `skills/`, behavioral patterns: | Protocol | Purpose | Location | |----------|---------|----------| | `PLAY-LEARN-LIFT` | The methodology | skills/play-learn-lift/ | | `SOUL-CHAT` | Everything speaks | skills/soul-chat/ | | `POSTEL` | Charitable interpretation | skills/postel/ | | `YAML-JAZZ` | Comments carry meaning | skills/yaml-jazz/ | ### Ad-Hoc Protocols (Emergent) Mentioned in conversations, examples, documents: ```markdown # In a soul-chat document We should apply KITCHEN-RULES here — no running, always clean up, respect the fridge contents. # In an adventure room This room follows GRUE-SAFE protocol — magically lit. ``` These don't need formal definitions. **Naming them makes them real.** --- ## Where Protocols Live ### Central Registry **[PROTOCOLS.yml](../../PROTOCOLS.yml)** — The master list: ```yaml YAML-JAZZ: meaning: "Comments carry meaning. Data is interpreted." location: kernel/constitution-core.md §3 invoke_when: "Writing or reading YAML with semantic intent" ``` ### Embedded in Files Protocols can be defined inline where they're used: ```yaml # In a room's ROOM.yml protocols: DARK-ROOM: meaning: "No light source = grue danger" applies_when: "player enters without lit lamp" ``` ### Referenced in Prose Just mention them: > This adventure follows REINCARNATION protocol — > death returns you to start, inventory preserved. The name IS the definition until someone formalizes it. --- ## Protocol Lifecycle ``` Mentioned → Documented → Formalized → Skill ↓ ↓ ↓ ↓ "Let's Entry in Directory Full try PROTOCOLS in PROTOTYPE FOOD-CHAIN" .yml skills/ + templates ``` Most protocols stay informal. Only crystallize when needed. --- ## K-REFs: File Pointers as K-Lines A **K-REF** is a K-line that points to a specific location in a file with optional metadata: ``` /path/to/file.md:42-58 # section_type | preview text ``` ### K-REF Format ``` PATH[:LINE[-END]][#ANCHOR][?SEARCH] # TYPE [LABEL] SEVERITY - DESCRIPTION EXCERPT or MASKED_VALUE ``` Components: - **PATH** — Absolute file path (if no line number, refers to whole file) - **LINE** — Line number (1-based), optionally with column `:LINE:COL`. Omit for whole file. - **END** — Optional end line or column for ranges - **#ANCHOR** — Jump to heading/section (like URL fragments) - **?SEARCH** — Find first match of pattern (like URL query) - **TYPE** — What was found (uuid, secret, user_prompt, tool_call, etc.) - **LABEL** — Optional redact label like `[SSH_KEY]` - **SEVERITY** — Icon: 🔴 critical, 🟠 high, 🟡 medium, 🔵 low, ℹ️ info - **DESCRIPTION** — Human-readable explanation - **EXCERPT** — Surrounding context or masked value ### Location Specifiers | Format | Meaning | Example | |--------|---------|---------| | `/path/file` | Whole file | `/etc/config.yml` | | `/path/file:42` | Line 42 | `/src/main.py:42` | | `/path/file:42-58` | Lines 42-58 | `/src/main.py:42-58` | | `/path/file:42:10-46` | Line 42, cols 10-46 | `/src/main.py:42:10-46` | | `/path/file#section` | Anchor/heading | `/docs/API.md#authentication` | | `/path/file?pattern` | Search for pattern | `/src/main.py?def process_data` | | `/path/file#section?pattern` | Section + search | `/docs/API.md#errors?401` | ### Anchor Types ``` # Markdown headings (slugified) /README.md#installation /SKILL.md#k-ref-format # YAML keys (dot-path) /config.yml#server.port /CARD.yml#methods.K-REF # Code symbols (function/class) /main.py#def:process_data /main.py#class:AuditRunner # HTML-style id /page.html#footer ``` ### Search Patterns ``` # Literal string (default) /file.py?def main # Regex (prefix with ~) /file.py?~def\s+\w+\( # First match of pattern /transcript.txt?user: # Combine with line context /file.py?def main # function - Jump to main() ``` ### Examples ``` /path/to/config.yml # config - Whole file reference /path/to/SKILL.md#k-refs # section - Jump to K-REFs section /path/to/main.py?def audit # search - Find audit function /path/transcript.txt:91:12-48 # uuid ℹ️ fe18ce96-5200-4e15-acd3-190695de6295 /path/transcript.txt:7528:18-45 # private_key ([PRIVATE_KEY]) 🔴 - Private key header ********** ******* ******** /path/config.json:15 # password ([PASSWORD]) 🟠 - Database password ************ ``` ### Images Work Too! Cursor (and other orchestrators) can read absolute file paths directly — **INCLUDING IMAGES**! ``` /Users/me/Screenshots/error.png # screenshot - Analyze this error /tmp/architecture-diagram.jpg # diagram - Explain this system /path/to/chart.png # data - What trends do you see? ``` No line numbers for images (obviously). Just the absolute path. **This is powerful:** - Point to any image on disk - Cursor reads and analyzes it - Works with screenshots, diagrams, charts, photos - Supports: jpeg, png, gif, webp **Security note:** Cursor can read ANYWHERE on your disk, not just the workspace. Powerful, but be aware. ### Sister Script Methodology **Problem**: LLMs have limited context windows. Dumping entire files wastes tokens. **Solution**: Sister scripts emit K-REFs (pointers) instead of full content: ```yaml # Sister script → K-REF → LLM flow kref_flow: source: "cursor-mirror (sister script)" produces: "K-REFs (pointers)" consumer: "LLM reads only what it needs" ``` 1. **Scan** — Script searches large data (transcripts, databases) 2. **Emit** — Output K-REFs pointing to relevant sections 3. **Read** — LLM selectively reads only the referenced ranges This is **reference by pointer, not by value** — parsimonious context. ### K-REF Output Modes | Mode | Command | Use Case | |------|---------|----------| | K-REF | `--emit-kref` | Human/LLM navigation | | REDACT | `--emit-redact` | Pipe to masking tool | | JSON | `--json` | Programmatic processing |