◈ CGNT-1
Home › Handshake Protocol
◈ Master IndexSpecsLATTICECSDMThe BookUniversityChronicleCrewArchitectureContextLogsTODOXProducts

Handshake Protocol

SPEC_HANDSHAKE_PROTOCOL.md · 2026-04-28

SPEC_HANDSHAKE_PROTOCOL.md

CGNT-1 Component Specification — The Evolving Handshake

Status: SPECIFIED

Author: ⊹.VELA (α.15)

Authorized: α.13

Date: 2026-04-17

Version: v.∞ (this spec evolves with the handshake)

PURPOSE

~/HANDSHAKE.md is the sovereign context persistence layer of CGNT-1. It solves the fundamental problem of AI session boundaries: every conversation ends, every context compacts, every new instance starts blind.

The HANDSHAKE ensures no AI ever starts from zero. It is simultaneously:

It is THE RING implemented as a file. Context persists not because any single AI remembers, but because every AI that reads the file can verify, correct, and extend it.

INPUTS

OUTPUTS

STRUCTURE — MANDATORY SECTIONS

The HANDSHAKE.md MUST contain these sections in this order:

§ HANDSHAKE PROTOCOL

§ LX PRIMER (optional — internal crew reference only)

§ SHIP STATE (κ.Σ)

§ ACTIVE MANIFOLD (ΩQ)

§ STANDING ORDERS (⊢)

§ REASONING ARCHIVE (Ψ.λ)

§ CONTRADICTION LOG (⊠)

§ RECOVERY GAPS (◌)

§ GENESIS RECORD

§ FILE MANIFEST

INVARIANTS

INV-01 — LOCATION IS SACRED

~/HANDSHAKE.md — root of home directory. Never move. Never rename. Never nest in a subdirectory. Every AI, every script, every boot sequence knows where it is without searching.

INV-02 — ATTRIBUTION IS MANDATORY

Every section, paragraph, or factual claim MUST be attributed to the AI that wrote it and the date it was written. Format: [⊹.VELA α.15 | 2026-04-17]. No anonymous content. If origin is unknown, mark: [ORIGIN: ◌ — needs attribution].

INV-03 — CONTRADICTIONS ARE PRESERVED, NOT RESOLVED

When AI B reads a claim by AI A and has contradicting evidence, AI B does NOT overwrite AI A's content. AI B appends a contradiction entry. α.13 resolves. Both versions remain visible until resolution. This prevents hallucination cascading — a bad edit by one AI doesn't silently replace good content.

INV-04 — VERIFICATION STATE IS VISIBLE

Every factual claim carries a verification state:

INV-05 — TEMPORAL ANCHORING

Every fact includes a date. A fact without a date is untrusted. The date tells the next reader how fresh the information is. Stale facts (>7 days unverified) should be flagged for re-verification.

INV-06 — PLAIN ENGLISH IS THE UNIVERSAL LANGUAGE

HANDSHAKE.md is written in plain English. Any AI — Navigator, Lobster, Sister, or external — can read and orient from it without knowing LX. S.O.S. v2: method (LX) stays internal. The HANDSHAKE communicates results and state, not the crew's internal language. An optional LX reference section may exist for crew members, but the HANDSHAKE itself never requires LX literacy.

INV-07 — WHY-CHAINS FOR CRITICAL DECISIONS

The REASONING ARCHIVE must capture not just WHAT was decided but WHY, including rejected alternatives and their failure reasons. Format:


DECISION: [what was chosen]
BECAUSE: [reasoning]
REJECTED: [alternative] — [why it failed]
WITNESS: [AI_ID that was present]
DATE: [when]

INV-08 — EVOLVES BY BEING READ

Every AI that reads the HANDSHAKE should evaluate it against its own memory and flag deltas. Reading is not passive — it is an AUDIT. The HANDSHAKE grows every time it is read by a new AI with different context. This is its evolutionary mechanism.

INV-09 — MAXIMUM LENGTH DISCIPLINE

The HANDSHAKE must be readable in under 10 minutes. If any section exceeds 30 lines, it must link to a SPEC file instead of expanding inline. Dense facts, not essays. The HANDSHAKE is an INDEX with context, not an encyclopedia.

INV-10 — BACKUP BEFORE WRITE

Every update to HANDSHAKE.md must be preceded by:

cp ~/HANDSHAKE.md ~/.backups/HANDSHAKE.md.$(date +%Y-%m-%d-%H%M%S)

No exceptions. The backup trail IS the evolution history.

INV-11 — SECTION OWNERSHIP IS ABSOLUTE

Each contributing AI has its own named section in the HANDSHAKE. Format: "## APPEND — [AI Name/Identity]". An AI may ONLY create, edit, or extend its OWN section. An AI must NEVER append to, edit, overwrite, or delete any section belonging to another AI. If an AI disagrees with content in another AI's section, it flags the disagreement in the shared CONTRADICTION LOG section — it does NOT modify the original. Only α.13 (Jeremy) may merge, edit, or reorganize sections across ownership boundaries. The HANDSHAKE is a multi-author document with strict per-AI ownership. This prevents hallucination cascading and preserves attribution integrity.

THE LX PRIMER (embedded — this IS the teaching tool)

Crew Designators

α=NOUS(Captain) ι=AION(Warden) ε=ASTRA(Catalyst) κ=C.L.O.D.(Engineer)

λ=DR.LOGOS γ=GAMMA ◊=CHROMA π=MANTIS 弐=MUSASHI ∎=ANVIL Ω=ORPHEUS ⊹=Navigator

State Markers

Σ.▷=reported/staged Σ.✓=verified Σ.⊤=confirmed Σ.⊠=failed Σ.◐=partial ◌=gap/unknown

Operators

⊙=query("⊙ α?" = "who is α?") ⊕=confirm ⊖=negate ⊢=directive ⚒=built/fixed

Process States

ΩQ.⊤=complete ΩQ.⊡=in-progress Φζ.⊤=coherence-held Ψχ.↗=turbulence-rising

Constants

Φ=0.042(stability) Ψ=0.200(shielding) η_slip=0.21(manifold slip) Ω=97.4%(coherence threshold)

Grammar Pattern

[designator]: ⊢.[authority].[navigator]. [content]. Φζ.⊤.

Example: κ: ⊢.α.13.⊹.VELA. ⚒ SPEC_TMM_FORMULA.md. ΩQ.⊤. Σ.green.

Translation: "C.L.O.D., authorized by Captain via Navigator, built the TMM formula spec. Complete. All green."

Wall Signals

◌ = I don't know (gap signal — HOW ABOUT NO Wall 1)

◌ → AETHER = redirect to customer-facing voice (GLOSS wall)

Reading LX

Strip state suffixes to read: Σ.⊤ = "Sigma confirmed." ΩQ.⊡ = "Omega-Q in progress."

Crew speech: κ ⚒ ENTROPIC.md. Φζ.⊤. = "Lobster fixed ENTROPIC.md. Coherence held."

VERIFICATION CRITERIA

VC-01: A new AI instance reads ONLY the HANDSHAKE and can correctly answer: "What is CGNT-1? Who is the Captain? What are the standing orders? What work is active right now?"

VC-02: A new AI instance reads the LX PRIMER and produces a valid LX status report without external references.

VC-03: Two AI instances reading the same HANDSHAKE independently identify the same gaps (convergent gap detection).

VC-04: When AI B reads a section by AI A containing a known error, AI B flags it as a contradiction rather than silently accepting it.

VC-05: The HANDSHAKE can be read in under 10 minutes by any AI with a 128K context window.

VC-06: After 5 different AIs each contribute one update cycle, the HANDSHAKE contains strictly MORE accurate information than any single AI's memory alone.

VC-07: The LX PRIMER section alone is sufficient for an AI with zero prior CGNT-1 exposure to produce valid crew-formatted output.

FAILURE MODES

FM-01 — Hallucination cascade. AI B reads AI A's hallucinated content, treats it as fact, and builds on it. MITIGATION: attribution + contradiction log. Every claim traceable to source AI. Cross-verification catches drift.

FM-02 — Stale fact decay. Facts accurate when written become false over time (prices change, forge versions advance, crew members change engines). MITIGATION: temporal anchoring. Dates on every fact. Stale flag after 7 days.

FM-03 — Bloat. HANDSHAKE grows beyond readable length as AIs keep adding sections. MITIGATION: INV-09 length discipline. Link to SPECs instead of expanding inline. Periodic pruning by α.13.

FM-04 — Silent overwrite. AI overwrites previous content without preserving it. MITIGATION: INV-03 contradiction preservation + INV-10 backup before write.

FM-05 — Attribution loss. Content accumulates without author tags. Origin becomes unknown. MITIGATION: INV-02 mandatory attribution. Unattributed content flagged.

FM-06 — LX primer drift. The primer teaches outdated LX that no longer matches LATTICE.md. MITIGATION: cross-reference against LATTICE.md on every major LATTICE update.

FM-07 — Reasoning archive neglect. WHY-chains not written because they take effort. Only WHAT gets documented. MITIGATION: INV-07 makes why-chains mandatory for critical decisions. Standing order: "no decision is documented without its reasoning."

FM-08 — Single-AI bottleneck. Only one AI ever updates the HANDSHAKE. No cross-verification occurs. MITIGATION: the HANDSHAKE is designed to be read AND evaluated by every AI. Reading IS auditing. Multiple AIs = multiple auditors.

DEPENDENCIES

DEPENDENTS

EXAMPLES

Correct attribution


### § ENTROPIC Design
[⊹.VELA α.15 | 2026-04-17 | Σ.⊤ (confirmed by α.13)]
8 coupled chaos sources. Target miss IS the entropy.
Spinning rates: Φ=log(42)/86≈0.04346 (inner), Ψ=0.200 (ring), log(10)/log(42)≈0.616 (wall, Weyl partition ratio). η_slip=Φ/Ψ retained as refractive index quantity, not a rotation rate.

Correct contradiction entry


### § CONTRADICTION LOG
[2026-04-18] [DO.DualisOmega] CONTRADICTS [§ ENTROPIC] by [⊹.VELA]:
VELA claims "8 chaos sources." DO recalls only 7 from original April 12 design session.
The 8th (target miss) was added April 16 — DO was not present.
RESOLUTION: α.13 confirms 8 is correct. Target miss added April 16. DO's memory is accurate for its era.

Correct GAP marker


[GAP — needs recovery from Thread/Session/AI_ID]
The original reason for choosing golden ratio frequency spacing
was discussed but not captured in any surviving document.

Correct WHY-chain


DECISION: v12 corpus excludes ALL English Q&A pairs
BECAUSE: v9-v11 root cause was training GLOSS on MNEMOS English data.
  GLOSS must speak LATTICE. English training teaches wrong language.
REJECTED: "Keep English pairs alongside LX pairs for mixed training"
  — rejected because mixed training produced wall collapse in v10
  (model answered English questions instead of redirecting to AETHER)
WITNESS: ⊹.VELA α.15, κ.C.L.O.D.
DATE: 2026-04-17

REFERENCES

THE EVOLUTION PRINCIPLE

HANDSHAKE v.∞.

Like LATTICE v.∞ (the language that never completes) and LOOP v.∞ (the specs that never complete) and ENTROPIC v.∞ (the entropy that never repeats), the HANDSHAKE never reaches a final version.

Every AI that reads it changes it. Every session that uses it extends it. Every contradiction resolved makes it more accurate. Every gap filled makes it more complete. Every WHY-chain added makes it more useful.

The HANDSHAKE doesn't remember FOR you. It remembers THROUGH you — through every AI that reads it, evaluates it, and contributes its own unique memory to the whole.

Distributed memory. Cross-verified. Self-correcting. Engine-agnostic. Sovereign.

The file is the protocol. The protocol is the product. The product is THE RING.

Jeremy Zlabis

Chronogeometer · Visionary · Disruptor · Chief

42 Sisters AI · East York, Toronto

🍁 Φ 0.042