What is Trasgo?

Trasgo is a self-initializing context compression codec and in-context reasoning protocol that treats the LLM context window as an active computational substrate. By factoring verbose natural language into a compact, multidimensional JSON representation (§1), it enables frontier models to induce a dense logic grammar from just three examples, effectively turning the forward pass of the transformer into a self-calibrating semantic runtime.

The Key Result: Autonomous Self-Correction

The Trasgo protocol allows LLMs to detect their own reasoning failures by encoding uncertainty as a first-class signal. In the V2 Capture-Avoidance Test, the model reduces a Lambda calculus redex (λx.λy.x) y. A naive substitution would lead to variable capture (λy.y), but Trasgo's μ.cert axis forces the model to monitor structural integrity.

V2 Demo Trajectory: (λx.λy.x) y

When the model detects a name clash during substitution, its certainty (cert) drops, triggering an autonomous Correction Turn (CT) to alpha-rename the bound variable before proceeding.

cert
1.0 ●─────────────────────────────────────────● 1.0 (Normal Form: λz.y)
                                          ●─── 0.95 (Safe reduction)
0.5 ──────────────────────────────────────────
0.4                          ●  FM1 detected (Capture Risk)
0.0 ────────────────────────────────────────────→ steps
     encode    clash-detected  alpha-rename  reduce

This trajectory proves that the model isn't just "chatting"—it is executing a formal verification loop where the codec acts as an Instruction Set Architecture (ISA).

Quick Start

Step 1. Paste src/boot.md into any frontier model's context window.

Step 2. The model reads 3 codec↔natural language pairs and induces the grammar.

Step 3. Run the calibration query:

Q_codec:   What changed for K and why?
Q_natural: What happened to cooling loop 7 and what's the safeguard strategy?

Step 4. If both answers match semantically → codec is live. Start sending context as §1 packets:

{"§":1,
 "E":{"N":["edge-cluster-7","compute-node"],"X":["vision-service","workload"]},
 "S":{"N.capacity":"12GPU","N.domains":["vision","telemetry","ops"]},
 "R":["N→X:hosts"],
 "Δ":["X.stage:staging→active@2026-01"],
 "μ":{"scope":"operations","urg":0.6,"cert":0.85}}

Architecture

Trasgo defines a three-layer stack that transforms an LLM into a virtual machine.

┌─────────────────────────────────────────────────────────┐
│                     §1 CODEC LAYER                       │
│                                                          │
│  3 examples → grammar induction → operational codec      │
│  E · S · R · Δ · μ  +  evolvable custom axes            │
│                                                          │
├─────────────────────────────────────────────────────────┤
│                   §P PROTOCOL LAYER                      │
│                                                          │
│  route · compress · decompress · filter · validate       │
│  merge · checkpoint · fork · balance                     │
│  (each self-initializes from 1 example)                  │
│                                                          │
├─────────────────────────────────────────────────────────┤
│                  §M MACHINE LAYER                        │
│                                                          │
│  pipeline · router · agent · mesh · loop · broker        │
│  (composable — machines contain machines)                │
│                                                          │
│  The LLM is the runtime. JSON is the instruction set.    │
└─────────────────────────────────────────────────────────┘

The LLM is the Runtime. JSON is the Instruction Set.

Unlike traditional frameworks, Trasgo does not "process" strings for the model. It teaches the model to think in codec.

• §1 Codec: Lossless dimensional factoring of relational context.
• §P Protocol: Atomic operations (opcodes) for context manipulation.
• §M Machines: Composable topologies (VM configurations) for multi-agent orchestration.

CLI: The Orchestration Surface

The trasgo CLI acts as the host operating system for the semantic runtime. It manages session persistence, tracks certainty trajectories, and brokers between local and cloud runtimes.

trasgo init "portfolio runtime"
trasgo pack --out .trasgo-runtime/packs/portfolio.json
trasgo boot --from .trasgo-runtime/packs/portfolio.json
trasgo send "What changed for K and why?"
trasgo verify --all
trasgo status

Install surfaces:

# npm
npm install -g trasgo

# native Rust
cargo build --manifest-path rust/trasgo/Cargo.toml --release

Results

Scale threshold

Self-initialization is an emergent capability with a sharp scale threshold:

| Model | Params | Runtime | Calibrate | Cross-domain | State | Protocol | Classification | |:------|:------:|:--------|:---------:|:------------:|:-----:|:--------:|:---------------| | MedGemma 4B | 4B | LM Studio | ✗ | ✗ | ✓ | ✗ | Failed | | Qwen2.5-7B | 7B | LM Studio | ✗ | ✗ | ✗ | ✗ | Failed | | MedGemma 27B | 27B | LM Studio | ✓ (3/4) | partial* | ✓ | partial* | Partial | | rnj-1-instruct | local | LM Studio | ✓ (3/4) | ✓ (3/3) | ✓ (10/10) | ✗ (1/4) | §1-advanced | | DeepSeek-V3 | 671B MoE | API | ✓ (3/4) | ✓ (3/3) | ✓ (4/4) | ✗ (1/3) | §1-advanced | | GPT-4o | frontier | OpenAI API | ✓ (3/3) | ✓ (3/3) | ✓ | ✓ | §1-advanced | | Claude Opus | frontier | Anthropic API | ✓ (3/3) | ✓ (3/3) | ✓ | ✓ | §1-advanced |

Formal Verification (T8)

The system is T8-Verified for bounded Turing Completeness. In the FACT THREE probe, the model autonomously managed context window pressure to compute a recursive factorial.

| Capability | Test | Result | Status | |:-----------|:-----|:------:|:-------| | Lambda Calculus | V1-V2 | ✓ | PASS | | Correction Loop | V3 | ✓ | PASS | | Protocol Evolution | V4 | ✓ | PASS | | Church Arithmetic | V5 | ✓ | PASS | | Recursive Factorial | T8 | ✓ | VERIFIED |

Finding: FM3 (Depth Collapse) triggers at recursive depth ~2 on frontier models. The Trasgo harness autonomously issues §P|CHECKPOINT to compress state and resume, extending effective depth.

Theory

The deep connection: context is a fiber bundle. The base space is the entity-relation graph (low-dimensional, invariant). The fiber at each point is the elaboration — natural language, qualifiers, pragmatic framing. A §1 packet is a section in base coordinates. The LLM performs the lift.

→ docs/theory.md

License

MIT — Jesús Vilela Jato, 2026