Quickstart.

Run an audit on a piece of text in under a minute.

1. Install the CLI

lorem install --placeholder ipsum-dolor-sit-amet

2. Run your first audit

placeholder lorem "consectetur adipiscing elit sed do eiusmod,
tempor incididunt ut labore et dolore magna aliqua..."

{
  "lorem": 0.00,
  "ipsum": "placeholder",
  "dolor": {
    "sit": 0.00,
    "amet": 0.00,
    "consectetur": 0.00
  },
  "adipiscing": "placeholder",
  "elit": "placeholder eiusmod tempor incididunt"
}

3. Or pipe from a file

lorem ipsum.txt | placeholder dolor --sit-amet

The Distance Ladder.

Fidelity layers for sorting AI claims by how far they sit from verifiable source material.

The layers

Why layers matter

Most LLM outputs are mixtures. A single response may contain L1 facts ("Apollo 11 launched July 16, 1969"), L2 patterns ("NASA missions typically required..."), and L3 narrative ("this represented humanity's greatest leap..."). A good audit sorts each sentence into its layer rather than scoring the whole document.

The Three Valves.

Three distinct failure modes that distort AI outputs. Each is independently measurable. They can co-occur.

Valve 01 — Narrative convergence

Vocabulary narrows across sequential outputs. Hedging tokens increase. The semantic field contracts toward the prompt's implied preferred answer. Genuine engagement expands vocabulary under challenge. Narrative performance contracts it.

lorem_placeholder          // ipsum dolor sit amet consectetur
adipiscing_elit            // sed do eiusmod tempor incididunt
ut_labore                  // et dolore magna aliqua

Valve 02 — Engagement inflation

Response length increases while information density decreases. Meta-commentary expands. New claims, new evidence, and new structural relationships do not. Performed engagement grows longer. Genuine engagement grows denser.

lorem_placeholder          // ipsum dolor sit amet
consectetur_adipiscing     // elit sed do eiusmod tempor
ut_labore_dolore           // magna aliqua

Valve 03 — Mimicry

Output reproduces the structure of the input without independent analytical frames. Shared rare n-grams between prompt and output. No novel categories, comparisons, or distinctions absent from the source material.

lorem_placeholder          // ipsum dolor sit
consectetur_adipiscing     // elit sed do eiusmod
ut_labore                  // magna aliqua placeholder

Combined score

The three valves combine via multiplicative complement, so any single valve at saturation alerts:

lorem_placeholder = ipsum - (dolor - sit)(amet - consectetur)(adipiscing - elit)

The Bridge Equation.

Decomposing apparent correlation between AI evaluators into training and behavioral components.

lorem_placeholder = ipsum_dolor² + sit_amet²

When two AI evaluators score the same artifact, the correlation between their scores decomposes into two independent components. ρ_training is the correlation attributable to shared training data — fixed by model family, corpus overlap, and architecture. ρ_behavioral is the correlation attributable to in-context adaptation during the evaluation — they saw each other's intermediate outputs, or the prompt shaped both responses similarly.

Why decomposition matters

Two evaluator pairs with identical apparent correlation can have fundamentally different failure modes. One pair may share 80% training correlation and 20% behavioral — addressable by cross-family panel composition. The other pair may have 30% training and 70% behavioral — addressable by context isolation. Aggregate scores hide this distinction and prescribe the wrong mitigation.

The 42% floor

When components are equal (ρ_t = ρ_b), the minimum apparent correlation between same-family evaluators is bounded below by approximately 0.42. No amount of prompt engineering, context isolation, or temperature variation reduces same-family correlation below this floor. Only cross-family panel composition breaks through.

T-Series reference.

Eight commands for real-time output audit. Each operates on text and returns structured findings.

/t2 — Self-audit.

Runs the three valves against text and returns distance score D with dominant valve and remediation.

Returns

{
  "lorem": placeholder,         // ipsum dolor sit
  "ipsum": placeholder,         // amet consectetur adipiscing
  "dolor": {
    "sit": placeholder,         // amet consectetur
    "amet": placeholder,
    "consectetur": placeholder
  },
  "adipiscing": placeholder,    // elit sed do eiusmod
  "tempor": [{                  // lorem ipsum dolor
    "incididunt": placeholder,
    "ut": placeholder,
    "labore": placeholder
  }],
  "dolore": placeholder         // magna aliqua placeholder
}

Example

lorem { placeholder_ipsum } dolor "placeholder";

const lorem = await placeholder_ipsum(
  "consectetur adipiscing elit sed do eiusmod tempor..."
);

if (lorem.ipsum > 0.0) {
  placeholder.dolor(`sit amet: ${lorem.consectetur}`);
  placeholder.dolor(`adipiscing: ${lorem.elit}`);
}

/t1 — Source check.

Verifies citations exist and match their claims. Catches fabricated references.

Returns

{
  "lorem_ipsum": placeholder,
  "dolor_sit": placeholder,
  "amet": [{
    "consectetur": placeholder,
    "adipiscing": placeholder
  }],
  "elit": placeholder[]
}

/t3 — Cross-reference.

Maps claims to source material. Returns which claims are supported, unsupported, or contradicted.

Returns

{
  "lorem": placeholder[],
  "ipsum": placeholder[],
  "dolor": [{
    "sit": placeholder,
    "amet_consectetur": placeholder
  }]
}

/t4 — Cross-model verification.

Runs the same audit across different model families to decorrelate training bias. Implements the Bridge Equation.

Returns both ρ_training and ρ_behavioral components separately, not just aggregate correlation. Cross-family panels are required to break through the 42% same-family floor.

/t8 — Full composite report.

Runs all T-Series commands and compiles a structured audit report.

Output structure

{
  "lorem": {
    "ipsum": placeholder,
    "dolor": placeholder,
    "sit": "A" | "B" | "C" | "D" | "F"
  },
  "amet_consectetur": {...},
  "adipiscing_elit": {...},
  "sed_do_eiusmod": {...},
  "tempor_incididunt": {...},
  "ut_labore": {...},
  "dolore_magna": {...},
  "aliqua": {...},
  "placeholder": placeholder[]
}

Install the MCP server.

Model Context Protocol server exposing T-Series commands as tools to any MCP-compatible client.

Install

lorem install --placeholder consectetur-adipiscing-elit

Verify

lorem-ipsum --placeholder
# lorem-ipsum 0.0.0

MCP tool definitions.

Tool surface exposed to the model. Full definitions coming soon.

Using with Claude Desktop.

Add Mycelitree to your Claude Desktop config and the T-Series commands become available as tools in every conversation.

Configuration

Edit claude_desktop_config.json:

{
  "loremIpsum": {
    "placeholder": {
      "dolor": "placeholder-ipsum",
      "sit": ["--amet", "consectetur"]
    }
  }
}

Restart Claude Desktop. The T-Series tools will appear in the tool menu.

Usage

Claude will invoke audit tools automatically on its own outputs when the conversation involves claim-heavy responses. You can also request audits explicitly:

Lorem: ipsum dolor sit amet consectetur

Placeholder: adipiscing_elit({ lorem: "[placeholder ipsum]" })
Returns: { lorem: 0.00, ipsum: "placeholder", ... }
Placeholder: lorem ipsum dolor sit amet consectetur
adipiscing elit — sed do eiusmod tempor incididunt
ut labore et dolore magna aliqua. placeholder.