TC
← All Research
PRD – MICL Sheet v1.0
ArchitectureConcept EncodingFeatured

PRD – MICL Sheet v1.0

**PRD – MICL Sheet v1.0**

2025-11-3014 min read2,295 words
PRD – MICL Sheet v1.0 Mechanistic Interpretability Concept Language  Date: 2025-11-30 Status: Draft – Ready for Implementation Author: Trent Carter (+ assistant gpt5.1) Supersedes: Concept Glyph v0.x line (kept as historical / engineering-only) #Patentable  

0. Intent This PRD defines MICL Sheets: a dual-layer representation for concepts that:
  • Provide a human-first, mechanistically structured view of a concept using familiar engineering artifacts (UML-style diagrams, tables, icon scenarios, plain text).
  • Embed the canonical high-dimensional concept vector (e.g., 768D) into a barcode frame with forward error correction (FEC) so the sheet itself is a scannable carrier of the concept’s latent representation.
  • We explicitly abandon the prior circular “Concept Glyph” dashboard as the primary UX. It can survive as an internal engineering visualization, but MICL Sheets are the new, user-facing standard.

    1. Scope 1.1 In-Scope

    MICL Sheet: a single image/canvas that represents one requirement-level concept.

    • Two orthogonal layers:

  • Vector Layer – 768D (or higher) concept vector with FEC, encoded as a barcode frame around the content.
  • Human Sheet Layer – a set of zero or more sections (summary, UML diagrams, tables, icon scenarios, etc.) that explain the concept.
  • • Sheets must:

    • Be interpretable by a human engineer without any tools.

    • Be scannable by a model to recover the concept vector even from sub-optimal photos/screenshots (via FEC).

    1.2 Out-of-Scope (v1.0)

    • Encoding implementation-specific details (exact PAS wiring, deployment configs) into the vector or barcode. Those are separate concepts and get their own sheets.

    • Multi-concept “scene” sheets (composite views) – v1.0 is strictly one concept per MICL Sheet.

    • Live/animated layouts. v1.0 is static; animations or multi-frame GIFs are a future extension.

    2. Terminology

    Concept Vector: High-dimensional embedding v \in \mathbb{R}^D for a requirement-level concept.

    MICL Sheet: The whole artifact (image + barcode + human sections).

    Section: A distinct human-readable panel on the sheet (summary, UML diagram, etc.).

    Section Type: Schema label for a section, e.g. summary, uml_diagram, table, scenario_icons.

    Renderer: Any system that turns a MICL JSON spec into pixels (or reads pixels back into structure).

    3. High-Level Requirements 3.1 Vector Layer Requirements
  • The MICL Sheet SHALL encode a single concept vector v \in \mathbb{R}^D (default D = 768) in a border barcode around the content area.
  • The vector encoding SHALL support forward error correction (FEC) such that:
  • • A scanner can reconstruct v with acceptable fidelity even if:

    • Up to X% of tiles are corrupted/occluded (X configurable; v1.0 target ≈ 10–20%).

    • Image resolution is reduced (e.g., phone photo) or slightly blurred.

  • The encoding SHALL be versioned and self-describing, so future decoders know which scheme and FEC parameters were used.
  • 3.2 Human Sheet Requirements
  • The MICL Sheet SHALL be human-first: all interior content is designed for human comprehension (engineers, non-experts) using:
  • • Plain-language summary text,

    • UML-style diagrams and flowcharts,

    • Tables of properties/interfaces,

    • Icon-based scenarios / cartoons.

  • The human sheet SHALL be extensible:
  • • Each section is typed and optional.

    • Unknown section types MUST be safely ignored by renderers without breaking layout.

  • The layout SHALL be responsive:
  • • Work on portrait (phone) and landscape (desktop/tablet).

    • Reflow sections (stack vs grid) based on aspect ratio and available space.

  • The minimum “useful” MICL Sheet SHALL be:
  • • A valid barcode frame + a single summary section.

    • All other sections are additive.

    4. Vector Barcode Frame Specification (v1.0) 4.1 Geometry

    • MICL Sheet canvas can be any resolution; recommended baselines:

    • Phone: ~1080×1920 (portrait)

    • Desktop: ~2160×3840 or 4K

    • Around the interior content area sits a white frame:

    • Outer frame background: pure white.

    • Just inside the outer edge: a 1–2 tile thick band of small black tiles forming a continuous “data strip” around all four sides.

    Visually: a white picture frame with a thin, high-density black barcode ring hugging the content.

    4.2 Dimension Mapping Vector dimensionality: D = 768 (v1.0 default). We map dimensions to tiles along edges:

    Top edge: dims 0–255

    Right edge: dims 256–383

    Bottom edge: dims 384–639

    Left edge: dims 640–767

    Each dimension is represented by one or more tiles, depending on chosen encoding.

    4.3 Payload + FEC Each dimension v_i is:
  • Normalized (e.g., to [-1, 1] or [0,1]).
  • Quantized to N bits (v1.0 suggestion: 16–24 bits per dim; 32 bits optional).
  • Packed into codewords with FEC, then mapped to tiles.
  • Abstract constraints (implementation-flexible):

    • The barcode encoding MUST:

    • Use a standard FEC scheme (e.g., Reed–Solomon, LDPC, or similar).

    • Provide enough parity bits to correct:

    • Random bit flips from noise/compression.

    • Erasures from occluded or unreadable tiles.

    • The FEC parameters (block size, parity rate) MUST be recorded in the metadata (Section 8).

    Visually, tiles appear as black/gray micro-squares; humans don’t read them, but a decoder can treat the full ring as a linear codeword.

    5. Human Sheet Schema (Logical) MICL Sheets are described in a JSON-like schema that is agnostic to rendering technology. 5.1 Top-Level Logical Schema

    {

      "concept_id": "string",

      "schema_version": "micl_v1.0",

      "vector_ref": {

        "vector_id": "string",

        "dim": 768,

        "barcode_encoding": "v1_fec_rs"

      },

      "layout": {

        "priority": ["summary", "scenario_icons", "uml_activity", "uml_sequence"],

        "orientation_hint": "auto"  // "portrait" | "landscape" | "auto"

      },

      "sections": [

        { "type": "...", "...": "..." }

      ]

    }

    5.2 Core Section Types (v1.0) All sections share:

    • type: required.

    • Additional fields determined by subtype.

    5.2.1 summary

    Plain text overview for humans. Can come directly from vec2text then post-edited by an LLM for concision/clarity.

    {

      "type": "summary",

      "title": "Taxation",

      "body": "Taxation is a system in which a governing body collects a portion of private income or value to fund shared public goods and services."

    }

    5.2.2 scenario_icons

    Cartoon/icon explanation for non-technical readers.

    {

      "type": "scenario_icons",

      "panels": [

        { "icons": ["person", "money"], "label": "Gross income" },

        { "icons": ["arrow", "percent"], "label": "Tax taken" },

        { "icons": ["road", "hospital", "school"], "label": "Public services" }

      ]

    }

    The icon set is a fixed vocabulary (extensible), but renderers are free to skin it. 5.2.3 uml_diagram

    General UML-like diagram.

    {

      "type": "uml_diagram",

      "subtype": "sequence",  // "activity", "state", "class", "component", etc.

      "actors": ["Taxpayer", "Tax Authority"],

      "messages": [

        ["Taxpayer", "Tax Authority", "report income"],

        ["Tax Authority", "Taxpayer", "issue tax notice"],

        ["Taxpayer", "Tax Authority", "pay tax"]

      ]

    }

    Other subtypes (activity/state/etc.) define their own minimal node/edge schemas. 5.2.4 table

    Structured tabular data.

    {

      "type": "table",

      "subtype": "features",  // "interfaces", "params", "hazards", etc.

      "columns": ["Property", "Value"],

      "rows": [

        ["Mandatory", "Yes"],

        ["Regulated", "Yes"],

        ["Progressive", "Often"]

      ]

    }

    5.2.5 free_text

    Arbitrary additional notes.

    {

      "type": "free_text",

      "body": "Notes about edge cases, political variation, or limitations..."

    }

    5.3 Extensibility

    • Any new section type can be introduced with a new type string (e.g., timeline, map_overlay, etc.).

    • Renderers MUST:

    • Render types they know.

    • Skip types they do not know, without breaking other sections.

    • This makes MICL forward-compatible like XML/JSON: unknown = safely ignored.

    6. Layout & Responsiveness 6.1 Concept

    • The sheet’s content area is all pixels inside the barcode frame.

    • A layout engine:

    • Reads layout.priority to know which section types are most important.

    • Chooses a grid or stack depending on aspect ratio and complexity.

    6.2 Portrait (Phone) Example

    • Sections are stacked vertically as cards:

  • Summary (full width)
  • Scenario icons
  • Activity diagram
  • Sequence diagram
  • Feature table
  • Interfaces table
  • • Long diagrams/tables can be scrollable or zoomable.

    • User can pinch-zoom into any panel; barcode remains around the edge.

    6.3 Landscape (Desktop / Tablet) Example

    • Use a 2×3 or 3×3 grid similar to the UML sheet you posted:

    Row 1:

    • Summary | Activity Diagram | Sequence Diagram

    Row 2:

    • Component/Class Diagram | Feature Table | Interfaces Table

    Row 3 (optional):

    • Scenario icons | Additional tables/notes | (empty or future section)

    The renderer is free to pick a fit, but the semantics of each section are fixed by their type.

    7. MICL vs Old Concept Glyph

    • Old Concept Glyph:

    • Circular, 64×64 macro/residual image.

    • Useful as an engineering visualization; not ideal for broad human interpretability.

    • New MICL Sheet:

    • Canonical human-facing representation using text + diagrams + icons.

    • Barcode frame encodes concept vector with FEC.

    • Internally, a MICL renderer could still generate a Concept Glyph as one more uml_diagram-like panel for engineers if we ever want it.

    We treat MICL as a new line, not a minor revision.

    8. Metadata & Versioning

    Each sheet carries metadata (internally or in a sidecar file):

    {

      "concept_id": "lnsp_core_platform",

      "micl_version": "1.0",

      "vector_ref": {

        "vector_id": "vec_lnsp_core_001",

        "dim": 768,

        "encoding": "barcode_v1_fec_rs_0.2",  // RS with 20% parity, for example

        "checksum": "sha256:..."

      },

      "generated_by": {

        "vec_source": "LNSP-embed-v4",

        "vec2text_model": "vec2text-jxe-2025-11",

        "llm_postedit": "GPT-5.1-DirEng"

      }

    }

    9. Naming / Branding Notes Working name for this spec:

    MICL Sheet – _Mechanistic Interpretability Concept Layout_

    • Short, pronounceable (“Michael”).

    • Emphasizes layout and interpretability.

    Alternative label for the actual artifact if you want it more product-y:

    MIC Card – _Mechanistic Interpretability Concept Card_

    MICC – _Mechanistic Interpretability Concept Card_ (internal acronym).

    You can have:

    • Technology name in docs: MICL v1.0

    • UX/marketing name: MIC Cards (“Every concept gets a MIC Card”).

    Sample Card Got it—let’s do a clean, production-style MIC Card for “Taxation” in portrait, with everything spelled out so a designer or image model could render it 1:1.

    MIC Card – “Taxation” (Portrait Layout) Note: Outer border is the white 768-D barcode frame with FEC. I’ll just call it out, not expand the encoding. 0. Frame (always present)

    Outer border: pure white.

    Inner rim: 1–2 tile tall band of tiny black squares all the way around → this is the 768-D concept barcode with FEC.

    • Inside that rim is the content area described below.

    1. Top Block – Title + Summary (full width) Title: TAXATION Summary text (2–3 sentences):

    Taxation is a system in which a governing body collects a portion of private income or value to fund shared public goods and services.

    It converts individual resources into collective capabilities such as infrastructure, security, and social programs.

    Visually: rounded rectangle, light background, bold title at top, body text below.

    2. Scenario Icons (full width, under summary) A simple strip with 3 cartoon panels:
  • Panel A – “Income”
  • • Icon: person with briefcase or laptop, coins or paycheck symbol.

    • Label under icon: Gross income

  • Panel B – “Tax slice”
  • • Icon: arrow from coins → government building + percent symbol.

    • Label: Taxes collected

  • Panel C – “Public services”
  • • Icons: road, hospital cross, school.

    • Label: Public services

    These three panels sit side-by-side in a single box titled SCENARIO (INTUITIVE VIEW).

    3. Middle Row – Activity & Sequence Diagrams (2 columns) 3.1 Left: Activity / Flow Diagram

    Box title: ACTIVITY / FLOW

    Simple top-down flow (UML-ish):

     [Earn income]

            |

            v

     [Calculate tax]

            |

            v

     [Pay tax]

            |

            v

     [Allocate budget to services]

    Maybe drawn as rounded UML action boxes with arrows. 3.2 Right: Sequence Diagram Box title: SEQUENCE (INTERACTION OVER TIME) Two lifelines:

    • Taxpayer | Tax Authority

    Messages:

    Taxpayer      → Tax Authority : report income

    Tax Authority → Taxpayer      : issue tax notice

    Taxpayer      → Tax Authority : pay tax

    Tax Authority → Public Budget : update revenue

    Drawn as classic UML sequence: vertical dashed lines, horizontal arrows with labels.

    4. Bottom Row – Tables (2 columns) 4.1 Left: Feature Table

    Box title: FEATURES

    Property Value Notes

    Mandatory Yes Required by law

    Regulated Yes Defined in legal/tax code

    Progressive Often Higher incomes taxed at higher %

    Systemicity High Touches nearly every transaction

    Purpose Public goods Infrastructure, services, welfare

    4.2 Right: Interfaces Table

    Box title: INTERFACES & ACTORS

    Actor Interface Payload

    Citizen Tax portal / forms Income details

    Employer Payroll system Withheld tax data

    Tax Authority Gov. budget sys Revenue totals

    Same Card as MICL JSON Spec (for your pipeline)

    Here’s the same thing in our MICL schema form:

    {

      "concept_id": "taxation",

      "schema_version": "micl_v1.0",

      "vector_ref": {

        "vector_id": "vec_taxation_001",

        "dim": 768,

        "barcode_encoding": "barcode_v1_fec_rs_0.2"

      },

      "layout": {

        "priority": [

          "summary",

          "scenario_icons",

          "uml_activity",

          "uml_sequence",

          "table_features",

          "table_interfaces"

        ],

        "orientation_hint": "portrait"

      },

      "sections": [

        {

          "type": "summary",

          "title": "Taxation",

          "body": "Taxation is a system in which a governing body collects a portion of private income or value to fund shared public goods and services. It converts individual resources into collective capabilities such as infrastructure, security, and social programs."

        },

        {

          "type": "scenario_icons",

          "panels": [

            { "icons": ["person", "money"], "label": "Gross income" },

            { "icons": ["arrow", "percent", "gov_building"], "label": "Taxes collected" },

            { "icons": ["road", "hospital", "school"], "label": "Public services" }

          ]

        },

        {

          "type": "uml_diagram",

          "subtype": "activity",

          "nodes": [

            "Earn income",

            "Calculate tax",

            "Pay tax",

            "Allocate budget to services"

          ],

          "edges": [

            ["Earn income", "Calculate tax"],

            ["Calculate tax", "Pay tax"],

            ["Pay tax", "Allocate budget to services"]

          ]

        },

        {

          "type": "uml_diagram",

          "subtype": "sequence",

          "actors": ["Taxpayer", "Tax Authority", "Public Budget"],

          "messages": [

            ["Taxpayer", "Tax Authority", "report income"],

            ["Tax Authority", "Taxpayer", "issue tax notice"],

            ["Taxpayer", "Tax Authority", "pay tax"],

            ["Tax Authority", "Public Budget", "update revenue"]

          ]

        },

        {

          "type": "table",

          "subtype": "features",

          "columns": ["Property", "Value", "Notes"],

          "rows": [

            ["Mandatory", "Yes", "Required by law"],

            ["Regulated", "Yes", "Defined in legal/tax code"],

            ["Progressive", "Often", "Higher incomes taxed at higher %"],

            ["Systemicity", "High", "Touches nearly every transaction"],

            ["Purpose", "Public goods", "Infrastructure, services, welfare"]

          ]

        },

        {

          "type": "table",

          "subtype": "interfaces",

          "columns": ["Actor", "Interface", "Payload"],

          "rows": [

            ["Citizen", "Tax portal / forms", "Income details"],

            ["Employer", "Payroll system", "Withheld tax data"],

            ["Tax Authority", "Gov. budget system", "Revenue totals"]

          ]

        }

      ]

    }

    This is a fully fleshed example MIC Card: you can hand the textual description to a designer or plug the JSON into a renderer. Next logical step, if you want, is to do the same for LNSP core so we see how this scales to something gnarly.

    Related Research