Forge Vision Densify

forge-vision-densify

Purpose

Bridge the Dimensionality Deficit between sparse human prompts and high-dimensional pixel output. Text-to-image generation is a thermodynamic leap: a 1024×1024 image is 1,048,576 hard pixel constraints that a 20-word prompt can typically license less than 1% of. The agent is forced to fabricate the remaining 99% from training priors, inflating ΔS.

This skill does NOT generate images. It is a governance layer that wraps every T2I dispatch in the federation. It enforces:

1. Prompt Densification — sparse prompts are deterministically expanded into dense constraint specifications before any diffusion call.
2. JSON Receipt Contract — every diffusion call returns a structured receipt with falsifiable density metrics. No naked {url, status}.
3. ΔS Thresholds — density bands gate the payload:
   • < 0.20: hard rejection, forced re-densification loop
   • 0.20–0.50: delivery with honest disclosure of hallucinated elements
   • ≥ 0.50: clean delivery
4. Anchor Routing Gates — geometric/structural prompts force a ControlNet/Depth/Ip-Adapter anchor request before diffusion (hard gate); creative prompts run soft-gate with post-hoc VLM witness.

The Dimensionality Deficit (doctrine)

Text operates in discrete semantic space where ambiguity is a feature. Pixels operate in continuous tensor space where every value is a verdict.

Compulsory Hallucinator: Unlike an LLM that can leave variables undefined ("a man sat quietly in a room"), a diffusion model MUST assign a specific RGB value to every pixel. Sparse prompt = forced fabrication.

Role of Image References: Anchors (ControlNet, Depth, Canny, Ip-Adapter, InstantID) bypass the text bottleneck by providing structural constraints in the modality of the output. Text prompt + anchor reduces the mathematical gap the model must fill.

The JSON Receipt Contract (immutable)

Every T2I dispatch returns this structure. The contract is load-bearing. Consumers (Hermes, the user) MUST receive all six fields.

{
  "f1_safe": true,
  "f2_adherence": 0.85,
  "prompt_density": 0.35,
  "hallucinated_elements": [
    "wallpaper",
    "ambient lighting",
    "chair geometry"
  ],
  "anchor_required": false,
  "anchor_suggestion": null
}

Density Thresholds (ΔS Boundaries)

Pinned numbers. Falsifiable state. No subjective judgement.

Anchor Routing Gates

Hard Gate (GEOX / Structural / Earth-science)

Trigger: Prompt contains explicit geometric, spatial, or numerical constraints. Detection rules:

• Numeric angle references: 45 degrees, 30°, isometric, top-down
• Named structures: K-DIP, fault line, wellbore trajectory, horizon
• Text-in-image demands: label the X, caption:, quoted text
• Coordinate/scale constraints: cm scale, NW–SE strike, 2.3 km depth
• 3D-rendering style with measurement: orthographic, technical drawing

Action: System intercepts BEFORE diffusion. anchor_required flips to true in the receipt, and Hermes is structurally forced to request a ControlNet/Depth/Ip-Adapter/Ip-Adapter-Face anchor. No silhouette of payload may be released.

Hard Gate (Human Physique Topography Gate / AAA-human-physique-topology)

Trigger: Prompt references human physique, body composition, athletic build, muscularity, fitness, athletic pose, bodybuilding, calisthenics, or anatomy. Detection rules:

• Musculoskeletal terms: muscular, shredded, abs, biceps, pectoral, deltoid, quadriceps, vascular, striated
• Anthropometric terms: physique, somatotype, body fat, V-taper, 8-head canon
• Fitness/Athletic settings: gym portrait, bodybuilder, fitness model, calisthenics

Action: System intercepts BEFORE diffusion. The prompt must be densified through the 5-stratum topography model (AAA-human-physique-topology), or bound with OpenPose/DensePose/Depth structural conditioning. If naive/sparse ("a muscular guy"), density_lower is strictly $< 0.20$, forcing re-densification via compile_physique_prompt.py to prevent plastic mannequin hallucination and anatomical drift.

Soft Gate (Creative / Marketing / Illustrations / Hero shots)

Trigger: Default. No geometric/numerical or human physique hard-gate criteria detected.

Action: Standard pipeline runs. VLM Tri-Witness evaluates output. If mismatch is severe (f2_adherence < 0.70 OR residual_entropy == HIGH), the receipt flags issue post-generation. Disclosure follows the density band rules above.

Shadow Mode Deployment (F1 Reversibility)

Initial deployment runs VLM Tri-Witness in shadow mode:

• Log f1_safe, f2_adherence, prompt_density, hallucinated_elements, anchor_required, anchor_suggestion for every render
• DO NOT gate the payload based on these values
• Calibrate false-positive rate for 7 days before promoting to hard gate

This prevents a malformed VLM prompt from breaking active image generation while we map the false-positive baseline.

Density Estimation Method (Hybrid)

The density estimate is NOT a single number from a single source. It is a hybrid:

Stage 1 — Heuristic Lower Bound (Hermes-side, no API call)

Pure math. Computed pre-dispatch. Concept-coverage formula:

density_lower = (
  has_subject_identity   × 1.00 +
  has_subject_pose       × 0.80 +
  has_primary_material   × 0.60 +
  has_background_scene   × 0.40 +
  has_lighting_lens      × 0.30 +
  has_atmospheric_tone   × 0.15
) / 3.25

Max possible score = 1.00 (all six concepts fully specified). Default raw prompt like "generate a portrait" → density_lower ≈ 0.10. Hard reject territory.

If density_lower ≥ 0.50:

• Skip VLM round-trip
• Final density = density_lower clamped to [0.50, 1.0]
• Save tokens, save latency

Stage 2 — VLM Alignment Pass (only if Stage 1 < 0.50)

VLM runs a three-question audit (see references/vlm-audit-prompt.md). The VLM does NOT estimate a percentage. It answers three discrete booleans:

1. Foreign object presence — are there visible elements NOT in the prompt?
2. Entity specification completeness — for each prompt entity, is its spatial/lighting/material/scale specified?
3. Residual entropy — could a second diffusion model produce a substantially different image?

Failure count → density band:

• 0 fails → density = 0.50
• 1 fail → density = 0.35
• 2 fails → density = 0.20
• 3 fails → density = 0.20 (also flagged for second re-densify)

Atmospheric Exemption List

To prevent VLM false-positives on lighting/mood primitives, these are NOT counted as foreign objects:

atmospheric_primitives = [
  "rim light", "fill light", "skin bloom", "lens flare",
  "depth of field", "bokeh", "haze", "atmospheric perspective",
  "chromatic aberration", "ambient occlusion", "soft shadow",
  "specular highlight", "key light direction", "backlight"
]

These may appear in hallucinated_elements for transparency, but they do NOT count toward the failure tally in Stage 2.

Constitutional Binding

Trigger / When to Use

Use when:

• Any T2I (text-to-image) dispatch in the federation
• Any vision request from Hermes that will materialize pixels
• Any workflow that produces RGB output from a text prompt

Do NOT use when:

• Image editing (inpainting, transparency, style transfer) — different dimensionality regime; use the multimodal router's edit path
• Pure VQA / image understanding — no pixel output
• Video generation — T2V has different temporal semantics; separate skill
• Audio/somatic — different modality entirely

Procedure (callers MUST follow this order)

1. Receive sparse or dense human prompt.
2. Classify — hard-gate geometry? Detect numeric/structural/key terms. If yes → set anchor_required: true, request anchor, halt here.
3. Densify — run recipes/densify.py to expand sparse prompt into constraint specification. Stage 1 heuristic computes density_lower.
4. Branch:
   • If density_lower ≥ 0.50 → proceed directly to diffusion with prompt_density: density_lower in receipt.
   • Else → run recipes/shadow_audit.py (VLM tri-witness) AFTER diffusion to populate f2_adherence, hallucinated_elements, anchor_required, final prompt_density.
5. Gate:
   • density < 0.20 → throw exception, re-densify loop, do NOT deliver
   • density ∈ [0.20, 0.50) → deliver WITH honest disclosure caption
   • density ≥ 0.50 → deliver clean
6. Receipt — return the JSON contract to caller. Every field populated.
7. Log — append to density_audit.csv for shadow-mode calibration.

Failure Modes

Key Paths

Telemetry per Run

{
  "skill_name": "forge-vision-densify",
  "version": "1.0.0",
  "prompt_original": "<hash>",
  "prompt_densified": "<hash>",
  "density_lower": 0.0,
  "density_final": 0.0,
  "vlm_stage_invoked": false,
  "vlm_failures": 0,
  "anchor_required": false,
  "f1_safe": null,
  "f2_adherence": null,
  "delivery_mode": "clean|disclose|reject",
  "shadow_mode": true,
  "latency_ms": 0
}

Verification

Two test prompts required before SEAL:

1. Sparse human prompt ("a man standing in a park"): density_lower should compute to < 0.20, hard-reject path exercised.
2. Dense constraint prompt ("Caucasian male, 30s, navy suit, white shirt, red tie, golden hour sidelight, 85mm portrait lens, bokeh, urban rooftop"): density_lower ≥ 0.75, VLM stage skipped, clean delivery.

Sealing Note

DITEMPA BUKAN DIBERI — Densification Receipt Forge

Skill forged 2026-08-27 under F13 directive. The pipe cannot leak. The limit is structural. The receipt is load-bearing. The VLM is shadow-mode for one week of false-positive calibration before hard promotion.

Acknowledged correction: "knowing limits" → "structural governance." Anti-Hantu: character attribution removed. Pipe compliance is the bar.