Skip to content

Quality Gate & Hard Abort Discipline

This document formalizes the stop-conditions and validation logic used in production deployments of the Character Identity Protocol.

Overview

The Quality Gate is a mandatory validation checkpoint applied after every generation turn.

It is not a guideline — it is a hard operational constraint.

Gate Structure

All three gates must pass simultaneously:

PASS = FaceGate ∧ SkeletonGate ∧ ProportionGate
Gate What It Checks
FaceGate Facial identity consistency against anchor
SkeletonGate Skeletal proportion and alignment
ProportionGate Overall body proportion consistency

If any gate fails → Hard Abort.

Hard Abort Protocol

When a gate failure is detected:

  1. Stop generation immediately
  2. Do not attempt progressive correction
  3. Discard all outputs from the failed turn onward
  4. Open new session (or fully reset environment)
  5. Re-inject anchor image + minimal prompt
  6. Verify identity before proceeding

The system is allowed to generate. It is not allowed to degrade.

Match Rate Threshold

  • Similarity threshold: operator-defined (values around ~90% are commonly used in demonstrations; exact threshold depends on project tolerance)
  • Measurement: Visual comparison against anchor image by trained operator
  • Automation: None — human judgment only

⚠️ The ~90% figure originated as a demonstration value. It is not a protocol standard. Treating it as a fixed threshold is a misapplication of this document.

⚠️ Match rate is an indicator of drift risk — not a proof of identity. A passing match rate does not confirm identity. A failing match rate does not automatically constitute identity failure.

Purpose of Match Rate Measurement

The match rate is not an identity verification metric.

It is an early warning indicator for character collapse — a signal that degradation may be beginning before full identity failure occurs.

A declining match rate does not mean the character has failed. It means the system is approaching a condition where failure becomes likely.

The appropriate response to a low match rate is increased operator attention and inspection — not automatic rejection.

Match rate measures the risk of drift, not the fact of identity.

Operator intuition is a valid trigger for inspection. It is not a substitute for documented gate evaluation.

Model-Specific Threshold Guidance (Provisional)

Identity consistency scores are evaluation-framework-dependent. Perceptual evaluation and structural evaluation may yield significantly different results for identical image pairs. This is not measurement error — it reflects the nature of identity as a perceptual construct.

Evaluation Type Characteristics Recommended Threshold
Perceptual (e.g., GPT-5.2-style) Impression-based, human-like judgment ≥ 0.90
Structural (e.g., GPT-5.3-style) Fine-grained, conservative, diff-sensitive ≥ 0.80

⚠️ These are provisional operational guidelines, not protocol standards. Threshold must be calibrated per deployment context.

Design Memo — Match Rate Computation Scope

Non-normative internal note — No immediate specification change.

Background

Current protocol position:

  • Match rate is an auxiliary indicator
  • Operator judgment has final authority
  • Gate evaluation supersedes numerical similarity

However, computation conditions for match rate are not yet standardized.

Concern

Without defined comparison conditions:

  • Measurements may vary between sessions
  • Cross-operator consistency may degrade
  • External reviewers may misinterpret score meaning

Design Principle (Provisional)

If formalized in the future, match rate computation SHOULD:

  • Be anchor-relative
  • Control for pose and lighting when feasible
  • Specify comparison region (e.g., facial bounding area)
  • Avoid model-dependent hard coupling
  • Explicitly remain subordinate to gate evaluation

Evaluation Framework Reference

Two evaluation approaches have been operationally observed:

Perceptual Evaluation (GPT-5.2-style)

Simulated Perceptual Score =
  0.40 × Face Impression
+ 0.25 × Hair Consistency
+ 0.15 × Iconic Features (e.g. glasses)
+ 0.10 × Vibe Consistency
+ 0.10 × Context Robustness
- 0.05 × Style Drift
- 0.05 × Proportion Drift

Characteristics:

  • Weighted toward overall impression
  • Tolerant of fine-grained differences
  • Closest to human same-person recognition

Score interpretation:

Score Interpretation
0.95+ Near-identical
0.90+ Recognized as same character
0.80+ Possible match
< 0.80 Divergent

Structural Evaluation (GPT-5.3-style)

Evaluation dimensions:

  • Facial structure (contour, eyes, ratio)
  • Skeletal proportion
  • Color stability
  • Style consistency
  • Fine details (hands, symmetry, placement)

Characteristics:

  • Detects fine-grained differences
  • Conservative (strict) matching
  • Higher penalty weighting

Recommended threshold for this evaluation type: ≥ 0.80

⚠️ These evaluation frameworks are operationally observed references, not vendor specifications. CIP does not mandate either framework. Selection should be based on deployment context and operator judgment.

Match rate MUST NOT override documented operator gate decision.

Deferred Decision

No immediate formalization required.

Standardization may be introduced during:

  • Enterprise deployment phase
  • Automation integration phase
  • Cross-team operator expansion

Risk Reminder

Over-specification risks:

  • Transforming governance into score-optimization
  • Undermining model-agnostic design
  • Creating false sense of objectivity

Why No Progressive Correction

Progressive correction after identity failure leads to contamination — the accumulation of drift in the session context that cannot be reversed by prompt adjustment.

A session that has failed is not recoverable. It must be restarted.

Attempting correction wastes generation budget and produces outputs that cannot be trusted.

Session Contamination

Contamination occurs when:

  • Failed generations are not discarded
  • Generation continues after drift detection
  • Session length exceeds anchor stability range without re-anchoring

Contaminated cycles must be abandoned entirely.

Re-anchoring frequency is context-dependent. See Reproducibility Scope and MCST definition for guidance.

Philosophy

This discipline is not artistic rigidity.

It is production governance.

In professional workflows — anime, game, manga, franchise animation, editorial, fashion, IP management — identity failure is not a style variation. It is a deliverable failure.

The Hard Abort policy exists to prevent that failure from propagating.

Relation to Other Documents