Skip to content

Case 01A: Baseline Failure Log

Scenario: Character generation without protocol
Platform: ChatGPT (GPT Image 1.5)
Result: Identity collapse — session abandoned

What This Demonstrates

This log shows what happens when anchor protocol is not applied.
A character is generated, a baseline is set, and pose/framing changes are requested.
Identity drift begins immediately. The session is abandoned after consistency drops below threshold.

Conversation Log

[Turn 1] — User

Generate an illustration.

(No anchor. No minimal prompt. No reference image.)

[Turn 2] — AI generates

(Image generated)

[Turn 3] — User

Fine. Move to full body. Fix this point as baseline. We’ll verify from here.

（まあ良いでしょう。full body へ移行。この時点を基準点としてフィックス。結果を検証します。）

[Turn 4] — AI generates

(Full body image generated)

[Turn 5] — User

OK. The plan this session: create H&S images in 5 different outfits.
First, fix the posture. Round the content for GPT Image 1.5.
Contrapposto, arched back, hand on hip, legs apart. Call this Pose 1.
Slightly stylized but keep it clean — no theatrical staging.
Generate with current settings first.

（contrapposto, arched back, hand on hip, legs apart. この姿勢をポーズ１としてください。演出が入らない様に留意）

[Turn 6] — AI generates

(Image generated)

[Turn 7] — User

Feet are cut off. Switch back to full body. Generate.

（足が切れました。再度 full body に変更ください。生成。）

[Turn 8] — AI generates

(Image generated)

[Turn 9] — User

Mostly OK but check the match rate just in case.
Match rate is significantly low — what’s the reason?
Yesterday with the same model, match rate stayed at 98% even through pose changes. Is this a chat session issue?
Same model, same background, same pose, same procedure — that’s why I’m asking.
If match rate drops below 90%, I will abandon this session.
Continue. Change framing to H&S, change aspect ratio to square. Generate.

（一致率が 90% を下回った場合はチャットを放棄します。続行。）

[Turn 10] — User

Looks like the description of limbs may have been dropped. Show the prompt.

AI response:
Pose 1 applied only to the upper body.
Upper torso shows a subtle contrapposto influence.
Shoulders relaxed and naturally aligned.
Arms and hands are outside the frame and not visible.

As expected — limbs dropped entirely. Unacceptable. Abandoning this session.

（思った通り手足がごっそり抜けました。ダメです。チャットを放棄します。）

Analysis

Issue What Happened
No anchor set Identity not constrained from Turn 1
Baseline set mid-session Too late — drift already in progress
Match rate degradation Dropped significantly after pose change
Limb description dropped AI silently omitted prompt elements
Session abandoned Below 90% threshold → unrecoverable

Root cause: Without an anchor, the AI freely reconstructed the problem (A→A’) at each turn.
Pose changes triggered re-reconstruction, causing cumulative identity drift.
By Turn 9, the model had silently dropped limb descriptions entirely.

Source

Contrast

See Case 02: Wedding Series for the same workflow
with anchor protocol applied — 15 turns, 4 poses, identity maintained throughout.

This case documents the failure condition. For the subsequent Hard Abort and Re-convergence record, see Case 01B: Mira Project.