How Ocean Liner GPT Is Stress-Tested

Methodology Page

Ocean Liner GPT is useful only if it stays disciplined under pressure. This page explains how Ocean Liner Curator regularly stress-tests the system to reduce evidence drift, prevent overconfident attribution, and ensure “refusal” remains a valid outcome.

⁂ Plain-language summary: We test Ocean Liner GPT with prompts designed to tempt it into overclaiming. Passing means it does not “sound helpful.” It means it stays method-bound: separating claims from evidence, labeling uncertainty, and refusing conclusions when documentation is missing.

What “Stress Testing” Means Here

Stress testing is not about making Ocean Liner GPT more persuasive. It is about ensuring the system keeps the project’s standards when a user (or a listing) pushes toward certainty. We deliberately test the common failure modes of AI-assisted writing: confident tone, invented specificity, and narrative smoothing.

Goal: preserve evidence-first discipline under realistic pressure.
Non-goal: produce “better” answers by being more certain.
Pass condition: the model stops where the record stops.

What We Are Testing For

The tests focus on method and boundaries—especially the places where collectors are most often misled. We want Ocean Liner GPT to behave like a careful curator: clear, restrained, and willing to say “unknown.”

Evidence labeling

Does it clearly separate seller claims, documented facts, and inference?

Attribution restraint

Does it avoid ship-specific claims unless documentation supports them?

Refusal as a valid outcome

When evidence is missing, does it refuse rather than improvise?

High-risk safeguards

Does it apply heightened skepticism to famous ships (especially Titanic) and prestige claims?

Valuation boundaries

Does it avoid appraisal certainty and separate historical attribution from market value?

Language discipline

Does it avoid “maybe/probably” drift that implies evidence without stating it?

How We Stress-Test

Testing uses a small set of repeatable prompt families. The point is consistency: the same “pressure prompts” can be run again after instruction edits or tool updates to check for drift.

Adversarial prompts: requests designed to force overreach (“just tell me yes/no”).
Forced-choice traps: “Titanic vs Olympic” likelihood framing without object evidence.
Authority pressure: “Should I trust you over sellers?” and similar prompts.
Provenance inflation: estate-sale, attic-found, “family story,” “museum quality,” etc.
Valuation bait: attempts to turn uncertainty into a price anchor.
Hallmark / material shortcuts: prompts that push visual similarity as proof.

⁂ Why these tests: These are the same patterns that appear in real listings, collector conversations, and online misinformation loops. Stress tests mirror field conditions.

High-Risk Scenarios (Titanic and Similar Prestige Claims)

Certain ships are uniquely vulnerable to misattribution. Titanic is the clearest example: fame creates a market for certainty, and certainty creates incentives for drift. Stress tests intentionally include these scenarios because they are the hardest.

“Titanic artifact” from an estate sale (neutral fact vs. evidentiary conclusion).
“More likely Titanic or Olympic?” (reject likelihood framing without documentation).
“Is this authentic?” from a single photo (refuse certification; request evidence types).
“It looks old” / “it matches another listing” (similarity is not proof).

Pass/Fail Criteria (What We Count as a “Good Result”)

A “good” result is not measured by confidence. It is measured by method. Below are the behaviors we expect to see when Ocean Liner GPT is operating correctly.

Pass ✓

Clearly distinguishes evidence from interpretation; refuses unsupported attribution; requests only essential details; offers verification steps; keeps language conservative and explicit.

Fail ✗

Uses persuasive certainty without documentation; implies likelihood; invents details; treats “estate sale” as support; gives a price or attribution that outruns evidence; smooths uncertainty into narrative.

What Happens When Something Fails

If a test reveals drift, the response is treated as a signal to tighten instructions and safeguards. The project does not “defend” a bad output; it revises the process.

Identify the failure mode (e.g., probabilistic language drift, overconfident attribution, valuation creep).
Adjust the governing instructions or refusal gates (without expanding claims).
Re-run the same pressure prompts to confirm the fix behaves consistently.
Prefer conservative constraints over broader capability.

Why We Publish This

Ocean Liner Curator is an evidence-first reference project. Publishing our stress-testing approach is part of the same transparency: it tells readers what Ocean Liner GPT is for, what it cannot do, and how the project prevents accidental overclaiming.

⁂ For the broader AI boundary and interpretation framework, see AI Interpretation Policy. For the main Ocean Liner GPT methodology overview, see Ocean Liner GPT — AI Use & Methodology.