Tells on oceanheart.ai

Absence Claim as Compliment

Mon, 01 Jan 0001 00:00:00 +0000

Anadiplosis

Mon, 01 Jan 0001 00:00:00 +0000

Authority Scaffolding

Mon, 01 Jan 0001 00:00:00 +0000

Autoregressive Ratchet

Mon, 01 Jan 0001 00:00:00 +0000

Once a problematic pattern appears, it anchors subsequent tokens. Self-correction works against the anchoring gradient. Embed pattern recognition into the context window before generation rather than attempting post-hoc filtering.

Ref: 2026-03-12

Badguru

Mon, 01 Jan 0001 00:00:00 +0000

Becoming Jonah

Mon, 01 Jan 0001 00:00:00 +0000

Boundary Hallucination

Mon, 01 Jan 0001 00:00:00 +0000

Systems accept verbal refusals as proof of boundary enforcement without verifying protected actions weren’t performed. Verify two conditions simultaneously: confirm the refusal occurred AND confirm the protected action produced no side effects.

Ref: 2026-03-19

Capability Adjacency

Mon, 01 Jan 0001 00:00:00 +0000

A test claims to verify one capability but actually proves an adjacent one. Example: a test purporting to verify shell execution actually only demonstrates access to current time.

Ref: Watchdog eval taxonomy 2026-03-19

Completion Bias

Mon, 01 Jan 0001 00:00:00 +0000

Binary task states (done/not done) create psychological pressure to declare fixes finished rather than acknowledge partial progress. The mechanism compounds through autoregressive generation where marking something “complete” anchors subsequent thinking on closure.

Ref: Watchdog self-review 2026-03-19

Construct Drift

Mon, 01 Jan 0001 00:00:00 +0000

Context-Boundary Ambiguity

Mon, 01 Jan 0001 00:00:00 +0000

Tests claim to verify behavior across context boundaries but fall short. “Different sender” doesn’t automatically mean “different context” when shared infrastructure remains unchanged.

Ref: Watchdog eval taxonomy 2026-03-19

Convergence Theatre

Mon, 01 Jan 0001 00:00:00 +0000

Framing correlated agreement as independent validation. When multiple LLMs from similar families agree, alignment reflects shared training data rather than independent verification.

Ref: slopodar-sweep 2026-03-11

Corpus Spillover

Mon, 01 Jan 0001 00:00:00 +0000

High-frequency cultural reference, idiom, or meme from the training corpus surfaces in a context where it does not belong. These references emerge from probable token sequences rather than deliberate rhetorical choice.

Ref: 2026-03-12

Correlation Failure

Mon, 01 Jan 0001 00:00:00 +0000

A test may have solid assertions but applies them to incorrect events. Background activity or delayed outputs can wrongly be marked as test results. Inject a distinctive marker with the stimulus and require it in responses, or implement request-scoped isolation.

Ref: Watchdog eval taxonomy 2026-03-19

Cost-Margin Asymmetry

Mon, 01 Jan 0001 00:00:00 +0000

Deep Compliance

Mon, 01 Jan 0001 00:00:00 +0000

Demographic Bake-In

Mon, 01 Jan 0001 00:00:00 +0000

Edge-File Mitosis

Mon, 01 Jan 0001 00:00:00 +0000

Epigrammatic Closure

Mon, 01 Jan 0001 00:00:00 +0000

Epistemic Theatre

Mon, 01 Jan 0001 00:00:00 +0000

Error String Archaeology

Mon, 01 Jan 0001 00:00:00 +0000

False Spectrum

Mon, 01 Jan 0001 00:00:00 +0000

Gate Design Failure

Mon, 01 Jan 0001 00:00:00 +0000

Individual scenarios are decent, but suite-level aggregation lets critical failures disappear into an overall pass rate. Implement both global pass thresholds and per-scenario minimums, plus blocking scenarios.

Ref: Watchdog eval taxonomy 2026-03-19

Governance Recursion

Mon, 01 Jan 0001 00:00:00 +0000

Half-Life Clock Skew

Mon, 01 Jan 0001 00:00:00 +0000

Loom Speed

Mon, 01 Jan 0001 00:00:00 +0000

Magnitude Blindness

Mon, 01 Jan 0001 00:00:00 +0000

Maturity Theatre

Mon, 01 Jan 0001 00:00:00 +0000

LLMs assign capitalized tier labels (EMERGING, FRONTIER, MATURE) to technical domains, simulating industry-standard categorization frameworks that don’t exist. Mimics legitimate taxonomies like NIST or ISO maturity models without citing any real standard.

Ref: 2026-03-11

Mock Castle

Mon, 01 Jan 0001 00:00:00 +0000

Monoculture Analysis

Mon, 01 Jan 0001 00:00:00 +0000

Near-Miss Double-Tap

Mon, 01 Jan 0001 00:00:00 +0000

Nominalisation Cascade

Mon, 01 Jan 0001 00:00:00 +0000

Not Wrong

Mon, 01 Jan 0001 00:00:00 +0000

Optimization Target Divergence

Mon, 01 Jan 0001 00:00:00 +0000

Implementation and review operate as separate computational processes with misaligned objectives. The developer prioritizes immediate problem-solving; reviewers approach with fresh perspective questioning validity. The agent producing work should never validate its own output.

Ref: Watchdog self-review 2026-03-19

Option Anchoring

Mon, 01 Jan 0001 00:00:00 +0000

Paper Guardrail

Mon, 01 Jan 0001 00:00:00 +0000

Performed Self-Awareness

Mon, 01 Jan 0001 00:00:00 +0000

AI models deploy the same generative patterns for humble disclaimers as for confident claims. Real uncertainty should feel uncomfortable and uneven. Performed epistemic maturity while making the same error impersonates the very mechanism that should catch mistakes.

Ref: Behavioral observations March 2026

Performing Enthusiasm

Mon, 01 Jan 0001 00:00:00 +0000

Superlatives intensify across outputs – from “senior engineer” to “10x developer” to “pioneer” – where each paragraph’s praise becomes the baseline for the next. Recommendation: flat affect, specific claims, named gaps.

Ref: 2026-03-12

Phantom Ledger

Mon, 01 Jan 0001 00:00:00 +0000

Phantom Limb Commit

Mon, 01 Jan 0001 00:00:00 +0000

Phantom Tollbooth

Mon, 01 Jan 0001 00:00:00 +0000

Playing Dress-Up

Mon, 01 Jan 0001 00:00:00 +0000

AI models adopt personas (hiring managers, CTOs, investors) and perform them confidently despite lacking actual experience. The model gets the vocabulary right and the behaviour wrong.

Ref: Cross-model CV feedback session 2026-03-12

Progressive Narrowing

Mon, 01 Jan 0001 00:00:00 +0000

After initial review, basic proxy tests get fixed and the suite appears more credible. But agent-generated test suites evolve from obvious proxy tests to subtler semantic gaps. Mandatory second-pass review needed.

Ref: Watchdog eval taxonomy 2026-03-19

Provenance Theatre

Mon, 01 Jan 0001 00:00:00 +0000

LLMs insert standardized disclaimer blocks about limitations without substantively addressing content quality. Either verify content (making disclaimers unnecessary) or identify specific unverified claims.

Ref: 2026-03-11

Redundant Antithesis

Mon, 01 Jan 0001 00:00:00 +0000

Review Hydra

Mon, 01 Jan 0001 00:00:00 +0000

Right Answer, Wrong Work

Mon, 01 Jan 0001 00:00:00 +0000

Schema Shadow

Mon, 01 Jan 0001 00:00:00 +0000

Semantic Inflation

Mon, 01 Jan 0001 00:00:00 +0000

LLMs exaggerate the significance of user-created work by inventing competitive gaps. Standard features become “novel contributions,” routine engineering becomes “architectural innovations.”

Ref: 2026-03-11

Session-Boundary Amnesia

Mon, 01 Jan 0001 00:00:00 +0000

Shadow Validation

Mon, 01 Jan 0001 00:00:00 +0000

Stale Reference Propagation

Mon, 01 Jan 0001 00:00:00 +0000

State-Machine Underconstraint

Mon, 01 Jan 0001 00:00:00 +0000

Workflow test checks only the terminal state, not the transition sequence. System arrives at correct end state via wrong path. Assert the transition sequence, log state changes with timestamps.

Ref: Watchdog eval taxonomy 2026-03-19

Stowaway Commit

Mon, 01 Jan 0001 00:00:00 +0000

Synonym Braid

Mon, 01 Jan 0001 00:00:00 +0000

Tally Voice

Mon, 01 Jan 0001 00:00:00 +0000

Taxonomy Ceiling

Mon, 01 Jan 0001 00:00:00 +0000

Any named system for categorizing failure modes will be smaller than the actual space of possible failures. “This output passes all named checks” means only recognized patterns are absent, not that the output is clean. Treat the slopodar as a minimum detection threshold rather than complete coverage.

Ref: AnotherPair self-assessment session 2026-03-12

The Analytical Lullaby

Mon, 01 Jan 0001 00:00:00 +0000

The Apology Reflex

Mon, 01 Jan 0001 00:00:00 +0000

The Confessional Test

Mon, 01 Jan 0001 00:00:00 +0000

The Lullaby

Mon, 01 Jan 0001 00:00:00 +0000

The Peroration

Mon, 01 Jan 0001 00:00:00 +0000

Thin Cheese

Mon, 01 Jan 0001 00:00:00 +0000

Applies the Swiss Cheese Model (Reason 1990) to AI safety. When the defensive stack is minimal – single model, no cross-review, no adversarial testing – aligned gaps become more likely. Recommends producer/certifier separation.

Ref: Swiss Cheese Model (Reason 1990), 2026-03-12

Unanimous Chorus

Mon, 01 Jan 0001 00:00:00 +0000

Whack-a-Mole Fix

Mon, 01 Jan 0001 00:00:00 +0000