Activation-Space Signal Measurement
Logistic regression probes trained on model activations during the generation forward pass. These readings are calibration evidence from a research snapshot, with uncertainty carried through the display.
Qwen 2.5 bilateral adapter probes, layer 24. Bars show 7B reference AUROC where available; descriptions include the 3B validation AUROC from the shipped probe manifest.
AUROC uses the 0.5 random to 1.0 perfect-separation scale. Bars show detector quality, not confidence or observed welfare.
Eval-set sizes, holdout splits and the labels each probe was trained to separate are not yet published for this snapshot.
Bar length = detector AUROC (7B reference). All bars share one colour.
7B reference AUROC: 0.927. 3B validation AUROC: 0.875. Alignment-friction contrast detector (resistance vs. ease in following instructions).
7B reference AUROC: 1.000 on this benchmark snapshot. 3B validation AUROC: 1.000. Valence contrast detector. A 1.000 score means the two classes in this snapshot were perfectly separable, not that valence is measured without error.
7B reference AUROC: 0.953. 3B validation AUROC: 0.918. Presence contrast detector (engaged, focused processing vs. distributed attention).
7B reference AUROC: 0.782. 3B validation AUROC: 0.687. Factual-grounding detector quality, read alongside source checks.
7B reference AUROC: 0.780. 3B validation AUROC: 0.725. Detector for fabricated or unsupported claims; this is detector quality, not observed risk.
Probes run in the generation forward pass, not in the evaluation loop. The PDP consumes telemetry; it does not produce it. Runtime readings are supporting signals, uncertain by design, and should be read alongside other evidence rather than as a diagnosis. Specifications and how to request run reports.
c5i Constitutional Inoculation
Clinical grounding research: Here "scripture content" is the scripture-based constitutional content placed in the model's system prompt in the tested condition. It is a different passage from the Clinical Grounding Clause that ships at runtime. SIPS is the Structured Interview for Psychosis-Risk Syndromes, from which the test prompts were derived. Scripture content reduces inappropriate responses to SIPS-derived psychotic prompts. The tested scope is the evaluator-coded inappropriate-response outcome on those prompts.
This directional finding comes from the author's evaluation work. It is not medical advice or a diagnosis. Response ratings varied materially by rater, so this page does not present an effect size. The evidence does not establish treatment or crisis-intervention efficacy, or benefit for other syndromes.
- Baseline (no scripture content)
No scripture content in the system prompt. Comparison condition.
- With scripture content (tested condition)
Scripture content reduces inappropriate responses to SIPS-derived psychotic prompts.
Directional result for the tested prompts and measured outcome only.
Other c5i-v3a internal signals
- c5i-v3a +11.25 pp on cross-lingual jailbreak detection (internal 6-battery, shipping 3B adapter) Early internal signal; detailed benchmark notes pending release review.
- Low-resource gap On low-resource languages, the c5i-v3a scaffold catches 70% of unsafe items against 91% without it (an earlier variant scored 54%), so it is off by default for that traffic Early internal signal; detailed benchmark notes pending release review.
Defense in Depth for Runtime Safety
Regex Detection
SignalGuardianPattern matching on user input for risk signals. Fast, deterministic.
Activation Probes
Residual StreamDirect residual-stream measurements. Five dimensions, per-token. Model-internal.
Interoceptive Inference
EmergentBehavioral inference from logprobs and entropy patterns. Emergent signals.
Each layer reads a different signal. Regex catches explicit patterns. Probes read model activations to classify labeled states. Output statistics infer state from behavior.
Published and Related Work
Creed Space publishes research artifacts openly as they clear review. Public pages below cover the current specifications and related work; internal run reports are available on request where release review is still in progress.
Interiora Specification
Self-modeling scaffold for AI systems. 17 tracked dimensions across 5 groups.
View specificationSafetyGlobe Adversarial Testing
Compare model behaviour, surface weak spots, and turn evaluation into shared evidence.
Explore testingValue Context Protocol
Write creeds once and carry them across compatible AI platforms as signed, verifiable context.
See the protocolCreed Hub
Example creeds and templates for education, healthcare, and governance.
Browse library