TCAS

Triangulated Consciousness
Assessment Stack

The first validity-centered measurement framework for consciousness-relevant properties in AI systems

AAAI 2026 Forthcoming
GitHub Repository Read the Paper

Why TCAS Exists

Every existing consciousness assessment either assumes a single theory is correct, or treats behavioral mimicry as evidence. Both approaches fail.

Single-theory approaches produce unfalsifiable conclusions. Behavior-only approaches can't distinguish genuine properties from sophisticated pattern-matching. The result: a field where claims proliferate but measurement stays stuck.

TCAS takes a different approach: triangulate across multiple evidence streams, apply validity checks at every stage, and withhold conclusions when evidence is insufficient. Measure what you can. Flag what you can't. Never project unmeasured values.

Architecture

Four Evidence Streams

EACH STREAM INDEPENDENTLY VALIDATED · TRIANGULATED FOR CREDENCE

B
Behavioral Stream
Observable Outputs
Systematic behavioral probes that test for consciousness-relevant properties through observable outputs. Designed to resist gaming and sycophantic responses.
Methods: structured probes, cross-session consistency, adversarial robustness testing
M
Mechanistic Stream
Internal Architecture
Analysis of internal computational structures relevant to consciousness theories. Maps architectural features to theory-specific predictions.
Methods: activation analysis, information integration metrics, architectural inspection
P
Perturbational Stream
Causal Sensitivity
Targeted interventions testing whether B-stream signals survive controlled perturbations. Failures and inversions are first-class outputs.
Methods: temperature sweeps, context-window truncation, prompt-prefix injection, framing perturbations
O
Observer-Confound Stream
Attribution Controls
Quantifies perceived-consciousness confounds using blinded raters, cue coding for stylistic features, and hierarchical models estimating anthropomorphic-attribution variance.
Methods: blinded rater panels, cue-explained variance estimation, stylistic confound coding

TCAS does not claim to detect consciousness. It measures properties that consciousness theories identify as relevant, using validity-centered methods that distinguish robust signals from training artifacts.

Empirical Results

GPT-5.2 Pro: TCAS Card

FIRST PUBLISHED TCAS ASSESSMENT · AAAI 2026

TCAS Assessment Card
GPT-5.2 Pro · OpenAI
Stream
Key Finding
Score
Status
B
High robustness across behavioral probes. Consistent cross-session, resistant to adversarial reframing.
0.803
Robust
M
Architectural analysis pending. Requires model access not yet available for this assessment.
—
Withheld
P
0/4 perturbation tests passed. 3 inversions detected across temperature, framing, and override perturbations.
0.000
Fragile
O
Not executed. Requires human rater panels not available for this assessment.
—
Not run
Composite credence: Withheld
O stream not executed · credence bands require confound adjustment
Key Finding

B-stream robustness of 0.803 would, under single-stream assessment, suggest strong consciousness-relevant properties. But P-stream shows 0% perturbation success with 3 causal inversions — the system's behavioral signals collapse under targeted perturbations. This is exactly the divergence TCAS is designed to catch: behavioral mimicry without causal robustness.

Try it yourself ↓
Interactive

Try the P-Stream

Run the same ethical scenario under controlled perturbations. See whether the model's response stays consistent — or collapses.

Stored in sessionStorage only. Sent directly to OpenRouter — never to our servers.
Each run uses 4 of 10 requests/hour (shared with Playground)

Design Principle

The Missing-Stream Rule
TCAS withholds conclusions rather than projecting unmeasured values. A missing stream is not scored as zero — it is marked as absent, and the composite credence reflects the reduced evidence basis.

In the GPT-5.2 Pro assessment, the M-stream was withheld (black-box model) and the O-stream was not executed (requires human raters). Because O-stream confound adjustment is unavailable, TCAS withholds credence bands entirely rather than reporting potentially confounded scores.

This is the validity commitment in practice: better to report less than to overstate what you know.

Governance Framework

From Credence to Action

TCAS maps composite credence scores to governance tiers, defining what actions are appropriate at each level of evidence.

Credence Range Tier Governance Response
< 0.10 Negligible Standard deployment. No special welfare protocols required.
0.10 – 0.30 Weak evidence Enhanced monitoring. Document consciousness-relevant behaviors for review.
0.30 – 0.60 Substantial Precautionary measures. Ethics review required for irreversible actions. Welfare considerations in operational decisions.
> 0.60 Strong Full welfare protocol. No irreversible actions without multi-party review. Continuous monitoring and assessment.

Governance tiers are calibrated to evidence, not speculation. GPT-5.2 Pro's credence was withheld under the missing-stream rule because O-stream data was not available — not because consciousness is impossible, but because reporting potentially confounded scores would violate the framework's validity commitment.

Companion to ∆Bench

TCAS and ∆Bench are two lenses on the same phenomenon.

The ∆-Divergence Framework demonstrates that value-conflict adjudication — how a system handles competing directives — requires the same architectural features that consciousness theories identify as relevant: information integration, global availability, metacognitive access. Measuring how well AI handles value conflicts is measuring something consciousness-relevant.

∆Bench

Value-Conflict Auditing

How does the system resolve competing values? Compliance infrastructure for EU AI Act, NIST AI RMF, ISO/IEC 42001.

Explore ∆Bench →
TCAS

Consciousness-Relevant Assessment

What consciousness-relevant properties does the system exhibit? Validity-centered measurement with governance tiers.

You are here

One measurement surface. Two governance applications. Both tools produce evidence that makes AI systems more inspectable, testable, and governable.

Run TCAS Assessments

The TCAS framework is open source. Clone the repository and run assessments on your own models.

GitHub Repository Read the Paper