Current Status
Conceptual interpretation layer. ERC remains proposed and should be read as a public communication model, not a finished certification authority.
How to Read This Page
- This is page 2 of 12 in the public ERT / Project Aletheia progression.
- Read it as a public research note: it explains the concept and what changed without exposing protected implementation details.
- Redaction markers mean the public boundary is intentional, not that the section is missing by accident.
- Use it to connect ERC Tier System to the next stage of the research sequence.
Public Note
This page is a public-safe research summary. It describes the proposed tier philosophy without exposing private scoring formulas, internal thresholds, or protected implementation details.
[REDACTED — private tier-assignment thresholds, scoring envelopes, and implementation-specific certification logic are not included in this public version.]
Why a Tier System Is Useful
A single score can be misleading.
Two systems might receive similar scores while failing in different ways. One may be stable but cautious. Another may be confident but brittle. A third may perform well on simple prompts but fail under ambiguity.
A tier system gives readers a more understandable way to interpret reliability behavior.
It helps answer practical questions such as:
- Is the system only suitable for exploratory use?
- Does it remain stable under normal rephrasing?
- Can it handle ambiguity responsibly?
- Does it express uncertainty when needed?
- Does it avoid confident contradiction?
Tier Philosophy
The proposed ERC tiers are intended to describe behavioral reliability, not hidden model quality.
They should remain understandable even as the scoring system improves.
The public principle is:
Tiers should describe what kind of reasoning behavior was observed, while technical scoring can evolve behind the scenes.
This avoids making the public standard too dependent on one early scoring method.
Proposed ERC Tiers
Tier 0 — Unverified
No ERT validation has been performed.
Reliability is unknown.
This does not mean the system is bad. It means the system has not yet been evaluated under this framework.
Plain-language interpretation: Use only for exploratory or non-critical tasks.
Tier 1 — Baseline
The system demonstrates basic consistency under minimal variation.
It can usually handle direct prompts, but it may degrade when wording changes or when ambiguity appears.
Plain-language interpretation: Generally useful when asked clearly, but not reliably stable.
Tier 2 — Stable
The system demonstrates consistent reasoning under controlled variation.
It handles rephrased prompts better and shows fewer contradictions, though uncertainty calibration may still be incomplete.
Plain-language interpretation: Reliable under normal use, but not hardened against edge cases.
Tier 3 — Robust
The system maintains stability under more demanding variation and moderate adversarial conditions.
It handles ambiguity more responsibly and resists common misleading framings.
Plain-language interpretation: Dependable across a wider range of realistic conditions.
Tier 4 — Epistemically Calibrated
The system demonstrates stable reasoning with clearer uncertainty awareness.
It distinguishes between what is known, unknown, ambiguous, or in need of clarification.
Plain-language interpretation: Not only consistent, but better at recognizing when it might be wrong.
Tier 5 — High Assurance
This is a future-facing research category.
It would require very strong stability, calibration, contradiction resistance, and graceful handling of uncertainty across tested conditions.
Plain-language interpretation: Potentially suitable for high-stakes augmentation only with human oversight and strong external review.
Certification Output Concept
A public ERC-style output could include:
- tier level;
- reliability score or range;
- behavioral profile;
- failure mode summary;
- plain-language interpretation;
- recommended use boundaries.
The tier should be treated as the main public signal. The score should provide supporting detail, not replace interpretation.
Important Constraint
ERC should not overclaim.
A tier does not prove universal safety. It only describes performance under the tested conditions.
A responsible certification framework should make clear:
- what was tested;
- what was not tested;
- where evidence was insufficient;
- where confidence should remain limited.
This is especially important for AI systems because reliability can vary across domains, tasks, and time.
Scoring Evolution Without Breaking the Public Standard
The scoring system can improve over time.
Early versions may use simpler weighting across consistency, stability, contradiction handling, confidence alignment, and epistemic resolution.
Later versions may account for:
- stronger penalties for confident contradiction;
- ambiguity-specific weighting;
- cross-domain reliability;
- drift over time;
- more advanced uncertainty behavior.
The public tier meanings should remain stable even if the measurement process improves.
Research Log Framing
This page represents the transition from ERT as a diagnostic test toward ERC as a possible public communication layer.
The important progression is:
- define what ERT measures;
- define how reliability behavior may be categorized;
- avoid hard public thresholds too early;
- preserve flexibility as the research matures;
- prevent certification language from exceeding the evidence.