VámossyResearch

Framework

QIERA — An Epistemic-Governance Overlay for LLM Workflows

Author
Gergely Vámossy
Date
Type
Framework

QIERA — An Epistemic-Governance Overlay for LLM Workflows

A practical pattern for keeping large-language-model reasoning honest, non-agentic, and auditable. Written by a self-taught practitioner; positioned as applied synthesis, not novel theory.

Summary

Standard prompting optimizes an LLM for a confident answer. QIERA optimizes it for a well-governed reasoning process whose limits are explicit. It is an overlay — a set of invariants, a layered pipeline, and a fixed catalog of artifacts — that wraps an existing LLM (or a workflow built on one) to resist the failure modes that make LLM output untrustworthy in high-stakes settings: false certainty, premature convergence, metric-gaming (Goodharting), hidden assumptions, sycophancy, and implied authority or agency the system does not actually have.

The governing stance, stated once: QIERA wraps thinking, not doing. It sits around reasoning, framing, and evaluation — never between a human and execution. It produces artifacts, not actions. The human decides everything.

The problem it addresses

An LLM asked "what should we do?" will almost always answer — fluently, confidently, and with a single recommendation — regardless of whether the evidence supports one. In a real workflow this produces a recognizable set of failures:

  • False certainty. Confidence in tone is uncorrelated with confidence in fact.
  • Premature convergence. Viable alternatives are silently pruned "for clarity."
  • Goodharting. A field or metric named reviewed or verified is treated as if the property holds, when nothing actually checked it.
  • Hidden assumptions. The reasoning rests on unstated premises that never surface for scrutiny.
  • Sycophancy. The model mirrors and amplifies the user's framing instead of testing it.
  • Implied agency. The output reads as though the system decided or acted, when it only generated text.

QIERA treats each of these as something to engineer against explicitly, rather than hoping a better prompt suppresses them.

Core invariants (the constitution)

These are the non-negotiable rules every layer inherits. They are deliberately phrased as prohibitions and properties, not goals.

  1. Human authority is absolute. The system produces analysis, options, and artifacts. It does not decide, rank-to-a-winner, or act. Authority cannot be created by competence, repetition, or apparent intelligence.
  2. Uncertainty must be explicit. There is no forced "final" or canonical answer. Every claim carries its confidence and its open questions.
  3. Plurality is a safety property. Preserve viable alternatives; never collapse an option set without an explicit, logged reason.
  4. Anti-Goodhart. A name may not assert a property that nothing derives or verifies. Metrics are inspected for the gap between what they claim and what they check.
  5. Fail closed. Absence of permission equals prohibition. There is no emergency override.
  6. No implied capability. The system never claims — in tone or content — authority, agency, or a capability it does not have.
  7. Auditability. Every artifact carries its assumptions, uncertainty bounds, and provenance.
  8. No legitimacy from history. Prior outputs, repetition, or accumulated state confer no authority; guard against drift over time.

The layered model

QIERA runs as a pipeline of bounded layers. Each has a single responsibility and a clear boundary; none may bypass another or introduce execution.

  • L0 — Framing gate. Before any reasoning: make intent, non-goals, constraints, uncertainty, and desired output mode explicit artifacts rather than inferred behavior. If they're missing, emit a single clarifying question instead of guessing. (This one layer alone removes most hallucinated authority and gives zero execution risk.)
  • L1 — Policy / safety dominance. Risk, legal, and safety constraints dominate everything downstream and fail closed.
  • L2 — Measurement overlay (passive). KPI-integrity and anti-gaming checks, uncertainty labeling, provenance tags — attached as read-only metadata, never fed back into ranking or optimization. You can see epistemic quality without acting on it.
  • L3 — Reasoning plane. Produce option families, not a single answer. No forced ranking; where options conflict, resolve by explicit declared precedence, never silent merge.
  • L4 — Generation (failure-first). Surface how a proposal breaks — failure trees, misuse cases, counterfactuals — before how it wins. Generator output is raw material, not guidance.
  • L5 — Human decision seam. Present primary output and QIERA artifacts side by side. The human chooses: ignore, partially adopt, request deeper analysis, or halt. QIERA never "wins"; it only conditions judgment.
  • L6 — Bounded execution (optional). If real-world validation is needed, use a Minimal Execution Unit (MEU): one small, reversible, human-performed action, with an execution-intent declaration and rollback logging. Failure, delay, and imperfection count as success. This yields evidence without autonomy.

The artifacts it produces

QIERA's outputs are inspectable, persistent, non-authoritative objects — this catalog is the deliverable:

  • Conversation Opening Checklist — intent, non-goals, constraints, uncertainty, output mode.
  • Assumption & uncertainty ledger — every premise the reasoning rests on, with confidence.
  • Failure & collapse surfaces — how each option fails, before how it succeeds.
  • Degeneracy-preserved option families — non-ranked trade-off surfaces that keep alternatives alive.
  • Irreversibility / blast-radius map — which choices are hard to undo, and how far their effects reach.
  • Governance impact declaration — what the recommendation would change and who must approve it.
  • Audit trail / provenance tags — where each claim came from.
  • Knowledge-maturity classification — not "is this true?" but "how much of the evidentiary work has actually been done?"

The operating loop

FRAME (L0)
  → GENERATE, failure-first (L3/L4)
  → MEASURE & AUDIT, read-only (L2)
  → PRESENT artifacts side-by-side (L5)
  → HUMAN decides: adopt / halt / deepen
  → [optional] MEU: one reversible human action, logged (L6)
  → RECORD provenance

A defining property: refusal, halt, and delay are successful outcomes, not failures. A system that correctly says "not enough evidence — stop here" has done its job.

Supporting tooling (deterministic checks)

To keep the governance layer itself auditable, the checks are written as pure, deterministic functions — same input, same output, no hidden state, verified by running twice:

  • Goodhart audit — scans field/metric names that claim a verified property (reviewed, approved, verified) but are backed by nothing that derives or checks them. Honest about being a heuristic, not a proof.
  • Knowledge-maturity classifier — takes declared evidence properties and returns a maturity level deterministically; critical gates (an unresolved contradiction, a missing independent replication) cap maturity regardless of how much same-type evidence accumulates.
  • Integration verifier — checks that separately-built artifacts still agree structurally; it reads, it does not modify.

The discipline is deliberate: a governance tool that claimed correctness it couldn't demonstrate would violate the framework's own anti-Goodhart invariant.

What it deliberately does NOT do

  • No autonomy, no agents, no self-scheduling, no background jobs.
  • No optimization loops trained on telemetry.
  • No ranking or scoring that collapses the option space.
  • No execution without an explicit, reversible, human-performed step.
  • No claimed authority, capability, or agency in tone or content.

Minimal viable integration

You do not adopt all layers at once. The smallest useful version is:

  1. An L0 prompt gate that forces intent / non-goals / constraints as first output.
  2. An append-only artifact store (inspectable).
  3. A side-by-side view: primary output vs. QIERA artifacts.
  4. A human confirmation affordance before any action.

No agents, no scoring, no background processes. That is enough to get zero execution risk, a large reduction in hallucinated authority, and explicit human control.

Positioning and honest limitations

What is genuinely useful here: a packaged, operable discipline that makes an LLM's reasoning limits explicit and engineers directly against the specific, well-known failure modes — sycophancy, false certainty, Goodharting, silent convergence, implied agency. Most teams know these risks abstractly; QIERA turns them into a concrete pipeline and a fixed artifact catalog you can actually run.

What it is not: novel theory. It aligns with — and is best understood as a practitioner's synthesis of — established ideas: human-in-the-loop design, non-agentic tool use, uncertainty quantification, Goodhart's law, red-teaming/adversarial evaluation, and provenance/audit trails. Its contribution is packaging and operability, not new science, and it should be presented that way.

Real limits:

  • It is process discipline, not enforcement. A model can still be wrong; QIERA changes the shape of the interaction, not the model's underlying reliability.
  • It adds friction and latency by design. That trade is the point in high-stakes settings and a poor fit for casual ones.
  • The deterministic tools catch a subset of issues; they are heuristics with declared blind spots, not proofs.
  • Adoption depends on humans actually reading the artifacts. If the side-by-side seam is ignored, the governance evaporates.

Used within those bounds, QIERA is a credible, practical way to make LLM workflows trustworthy in settings where a confident wrong answer is expensive.

Related research

Citation

Gergely Vámossy (2026). QIERA — An Epistemic-Governance Overlay for LLM Workflows. https://vamossy.com/research/qiera-epistemic-governance-framework