← Research Hub
Proxy Seduction Framework · Paper Pipeline · Queued

The Recursive Proxy

AI Observability, LLM-as-Judge, and the Limits of Self-Evaluating Systems. Observability dashboards relocate the proxy ladder rather than eliminating it. The instrument tasked with detecting the dip is susceptible to the same erosion as the practices it monitors.

QUEUED Held until OSP and IfM submissions clear Memo dated 2 May 2026

Genesis

How the paper concept emerged

A practitioner social media post on 2 May 2026 explained AI observability as the engineering response to bad LLM-based applications. The diagnosis it offered (system alive does not equal answer correct, monitoring tells you uptime not behavior) shared the structural form of PSF's opening move. The solution it offered, however, was not what it appeared to be.

Observability does not eliminate the proxy problem. It relocates the proxy ladder one rung up. Where monitoring substitutes "is it working" with "is it up," observability substitutes "is it producing good outcomes" with "does it score well on faithfulness, groundedness, answer relevance." Each of these scores is itself a model judgment computed by another LLM acting as judge. The judge is reading criteria the model can read, applying them to outputs the model produced, and the calibration of the judge is itself a product of engagement-shaped training. The loop closes around what models can recognize, which is precisely the variable PSF holds in question.

Thesis

AI observability emerged as the engineering response to a real diagnosis: traditional monitoring measures whether systems are alive but not whether their outputs are correct. The vendor stack now treats faithfulness, groundedness, and answer relevance as the missing layer. The architecture that produces these metrics is recursive. An LLM judges another LLM's output against criteria written for both. The loop closes around what models can recognize, and the calibration of the judge is itself a product of engagement-shaped training. The dashboard reports model opinion as measurement.

The recursive form is not avoidable in any system that uses models to evaluate models, and refusing all such systems is not a serious recommendation. What can be required is the conditions that make the recursion legible. Anthropic's Constitutional AI is the bounded case: published criteria, explicit acknowledgment of the recursion, training-time use with external scrutiny. Production observability vendors (LangSmith, Arize Phoenix, Galileo, Braintrust, Patronus, RAGAS, DeepEval) deploy the same form with criteria hidden, the recursion unmarked, and the metric treated as fact. The paper develops four conditions for legible recursion and argues that the current vendor stack fails each of them.

Three Architectural Connections to PSF

Connection 1 · Calculative apparatus

Cabantous and Gond fit

The dashboard is a calculative apparatus enacting a particular version of "AI quality" by making certain things visible (faithfulness 72%, latency 1.42s) and certain things invisible (whether the answer eroded the user's capacity to catch errors of this kind independently). The three mechanisms (conventionalising, engineering, commodifying) operate visibly across LangSmith, Arize, RAGAS, and the broader vendor stack.

Connection 2 · Engagement as criterion

User feedback widget

"User Feedback 2.8/5" and thumbs-up/down ratings on observability dashboards are the canonical engagement-as-criterion move PSF specifically targets. They treat users as stable evaluators of quality at the precise point where PSF posits the evaluative capacity is being eroded. The session that registered as a thumbs-up may be precisely the session that confirmed the erosion.

Connection 3 · Recursive proxy

LLM-as-judge architecture

"Groundedness 92%" is typically computed by another LLM acting as judge (RAGAS, DeepEval, LangSmith QAEvalChain, Patronus, Galileo, Braintrust). The judge is reading criteria the model can read, applying them to outputs the model produced, and the judge's calibration is itself an artifact of engagement-shaped training. The meta-evaluator is the kind of thing whose evaluative capacity is in question.

Constitutional AI as the Bounded Case

Anthropic's Constitutional AI (Bai et al. 2022) is worth naming because it makes the recursion explicit and bounded. The architecture is identical in form: model produces output, model judges output against written principles, judgment becomes training signal. Three features distinguish it from production observability stacks.

1. The criteria are inspectable

The constitution is published, contestable, revisable. Anyone can read what Claude is being trained against and disagree with it. Critics have. Vendor LLM-as-judge prompts are typically proprietary, embedded in evaluation libraries, and treated as implementation detail. The criteria the dashboard reports against are not surfaced to the user reading the dashboard.

2. The recursion is treated as the methodological problem

Anthropic argues for the design rather than reporting its outputs as measurements. The CAI paper is in part an argument about why model self-critique under explicit principles is preferable to opaque human aggregation. Vendor dashboards present groundedness scores as measurements, not as one model's opinion of another model's output under unstated criteria.

3. The loop is bounded

CAI is one stage in a pipeline that includes red-teaming, human helpfulness labels, evaluation suites, and external research scrutiny. Vendor LLM-as-judge runs continuously in production as the primary signal of "AI quality," and the dashboard treats its output as fact rather than as judgment.

Empirical Anchors (Named, Tractable)

Four Conditions for Legible Recursion

1. Written criteria the user of the metric can read

The judge prompt must be inspectable by whoever consumes the dashboard. Anthropic publishes the constitution. Vendor stacks do not publish the eval prompts.

2. Explicit acknowledgment that the metric is a model judgment

The dashboard must mark "groundedness 92%" as model opinion, not measurement. Current UIs do the opposite.

3. Periodic out-of-distribution calibration by humans holding the criterion

Sampled outside the loop, with explicit attention to confident wrong answers that the judge rated highly. Cannot be the same humans who tuned the eval prompt.

4. Refusal to treat the meta-evaluator as authoritative

On questions whose answers determine whether the meta-evaluator itself is functioning. The judge cannot adjudicate its own calibration.

Possible Venues

Big Data and Society

Constitutive critique of measurement infrastructure. Sociotechnical fit. Engages the named tools as artefacts rather than as black-box utilities.

Information and Organization

Organisational consequences of evaluation tooling. Recent sociomaterial turn fits the LLM-as-judge architecture argument.

AI and Ethics

Faster turnaround, less prestigious. Useful if the timing matters relative to the EU AI Act and AISI evaluation rollouts.

MIT Sloan Management Review

Practitioner crossover possible but defer until SMR queue clears (AI Alibi already targets SMR).

Sequencing Contingencies

Park. Revisit after OSP and IfM clear.

If a reviewer surfaces the observability question during OSP or IJMR review and the response from the reviewer-response stub lands well, that exchange becomes the seed of the paper, and the framing sharpens to whatever the reviewer's specific concern was.

Otherwise, the natural window opens after the OSP and IJMR submissions are out and the IfM First Year Conference is complete. At that point the empirical surface is unusually tractable: LangSmith, Arize, Galileo, Braintrust, Patronus, RAGAS, DeepEval are all named, documented, instrumentable.

Reviewer Response Stub (Held in Reserve)

A short defensive reviewer response paragraph is drafted and held in the responses-to-reviewers reserve. The point is to have a ready answer if a reviewer asks "what about LLM-as-judge?" or "doesn't observability solve this?" without bolting the argument onto the OSP paper.

Open Questions