Thesis note · whitepaper in progress

ARCH

Autonomous Reasoning & Causal Hierarchy

Vaishnavi Yeruva · July 2026 · vaish.ai

§1 · TL;DR

The demo-to-production gap in agentic AI is not a capability gap; it is a reasoning-reliability gap: models answer interventional questions with associational machinery, and that difference is where production failures live. ARCH is a metacognitive executive layer for frozen models: an explicit structural causal model, a typed working-memory buffer, and an RL-trained controller that decides when to reason harder. The base model is never retrained, and no conclusion commits without passing a counterfactual consistency check.

§2 · The Problem

I build production AI systems. At one of the world's largest consumer internet platforms, I built agentic root-cause-analysis systems; more recently I re-engineered a Fortune 500 bank's development lifecycle around agentic tooling, taking feature cycle time from a week to a day. Across both, the production failures I saw were rarely capability failures. The model knew the relevant facts. It recommended the action associated with success in data like its training data, not the action that would cause success here.

Deployment is an intervention. When an agent acts, it perturbs the system it learned from: traffic shifts, users adapt, dependencies see new load. A model trained on observation answers P(Y | X): what tends to hold when X is seen. The production question is P(Y | do(X)): what becomes true when X is done. The two diverge exactly where confounders live, and more data does not close the gap. This is standard Pearl. Agent stacks mostly ignore it.

Figure 1. One confounder, two queries. Observation leaves the backdoor path open; intervention cuts it. Every claim in this note traces back to this distinction.

The reliability gap, stated plainly: agents commit to conclusions their machinery cannot support. They do not know which rung of the causal hierarchy they are standing on.

§3 · The Framework

ARCH is four components around one loop. The base model proposes; the executive decides; the gate permits.

The proposer. A frozen foundation model. It generates hypotheses, decompositions, and candidate causal structure. It is never fine-tuned. ARCH treats it as a capable but unreliable witness.

The causal layer. An explicit SCM over task variables, elicited from the proposer and refined against evidence. Do-calculus runs here: identification, interventional estimates, and abduction-action-prediction for counterfactuals. The graph is a live hypothesis, not an oracle.

The working-memory buffer. Not a scratchpad. A schema-constrained state object outside the context window: the current graph hypothesis, competing hypotheses with scores, counterfactual traces, and an evidence ledger. The proposer receives serialized views of it; only the executive writes to it. A scratchpad is free text the model owns. The buffer is typed state the executive owns, persistent across calls.

The executive. A learned controller (PPO/GRPO) over a discrete action space: elicit or refine structure, query the proposer, run a counterfactual check, commit, abstain. Its reward design is the open problem in ARCH. Two candidate directions: verifiable downstream outcomes where tasks provide them; abstention-calibrated accuracy where they don't. I state this as unsolved because it is.

Figure 2. The ARCH loop. The executive owns state and control; the proposer is a stateless witness; nothing commits without passing the gate. Rust marks the places where the design is honest about risk: the do-operator's cuts and the open reward problem.

The gate. A conclusion commits only if it survives abduction-action-prediction under the current graph: the counterfactual the conclusion implies must be consistent with observed evidence. Conclusions that fail send the executive back into the loop to refine the graph and gather evidence, or terminate in abstention.

Figure 3. The gate criterion. The counterfactual a conclusion implies is computed by abduction-action-prediction and checked against what was actually observed. Failure re-enters the loop; it never commits silently.

"Hierarchical" means climbing Pearl's ladder deliberately: associational retrieval, then interventional estimates, then counterfactual checks, rather than answering Rung-3 syntax with Rung-1 machinery.

Figure 4. Two responses to unreliable reasoning. Scaling test-time compute buys more of rung 1. ARCH climbs, and the gate sits at the top. Claim C2 makes this picture falsifiable.

Stated limitation

The gate is only as sound as the graph it runs on. A wrong elicited graph produces confidently wrong counterfactuals. ARCH converts "wrong answer" into "wrong answer with a legible causal justification": auditable, not automatically correct. Reducing graph error is an evaluation target, not a solved problem.

§4 · Falsifiable Claims

Claim C1 · errors convert to abstentions under novel confounding

Inject a deployment-time confounder absent from the training-like distribution. A correlational agent's error rate rises; an ARCH-gated agent's abstention rate rises instead, with its confident-error rate roughly flat.

Pivots on: whether the elicited graph captures the new confounder.

Falsified if: gated agents fail confidently at the same rate as ungated ones.

Claim C2 · test-time compute does not climb the ladder

Scaling samples, chain length, or search closes gaps on Rung-1 tasks but not on Rung-2/3 tasks; adding explicit causal structure does.

Pivots on: ladder failures being structural rather than compute-bounded.

Falsified if: a pure test-time-compute method matches a causally structured agent on interventional queries.

Claim C3 · the gate carries the calibration

Ablate only the counterfactual gate, keeping the graph and the buffer. Confidence calibration on causal queries degrades more than raw accuracy does.

Pivots on: the gate specifically, not the graph, being the source of calibration.

Falsified if: calibration is unchanged without the gate.

§5 · Related Work

Borrowed, wholesale: Pearl's causal hierarchy and identification machinery: SCMs, do-calculus, abduction-action-prediction. ARCH adds no new causal theory and claims none.

Adjacent: the Bengio line of causal representation learning asks how to learn causal variables from raw data. ARCH assumes the variables are nameable, as they usually are in agent tasks, where tools, metrics, and entities are explicit, and asks how to control inference over them. The two are complementary; progress in representation learning would upgrade ARCH's weakest input, the elicited graph.

Distinct from: test-time compute and process supervision. Sampling more chains, searching deeper, or rewarding better traces scales and audits the reasoning without changing its epistemic type: more Rung 1, checked Rung 1. ARCH's bet is that type, not volume, is the binding constraint on reliability.

Claimed as new: framing agent reliability as a control problem over a frozen proposer; the executive / buffer / gate architecture; counterfactual consistency as the commit criterion. If any of these have prior art I've missed, I want to know.

§6 · Evaluation Plan

ARCH will be stress-tested by a dedicated reliability evaluation suite, currently in design. Three failure modes it will measure: confident-error rate under injected confounding (does the agent fail loudly or quietly when causal structure shifts); the calibration gap between associational and interventional queries (does confidence track the rung); and abstention appropriateness (does the agent abstain when its graph cannot support a conclusion, and only then).

§7 · Status

ARCH is at framework stage. Specified: the buffer schema and the gate criterion. Open: the executive's reward design. Designed, not implemented: the evaluation suite. No results exist yet, and none are implied above. The whitepaper, covering the full formalism, executive training setup, and evaluation protocol, is in progress.

If you disagree with the claims, I want to hear it.