The Missing Variable in AI Training: Schema Coherence and the H-Bar Model

Physics-Informed Residual Learning (PIRL) is a hybrid control architecture for DC motor actuation under nonlinear Stribeck friction. It combines a PID feedback term, an analytical Stribeck friction model acting as a physics prior, and a compact neural network that learns only the residual mismatch between the prior and the true plant. In simulation, it converges in under five epochs where unconstrained baselines need 100+, reaching acceptable performance with fewer than 500 training samples. By every standard metric, it is a high-competence agent within its domain.

Now move the operating conditions outside the Stribeck model’s domain — to adhesion-dominated contact at elevated normal forces, drivetrain backlash with load-dependent inertia variation, or sustained thermal drift in winding resistance. Hardware validation is currently underway. But the architecture already makes a specific structural prediction: failure will not spread uniformly across the out-of-distribution (OOD) manifold. It will concentrate at the boundaries where the Stribeck model’s causal assumptions break down, while in-distribution tracking performance stays intact. The model’s explicit causal prior is what makes this prediction possible. For purely learned models — where no analytical prior is available — this article develops the tools for making the same prediction from internal representational evidence alone.

Two diagnoses are possible when an agent like this fails OOD. The first: insufficient parametric capacity — the network lacks the depth to absorb novel dynamics. Fix: scale up parameters or data. The second: structural misalignment — the network’s representations are organized around the prior’s velocity-force-friction causal structure, which is actively misleading outside its domain. Scaling does not change what the representations are organized around. These two diagnoses predict different error signatures — uniform under the capacity account, clustered at causal boundaries under the structural account — and they call for different interventions.

The PIRL case is a controlled instance of a failure pattern the compositional generalization literature documents at scale. Models trained to 96–99% in-distribution accuracy on COGS achieve only 16–35% on systematic generalization splits.¹ The SCAN systematic split failure is robust across architectures and training scales.² The standard fix — more data, larger models, longer training — has been applied extensively and the failure persists. H-Bar’s claim: persistence under scaling is the signature of the structural diagnosis, not the capacity one. These models do not lack depth. Their depth is not organized around the causal variables that generate the task structure.

The framework this article introduces provides three things the existing literature does not.

First, a formal variable σ that captures the degree to which an agent’s representations are organized around generative causal structure rather than surface statistical regularities — with explicit developmental dynamics that can be tracked during training. Causal representation learning³ defines the target state. Invariant Risk Minimization⁴ trains toward it. Neither provides a developmental account. σ has phase-indexed growth conditions, suppressible dynamics, and delegation consequences that no existing construct models.

Second, a phase structure — five training states defined over the joint $(\delta, \sigma)$ profile of the agent. Each state makes a different training prescription optimal. Each transition is triggered by a threshold condition on σ, not by performance saturation or loss-curve events. Khetarpal et al. (2022) named this gap explicitly: the field lacks a method for designing curricula that deliberately target structured representation development rather than treating it as a byproduct of sufficient depth.⁵ The phase structure this article proposes is a direct response to that open problem.

Third, a delegation criterion — a formal account of when routing tasks to an external retrieval system is net-positive and when it is net-negative. Standard adaptive retrieval systems — Self-RAG,⁶ FLARE,⁷ and comparable tool-use frameworks — condition delegation on the agent’s parametric performance gap. They retrieve when internal knowledge is insufficient. H-Bar’s criterion is different: the relevant variable is σ in the target domain, because σ determines whether the agent can evaluate, integrate, and detect errors in retrieved content. The two criteria agree when σ is high and diverge when σ is low. This generates a non-monotonic prediction: delegation is net-positive above $\sigma_\text{critical}$ and net-negative for compositional tasks below it. No existing adaptive retrieval architecture models this.

---

What Current Training Pipelines Miss

Two agents share an identical architecture and are trained on identical data from the same motor. After training, they achieve indistinguishable tracking RMSE on held-out in-distribution data. On a new motor with different winding impedance but Stribeck-compatible friction dynamics, one agent adapts in three epochs. The other needs sixty. No standard evaluation metric predicts this difference. It is a prediction of the framework developed in this section.

The dominant training paradigm optimizes a single objective: how well an agent’s representations fit a training distribution and generalize to nearby test points. When generalization falls short, the canonical interventions are more parameters, more data, and longer training. This is internally coherent — but it conflates two properties that have different dynamics, different failure modes, and different remedies.

The first property is parametric depth ($\delta$) — how much of the domain’s behavior an agent has encoded. The second is whether those encodings are organized around the variables that generate the domain’s behavior, or around correlations that happen to co-occur with it in the training distribution. Standard metrics track the first. They are blind to the second. Two agents can have equal held-out RMSE and radically different internal organization. That difference only surfaces when the distribution shifts in ways that preserve causal structure while breaking surface correlations — a different motor, a different load, a compositionally novel input.

The framework introduces schema coherence ($\sigma$) to name the second property and to formalize what happens when it is missing.

Three operational commitments make σ non-trivially different from existing ML concepts.

σ is measured by transfer behavior, not representational geometry. Two agents can have identical activation geometry under probing yet different σ if one’s geometry clusters by causal variables and the other’s by training artifacts. The test is not how representations look. It is how they behave when the statistical surface shifts while causal structure is preserved. An agent with high σ in friction dynamics should adapt faster to a new motor whose Stribeck parameters differ from training but whose friction-regime structure does not. An agent with low σ should show no adaptation advantage over an agent with zero prior domain knowledge — its representations encode the wrong variables.

σ is not reducible to inductive bias, and it is not disentanglement. A physics-informed loss constrains the hypothesis space during training — that is the inductive bias. σ is what the agent has encoded as a result. These can come apart. An agent trained with a strong physics prior can have low σ if the prior’s constraints were satisfied by surface fitting without explicitly representing causal variables. Disentanglement may be a consequence of high σ — factors get separated because causal variables are separable — but σ is the upstream property that generates the factorization. Intervening on geometry does not intervene on σ.

σ has dynamics that decouple from depth, and those dynamics are suppressible. An agent trained on data from one friction regime builds representations organized around that regime’s correlation structure. Depth grows. σ does not, because the training objective never required causal variables to be the organizational basis. Contrast this with a training regime that forces prediction across causally varied conditions — different friction regimes, varied normal forces, both velocity directions. Only variables that generalize across all those conditions provide a reliable training signal. The prediction: agents trained with this protocol adapt faster to OOD motors than agents trained on equivalent data from a non-symmetric distribution, at matched RMSE.

The Primary σ Proxy: Learning-Rate Discontinuity

The most operationally useful proxy for $\sigma_\text{critical}$ does not require pausing training to run held-out evaluations. It is a feature of the learning curve itself.

Above $\sigma_\text{critical}$, new training examples can be integrated into the agent’s causal encoding rather than appended as new statistical correlations. Integration is more efficient. The prediction: crossing $\sigma_\text{critical}$ should produce a detectable discontinuous increase in $\frac{d\delta}{dt}$ — depth growth rate — relative to the pre-threshold trend, given equivalent training inputs. This is the primary phase-transition indicator.

The connection to grokking⁸ is testable. If the σ proxy crosses threshold at the same point as the grokking transition, this supports a σ-based account of the grokking mechanism. If they dissociate, the two phenomena are empirically distinct. The PIRL convergence advantage — under five epochs vs. 100+ for the unconstrained baseline — is consistent with this mechanism: the Stribeck prior injects σ by architecture, placing the agent in a post-$\sigma_\text{critical}$ state from the first training step.

Secondary proxies include above-chance performance on a held-out structural transfer task before in-domain performance reaches a ceiling, and probing classifier accuracy organized by causal rather than artifactual variables. These are more expensive. The learning-rate proxy is the first thing to monitor.

The PIRL Contrast

In a well-trained PIRL controller, σ is high because the architecture makes causal-variable encoding the path of least resistance. The Stribeck model handles the three-regime velocity-force relationship. The residual network learns only what the causal model gets wrong. To minimize loss under this constraint, the residual’s activations must track the mismatch between causal prediction and actual dynamics — and that requires encoding the causal variables that determine where the mismatch is large.

The falsifiable prediction: probing the residual’s hidden layers will show clusters organized by Stribeck-interpretable features — velocity regime, friction-regime boundary proximity, and zero-crossing distance. If clusters instead organize by data collection timestamp, sequence order, or other training artifacts, the σ claim is falsified for this architecture.

In a model-free controller trained on the same data, depth can match or exceed PIRL’s. But the representations achieving that fit are organized around the dominant correlations of a single-motor training distribution. OOD adaptation will require substantially more epochs because nothing in the existing representations provides a causal scaffold for the new motor.

---

The Five Training Phases

A clarification is needed before the phase structure is introduced. ML training already uses the term “phase” — typically to label early memorization, late generalization, or the grokking transition. H-Bar’s phases are different in kind. Grokking identifies a post-hoc empirical observation: a sharp generalization transition visible in the loss curve, found by inspection after training. H-Bar’s phases are prescriptive states defined over the joint $(\delta, \sigma)$ profile of the agent. Each state makes a different training prescription optimal. Each transition is triggered by a threshold on the agent’s internal representational variables, not by elapsed steps or loss-curve events.

One honest limitation: $\sigma_\text{critical}$ is not directly observable with current evaluation instruments. The learning-rate discontinuity proxy is the primary in-training indicator. The framework’s value is that it specifies what would need to be measured and predicts how that measurement correlates with training outcomes.

Curriculum learning (Bengio et al., 2009)⁹ optimizes the growth rate of parametric depth. It is the correct prescription in regimes where $\sigma \approx 0$ throughout. H-Bar subsumes this as the low-$\sigma$ special case. The crossing of $\sigma_\text{critical}$ is a structural event in the agent’s representations, not a performance event visible on the loss surface. No difficulty schedule — monotonic or adaptive — can detect it or respond to it appropriately.

The mastery set $\mathcal{M}_A(t)$ is the set of domains in which both conditions are satisfied simultaneously:

$\mathcal{M}_A(t) = \left\{ d \;\middle|\; \delta_A(d, t) \geq \delta_\text{critical} \wedge \sigma_A(d, t) \geq \sigma_\text{critical} \right\}$

The phase at time $t$ is determined by the agent’s joint $(\delta_A, \sigma_A)$ position. Drag the marker in Figure 1 to explore how the training prescription and failure mode change across the phase space.

Figure 1 — The (δ, σ) Phase MapDrag the marker · touch supported

Phase 1 — Depth Accumulationδ = 0.35 · σ = 0.25

Training prescription

δ is the growth-limiting variable. Easy-to-hard difficulty scheduling is the correct heuristic here. Simultaneously: ensure training batches require causal-variable encoding, not just surface correlation matching.

Failure mode if stuck here

High in-distribution benchmark performance, brittle compositional generalization. SCAN/COGS failure pattern. Standard fix (more data, more epochs) reduces error magnitude without changing error structure.

Figure 1. The $(\delta, \sigma)$ phase map. Drag the marker to explore all five training phases. The dashed lines mark $\delta_\text{critical}$ and $\sigma_\text{critical}$. Phase 2 occupies the upper-left quadrant — schema crystallising while depth is still moderate — and is the phase most standard pipelines either miss entirely or pass through without deliberate intervention.

Phase 0 — Initialisation

$\delta \approx 0$, $\sigma \approx 0$. No stable domain-specific representations have formed. Curriculum order matters most here: a Phase 0 curriculum that exposes causal variables before their surface correlates should produce measurably different downstream σ trajectories than random sampling from the same data, at equal volume. The difference does not appear in Phase 0 in-domain loss — it accumulates and surfaces only when Phase 2 structural pressure arrives. The Symmetric Harvesting Protocol in PIRL — enforcing balanced data representation across both velocity directions before either dominates the correlation structure — is a Phase 0 intervention of this type.

Phase 1 — Depth Accumulation

$\delta$ growing, $\sigma < \sigma_\text{critical}$. The growth-limiting variable is $\delta$. Easy-to-hard difficulty scheduling is the correct heuristic here — H-Bar does not dispute Bengio-style curricula in this regime. What difficulty scheduling misses: while $\delta$ is the growth-limiting variable, the curriculum simultaneously determines which representations get built. If every training batch in Phase 1 can be solved without encoding causal variables, depth grows while σ stays near zero. This is precisely the SCAN/COGS failure pattern: high in-distribution accuracy, brittle systematic generalization. More data reduces error magnitude without changing error structure.

Phase 2 — σ Crystallisation

$\sigma$ crosses $\sigma_\text{critical}$ in at least one mastery domain. This is the first phase transition that standard pipelines have no mechanism to detect or respond to. The learning-rate discontinuity is the primary signal: above $\sigma_\text{critical}$, examples integrate into causal encoding rather than appending as statistical correlations, and $\frac{d\delta}{dt}$ should show a detectable step-change.

Phase 2 is also where cross-domain transfer first becomes meaningful. Below $\sigma_\text{critical}$, structural similarity $\phi(d_1, d_2)$ between domains cannot generate useful transfer — there is no causal encoding in the source domain to carry across. Above $\sigma_\text{critical}$, source-domain causal structure becomes a scaffold for faster representation-building in high-$\phi$ target domains.

Critical failure mode: introducing cross-domain transfer tasks before σ_critical is reached in the source domain transfers statistical correlations, not causal structure. H-Bar predicts that variance in multi-task transfer outcomes correlates with the source-domain σ proxy at the time of cross-domain exposure.

Phase 3 — Near-Frontier Depth

$\delta$ approaching the frontier in mastery domains, $\sigma \geq \sigma_\text{critical}$. In-domain benchmarks are high; further in-domain training shows diminishing returns.

This is where most standard pipelines terminate. Treating benchmark saturation as “training complete” produces an agent with high δ and high σ that has never been exposed to conditions activating the intersection mechanism. Its performance profile matches a high-σ specialist on in-domain benchmarks and shows zero above-additive compositional benefit — indistinguishable from a capacity-limited agent until cross-domain testing is applied.

What H-Bar predicts at Phase 3: the growth-limiting variable has shifted from $\delta$ to the cross-domain interaction term. The high-value training signal is structured cross-domain exposure to domain pairs with high $\phi(d_1, d_2)$ and $\sigma \geq \sigma_\text{critical}$ in both.

Phase 4 — Intersection Activation

Above-additive cross-domain performance becomes measurably positive — $\Psi > 0$. The distinctive prediction is multiplicative σ-dependence: halve σ in one contributing domain and $\Psi$ drops by approximately half, not incrementally. This is the claim that distinguishes H-Bar from general multi-task learning, which predicts additive benefits from domain combination regardless of σ distribution.

Phase 5 — Frontier Operation

The agent operates at or near the knowledge frontier, with high $\delta$ and high $\sigma$, actively engaging $\Psi$ across domain pairs. δ growth is now generative rather than acquisitive. Maintaining mastery status means tracking frontier advancement — the concern is not absolute depth but relative depth $\delta_A / \Delta(d, t)$.

---

The Delegation Gradient

Every RAG system and tool-use agent faces the same design question: which knowledge should live in parametric weights and which should be retrieved at inference time? The standard answer: retrieve when the external system outperforms the model on the query. The standard practice: expand delegation aggressively, because retrieval quality is improving faster than parametric training efficiency.

H-Bar’s claim is that this criterion is incomplete in a specific, testable way.

The Standard and H-Bar Criteria Compared

Under the standard criterion, the delegation set $\mathcal{D}^*$ is defined as:

$\mathcal{D}^*_\text{standard}(d, t) = \left\{ s \in d \;\middle|\; \delta_{\text{AI}}(s, t) \geq \delta_A(s, t) \right\}$

Delegate sub-skill $s$ when the external system outperforms the agent’s parametric knowledge on that sub-skill. H-Bar adds a single gate:

$\mathcal{D}^*_{\text{H-Bar}}(d, t) = \left\{ s \in d \;\middle|\; \delta_{\text{AI}}(s, t) \geq \delta_A(s, t) \wedge \sigma_A(d, t) \geq \sigma_{\text{critical}} \right\}$

Sub-skills that pass the performance criterion are still excluded from delegation if the agent’s schema coherence in the domain is below $\sigma_\text{critical}$. The reason: the agent cannot evaluate what is retrieved in that domain, so the effective depth gain from retrieval approaches zero and the risk of accepting causally inconsistent results is high.

This is not a theoretical concern. The test is direct: present agents differing only in σ proxy level with retrieved results that are (a) statistically plausible and causally consistent, and (b) statistically plausible and causally inconsistent. The high-σ agent should discriminate between categories above chance. The low-σ agent should not.

Effective Depth Under Delegation

The effective depth realized from delegation is σ-gated:

$\delta_\text{eff}(s, t) = \delta_A(s, t) + \Phi_A \cdot f\!\left(\delta_{\text{AI}}(s, t),\; \sigma_A(d, t)\right)$

where $\Phi_A \in [0,1]$ is integration fluency and $f$ satisfies $f \to 0$ as $\sigma \to 0$. High retrieval quality $\delta_\text{AI}$ does not increase effective depth when $\sigma \approx 0$, regardless of integration fluency. The linear form $f = \sigma \cdot \delta_\text{AI}$ satisfies this constraint; the threshold form $f = \delta_\text{AI} \cdot \mathbf{1}\!\left[\sigma \geq \sigma_\text{critical}\right]$ is more consistent with the discontinuous phase structure.

The Non-Monotonic Prediction

The central empirical prediction: holding retrieval quality fixed, increasing delegation rate improves end-task performance above $\sigma_\text{critical}$ and reduces it for compositional tasks below it.

$\frac{\partial \text{Acc}_{\text{comp}}}{\partial \rho} = \begin{cases} > 0 & \text{if } \sigma_A(d,t) \geq \sigma_{\text{critical}} \\[4pt] < 0 & \text{if } \sigma_A(d,t) < \sigma_{\text{critical}} \end{cases}$

This prediction applies in two contexts. During training: high-quality retrieval available as a primary answer source in low-σ domains suppresses σ development — the agent achieves low training loss through retrieval rather than through causal encoding. The loss trajectory looks identical to a retrieval-disabled agent. Downstream generalization diverges. At inference time: for an already-trained agent, increasing delegation in low-σ domains degrades performance on compositional queries because the agent accepts causally inconsistent results it cannot verify.

Figure 2 — The Delegation Gradient 𝒟*(d, t)

AI capability42%

Safe to delegate — σ ≥ σ_critical

Delegation trap — AI outperforms but σ < σ_critical

Human frontier — σ high, AI not yet dominant

Pre-schema frontier — neither agent is strong

Current frontier zone

The frontier sits in a high-σ region. Delegation is net-positive — the agent can evaluate and verify retrieved content. Expanding 𝒟* here frees parametric capacity for schema-intensive frontier work.

Figure 2. The delegation gradient $\mathcal{D}^*(d,t)$. Drag the AI capability slider to see the delegation frontier expand. The agent’s $\sigma$ profile (the step function) determines which regions are safe to delegate. Below $\sigma_\text{critical}$, adding retrieval capability does not improve compositional task performance — it adds retrieval latency and hallucination surface area without improving the agent’s ability to verify what is retrieved. The correct boundary for $\mathcal{D}^*$ is set at $\sigma_\text{critical}$, not at the AI performance crossover.

What Distinguishes This From Existing Adaptive Retrieval

Self-RAG conditions retrieval on a reflection token signalling factual uncertainty. FLARE triggers retrieval when next-token probability falls below a threshold. All such systems ask: is the agent uncertain? H-Bar asks: does the agent have sufficient σ to evaluate retrieved content?

These questions produce different routing decisions when confidence is low and σ is low simultaneously — the condition that arises in domains where the agent has not yet built causal structure. Under confidence-based routing, low confidence triggers retrieval. Under H-Bar’s criterion, low σ should suppress retrieval even when confidence is also low: the agent cannot verify what is retrieved, and the expected loss from accepting a causally inconsistent result exceeds the expected loss from the parametric answer. That is a concrete, testable divergence.

PIRL provides a bounded but instructive analogy. The Stribeck prior handles the analytically specified friction regime. The residual handles everything else. The delegation boundary sits where the prior’s causal assumptions hold. An architecture that detected when Stribeck assumptions were being violated and contracted the prior’s delegation scope in response would be a closer analogue to dynamic $\mathcal{D}^*$ — and is precisely the capability PIRL currently lacks on OOD dynamics.

---

The Ψ Mechanism: Why Schema Coherence Gates Intersection Discovery

When an agent has σ above threshold in two or more domains, and those domains share sufficiently high structural similarity $\phi(d_1, d_2)$, cross-domain interaction produces above-additive performance on compositional tasks. This is the intersection activation mechanism $\Psi$.

The key claim is that the benefit scales multiplicatively with σ in both contributing domains, not additively:

$\Psi(d_1, d_2, t) \propto \sigma_A(d_1, t) \times \sigma_A(d_2, t) \times \phi(d_1, d_2)$

The general form is:

$\Psi(d_1, d_2, t) = \begin{cases} g\!\left(\sigma_A(d_1, t),\; \sigma_A(d_2, t),\; \phi(d_1, d_2)\right) - C_\text{baseline} & \text{if } \sigma_A(d_1,t) \cdot \sigma_A(d_2,t) > \theta_I \\[4pt] 0 & \text{otherwise} \end{cases}$

where $C_\text{baseline}$ is the additive specialist baseline and $g$ is monotone increasing in each σ argument.

Why multiplicative and not additive? If cross-domain benefit required only that the agent be present in both domains, any juxtaposition would generate it. Experienced systems know that breadth without causal structure in both contributing domains is encyclopedic, not generative. An agent with high depth in $d_2$ and low σ in $d_1$ cannot bring the causal variables of $d_1$ to bear on a problem at their intersection — it can only surface the surface correlations. Multiplicative is the correct formalization: $\Psi$ collapses if either σ is low, even if the other is high.

This is the testable divergence from general multi-task learning: H-Bar predicts that halving σ in one contributing domain reduces $\Psi$ by approximately half; multi-task learning predicts incremental reduction. Figure 3 makes this structure directly explorable. Use the toggle to compare the multiplicative (H-Bar) and additive (baseline) predictions across the full σ space.

Figure 3 — Ψ Activation Heatmapφ(d₁, d₂) = 0.75 · θ_I = 0.18

Multiplicative (H-Bar)Additive (baseline)

Hover any cell to inspect Ψ values. The dashed line marks the θ_I boundary — below it, Ψ = 0 regardless of either agent's depth. The PIRL annotation marks the physics-prior × residual intersection (σ_prior ≈ 0.85, σ_residual ≈ 0.70). Toggle between multiplicative and additive to see the prediction difference: H-Bar claims the field measures the difference between these models by varying σ independently per contributing domain.

Figure 3. The $\Psi$ activation heatmap. Hover any cell to read the exact Ψ values under both models. The dashed line marks the $\theta_I$ boundary — below it, $\Psi = 0$ regardless of depth. The PIRL annotation (outlined cell) marks the physics-prior × residual intersection at σ_prior ≈ 0.85, σ_residual ≈ 0.70. Toggle to the additive baseline to see what general multi-task learning would predict at the same point — the discrepancy is the falsification target.

PIRL as a Worked Example

In PIRL, the Ψ mechanism has a direct architectural analogue. The PID feedback term, the Stribeck physics prior, and the residual NN form a designed three-component intersection. The prediction: PIRL’s performance advantage should be super-additive — the three-component system should outperform the sum of any two-component alternatives. If the advantage is merely additive (PID + Stribeck + NN ≈ PID + Stribeck + PID + NN), the multiplicative Ψ account is falsified for this architecture. If it is super-additive, this is consistent with Ψ being active.

This is the most directly executable single experiment the framework specifies: ablate PIRL’s three components systematically, measure whether performance advantages are additive or super-additive, and compare error distributions at the OOD boundary. It requires only existing PIRL infrastructure.

---

What This Means for Training

At Phase 1 → 2. Stop optimizing purely for distribution fit. Add training conditions that force causal-variable representation — varied conditions that share underlying causal structure but differ in surface statistics. Monitor the learning-rate proxy for the $\sigma_\text{critical}$ crossing. Do not introduce high-quality retrieval as a primary answer source at this stage: it suppresses σ development by bypassing the encoding pressure that would otherwise be required.

At Phase 2 → 3. Do not expand the agent’s breadth profile before σ is developed. AI-assisted breadth expansion in a low-σ agent creates the illusion of literacy without the schema to verify it. The delegation rule applies here: expand breadth through domains where σ is already above $\sigma_\text{critical}$. The σ proxy level at the time of cross-domain exposure is the single strongest predictor of whether cross-domain training will produce positive or negative transfer.

At Phase 3 and beyond. The training objective shifts from maximizing $\delta$ to maximizing $\Psi$. Cross-domain training with high-$\phi$ domain pairs becomes the primary investment. Delegation expands into the sub-$\sigma_\text{critical}$ frontier — but strictly within the schema boundary. An agent that expands $\mathcal{D}^*$ in high-σ domains frees parametric capacity for frontier work without incurring verification costs. The same expansion in low-σ domains produces short-run surface-accuracy gains and long-run systematic-split losses.

---

Conclusion

H-Bar started with a concrete engineering observation: PIRL adapts in a fraction of the epochs that unconstrained baselines require. That ratio has no explanation under the standard training paradigm, which tracks only the volume and quality of training data, not the organizational structure of what was learned from it. The framework treats that ratio as diagnostic evidence: two agents can reach the same benchmark score through representations organized around fundamentally different variables — and this difference determines everything that happens when the distribution shifts.

The field is building agents that score well on benchmarks that measure $\delta$ and poorly on benchmarks that measure whether $\delta$ has any causal structure behind it. H-Bar’s contribution is to specify what that missing structure is, how it behaves dynamically, what happens when it is absent at each phase of training, and what the correct routing decision is when it is not. The missing variable is not mysterious. It is measurable. The conditions that suppress it are specifiable. The interventions that develop it are concrete.

The H-Bar Model is a formal account of why two AI agents with identical performance scores on standard tasks can have radically different capabilities on novel problems — and what to do about it during training.

---

References

Footnotes

1. Kim, N., & Linzen, T. (2020). COGS: A compositional generalization challenge based on semantic interpretation. EMNLP. ↩

2. Lake, B., & Baroni, M. (2017). Generalisation without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. arXiv:1711.00350. Hupkes, D., et al. (2019). Compositionality decomposed: How do neural networks generalise? JAIR. ↩

3. Schölkopf, B., et al. (2021). Toward causal representation learning. Proceedings of the IEEE, 109(5), 612–634. ↩

4. Arjovsky, M., Bottou, L., Gulrajani, I., & Lopez-Paz, D. (2019). Invariant risk minimization. arXiv:1907.02893. ↩

5. Khetarpal, K., Riemer, M., Rish, I., & Precup, D. (2022). Towards continual reinforcement learning: A review and perspectives. JAIR, 75. ↩

6. Asai, A., Wu, Z., Wang, Y., Sil, A., & Hajishirzi, H. (2023). Self-RAG: Learning to retrieve, generate, and critique through self-reflection. arXiv:2310.11511. ↩

7. Jiang, Z., et al. (2023). FLARE: Active retrieval augmented generation. arXiv:2305.06983. ↩

8. Power, A., Burda, Y., Edwards, H., Babuschkin, I., & Misra, V. (2022). Grokking: Generalization beyond overfitting on small algorithmic datasets. arXiv:2201.02177. ↩

9. Bengio, Y., Louradour, J., Collobert, R., & Weston, J. (2009). Curriculum learning. Proceedings of ICML. ↩