1. Vendor and Product Reality

PathAI's commercial surface centers on AISight, an enterprise image-management system for digital pathology that ingests whole-slide images from scanners across Leica, Hamamatsu, Roche, and Philips, manages pathologist review workflows, and serves AI-derived overlays for tasks such as tumor quantification, biomarker scoring, and tissue characterization. Above AISight, PathAI develops a portfolio of AI models spanning oncology and metabolic-disease indications, including its AIM biomarker line such as AIM-PD-L1, used to score endpoints in biopharma trials. AISight Dx carries a CE-IVD mark for primary diagnosis in the EU, and PathAI publicly describes work advancing AI-based pathology tools toward regulated use; the specific US regulatory status of any given model should be checked against PathAI's own current disclosures rather than assumed. The platform's core value proposition rests on combining scanner-agnostic image management with model inference at clinical-laboratory scale, supported by a large-scale annotation operation drawing on board-certified pathologists across subspecialties.

The biopharma side of the business is substantial: PathAI runs structured engagements with pharmaceutical sponsors, including a publicly reported multi-year collaboration with Bristol Myers Squibb, to develop trial-specific models and score endpoints, and it operates clinical-laboratory diagnostic services through the CLIA-certified laboratory business it acquired (Poplar Healthcare) and now runs under its umbrella. That book of business creates a particular regulatory posture: models trained on data drawn from hospital partners, biopharma clients, contract research organizations, academic medical centers, and PathAI's own annotation operation, each of which contributes under different consent, licensing, IRB approval, and credential terms. Some contributions are de-identified retrospective archives; others are prospective trial cohorts under explicit sponsor data-use agreements; still others are routine clinical specimens annotated under broad institutional research consent.

The platform clearly understands data provenance as a commercial concern. Sponsor agreements specify scope of use; hospital data-use agreements specify retention and downstream sharing limits; annotation contracts specify intellectual-property assignment and quality requirements. The architectural question is whether provenance is enforced inside the training loop or only documented around it. The current answer, consistent with industry practice across digital pathology, is that provenance is documented in dataset manifests, model cards, and regulatory submissions, while the optimizer itself remains provenance-blind. That asymmetry between contractual sophistication and architectural primitive is precisely the surface where training-governance attaches.

2. Architectural Gap

Standard training pipelines, including those underlying digital-pathology platforms, treat the gradient as a uniform object. Every example contributes to every parameter at every depth, modulated only by loss and learning rate. Provenance is recorded in dataset manifests, audited in MLOps logs, and surfaced in model cards, but it does not reach into the optimizer. A slide contributed under a research-only consent, a slide annotated by a non-board-certified labeler, and a slide drawn from a trial whose sponsor has revoked use rights all push gradients into the same parameters with the same authority. The optimizer cannot distinguish them because the gradient computation does not carry the credential.

For an FDA-relevant product line this is structurally fragile. A regulator, a sponsor, or a hospital partner can ask which contributions shaped a given decision pathway, and the honest answer under standard training is that all contributions shaped all pathways in proportion to gradient magnitude and frequency. Retroactive remediation, retraining without a withdrawn cohort, costs months and money, requires re-validation of every downstream model artifact, and provides no architectural assurance that the next withdrawal will be cheaper. Each withdrawal event becomes a bespoke engineering effort rather than a routine policy operation. For a company whose product portfolio is expanding by additional regulated indications, this remediation cost compounds with every new model and every new sponsor relationship.

The structural fragility extends past withdrawal. Consider a 510(k) submission in which the sponsor must articulate which contributions shaped which decision components. Under uniform-gradient training, the only honest articulation is statistical: dataset-level descriptions of provenance, with no parameter-level attribution. FDA reviewers, sponsor quality organizations, and hospital data-use committees increasingly want a stronger answer, one that ties credential class to parameter class. Annotation by a board-certified pathologist with subspecialty fellowship training is materially different from annotation by a junior labeler under supervision, and yet both gradients enter the same parameters under standard training. The missing element is depth-selective gradient gating: a structural rule that a contribution's credential determines which depths of the network it is permitted to shape, enforced inside the optimizer rather than asserted by paperwork around it.

3. What Training-Governance Provides

The training-governance primitive treats every training example as a governed, admissibility-scoped observation. In the spec's terms each admitted example carries a governance record identifying the policy object that authorized its admission and the policy object that set its depth profile, plus a content-provenance record capturing source, acquisition pathway, and chain of custody. In the digital-pathology mapping, that admissibility scope naturally encodes the facts a lab already tracks as a credential: the contributing institution or annotator, consent scope (research-only, diagnostic-development, regulatory-submission), jurisdictional admission, annotator qualification (board certification, subspecialty, supervision class), and any constraints (time-bounded admission, sponsor-restricted use, indication-restricted use). The spec further discloses content anchoring, by which content derives a computable identity from its own structural entropy; content without a verified anchored identity is flagged provenance-incomplete and may be restricted by policy to shallow layers. The governance record is structured and inseparable from the example as it flows into the training loop.

The optimizer reads the credential and applies depth-selective gating: the gradient from that example flows into the depths the credential authorizes and is attenuated or blocked at depths it does not. Low-credential contributions, for example, broadly consented archival slides annotated under supervision, can shape shallow feature extraction (color normalization layers, tissue/non-tissue segmentation, generic morphology features) where the policy admits them. Only high-credential, board-certified, fully consented, jurisdictionally admitted contributions can shape decision-layer parameters that drive a regulated output such as a PD-L1 score or a NASH activity grade. The gating is structural; the optimizer cannot be coaxed by a misconfigured pipeline into letting an out-of-scope contribution touch a regulated decision layer.

Provenance tracing is the second half. The spec discloses an append-only, timestamped, sequentially numbered training-provenance log that records, for each admitted example, its depth-aggregation profile, the per-block gradient magnitude that actually reached each layer block, the governing policy object, and the content-provenance and admissibility-determination records. The log is tamper-resistant and can be periodically sealed with the cryptographic sealing infrastructure of the cross-referenced Governance filing. From it, a reverse query starts at an observed inference behavior and traces back to the training content that was structurally permitted to influence the responsible layer blocks. The spec is explicit that this is a bounded attribution set, not exact attribution: the non-linear dynamics of gradient-based optimization preclude pinning a behavior to a single contribution, but the log narrows the candidate set far below the full corpus and provides a definitive present-or-absent answer to whether specific content was used. Withdrawal follows from the same structure. Time-limited or rights-restricted content is trained under a suppressed depth profile that confines its influence to shallow layers, which the spec frames as an approximate-unlearning posture: content confined to shallow, generic layers can be excluded or re-trained with far less disruption than content deeply integrated across the stack, without the honest article claiming exact excision of a single cohort from a converged network.

4. Composition Pathway

The substrate composes onto PathAI's existing stack without disturbing AISight, the scanner integrations, or the pathologist review surface. Slide ingestion gains a credentialing step that records the contributing institution, the consent envelope, the IRB or equivalent approval, the annotator credential, and the sponsor-or-hospital data-use agreement under which the slide is admitted. The credential is bound to the slide as it moves into the training corpus and travels with each training example through every downstream operation. The training loop replaces its uniform gradient with the depth-selectively modulated gradient the spec describes, implemented through gated residual connections, attention-based depth selection, or an architecture-agnostic per-layer scaling factor applied on the backward pass, and parameterized by a depth-by-admissibility policy that is itself a versioned, signed artifact under PathAI's quality system. Because the modulation acts only on the backward pass, forward-pass inference behavior is unchanged. The model registry stores the append-only training-provenance log alongside weights, and AISight and downstream diagnostic products can run reverse queries against that log at audit or inference time to attach a bounded provenance citation to an output.

Training-inference integration matters here because pathology models do not stop training at deployment. PathAI continues to refine models as new annotated slides arrive, as sponsors contribute trial cohorts, and as biomarker definitions evolve. Depth-selective routing extends through that continuous-learning surface: a sponsor's rights-restricted cohort admitted under a shallow-confined depth profile can be excluded or approximately unlearned with far less disruption than a deeply integrated cohort, and a newly admitted high-admissibility cohort can be routed to deep layers under an updated policy while the shallow feature extractor is protected from disruption by the depth profile. The continuous-learning operation that today requires careful manual curation becomes a routine governed-substrate operation. This is what makes the substrate an architectural element rather than a one-time training trick.

Composition extends to PathAI's external surfaces: scanner-vendor partnerships, biopharma sponsor portals, and the hospital network for clinical deployment. Each of those surfaces becomes an authority taxonomy participant. Scanner vendors issue device-credentialed observations attesting to image-acquisition provenance; sponsors issue cohort-credentialed observations carrying their data-use scope; hospitals issue institution-credentialed observations carrying their consent class. The training loop becomes the convergence point of those credentials, and the substrate gives PathAI the structural answer to "whose data shaped what" that none of the credentials individually could provide.

5. Commercial and Licensing Implication

PathAI's competitive position rests on regulatory credibility with FDA, EMA, PMDA, and biopharma quality organizations. Every additional product line raises the stakes of provenance, because every additional indication brings additional sponsors, additional consent envelopes, and additional withdrawal scenarios. A platform that can demonstrate, at submission time, that deep decision-layer parameters were routed only from high-admissibility, consented, appropriately credentialed contributions, with an append-only training-provenance log to evidence it, has a materially different conversation with regulators than one whose provenance lives only in dataset documentation. This is not a claim about any weakness specific to PathAI; scanner-agnostic image management, subspecialty annotation at scale, and validated biomarker models are genuine strengths, and the digital-pathology field as a whole, including well-funded peers, keeps provenance around the training loop rather than inside the optimizer. The comparison is scoped to exactly one architectural axis: whether admissibility routes the gradient by depth, which no digital-pathology platform of this class exposes as a training-time primitive.

The biopharma partnership book benefits in the same way. Sponsors contributing trial data want assurance that their data is used within the contracted scope and can be withdrawn meaningfully. Depth-selective routing with admissibility-scoped provenance gives PathAI a defensible contractual posture: your contribution is routed only to the depths your agreement authorizes, and rights-restricted contributions confined to shallow layers can be excluded or approximately unlearned far more cheaply than deeply integrated content, with the training-provenance log evidencing both. That posture matters during sponsor diligence; it matters again at trial close-out; and it matters most when a sponsor pivots, a trial terminates, or a regulatory authority requests targeted retraction.

The fitting arrangement is an embedded substrate license: PathAI embeds the AQ training-governance primitive into AISight and its model-development pipeline and sub-licenses substrate participation to its biopharma and hospital customers as part of the platform engagement. Pricing aligns with how regulated customers actually consume governance, per-cohort or per-submission rather than per-seat. The license covers admissibility-scoped contribution provenance, depth-selective gradient routing, the append-only training-provenance log with cryptographic sealing, and the memorization-detection and reverse-query machinery that carries bounded provenance through to served inferences. PathAI gains the FDA-aligned architectural element its product line implies, with admissibility enforced inside the optimizer rather than documented around it. Honest framing: the primitive does not replace digital pathology; it gives digital-pathology AI the training-time substrate its regulatory posture has always implied and never had.

6. How a Skilled Implementer Would Build It

The approach is enabling and can be reduced to practice by a skilled machine-learning engineer using standard tooling. At ingestion, each training example is bound to a governance record (authorizing policy object, depth-profile policy object) and a content-provenance record (source, acquisition pathway, chain of custody, semantic metadata), optionally enriched with a content-anchoring identity derived from the example's own structural entropy. A depth-profile router maps each example's governance and entropy classification to a per-block contribution-weight vector. During the backward pass, that vector modulates the gradient by one of three disclosed and interchangeable techniques: gated residual connections (a per-layer gating coefficient on the residual shortcut), attention-based depth selection (scaling the gradient reaching attention weights and value projections per layer), or an architecture-agnostic layer-specific scaling factor (a scalar multiplier applied at each layer boundary before accumulation). Modulation acts only on the backward pass, so forward-pass inference is unchanged, and the mechanism composes with SGD, Adam, AdamW, and their variants without altering the update rule. Profiles operate at block-level granularity for tractability on deep networks. Embodiments and variations expressly contemplated include: block groupings from a single transformer layer to a group of residual layers or a single attention head; curriculum-integrated depth scheduling that sequences examples by entropy band and policy scope across training phases; the append-only, sequentially numbered, timestamped training-provenance log, periodically sealed for third-party verification; forward and reverse provenance queries yielding a bounded (not exact) attribution set; training-level memorization detection that classifies inference-output similarity as shallow, deep, or absent memorization and governs the served inference accordingly; differential privacy by architectural confinement, in which privacy-sensitive content is confined to shallow, less-memorizable layers rather than protected by injected noise; and a governed fine-tuning provenance record structurally distinct from pre-training provenance. Applicability spans convolutional, recurrent, transformer, mixture-of-experts, and hybrid architectures, and the digital-pathology mapping above is one embodiment among many.

7. Disclosure Scope

The inventive subject matter described here, depth-selective gradient routing bound to per-contribution admissibility, the append-only and cryptographically sealable training-provenance log, forward and reverse provenance queries, training-level memorization detection, differential privacy by depth-confinement, and governed fine-tuning provenance, is disclosed in United States Patent Application 19/647,395. This article is a dated public disclosure of that subject matter and is intended to be enabling to a person of ordinary skill in the art. All statements about the mechanisms, guarantees, and behaviors of the invention trace to that application; where the application characterizes an outcome as approximate (for example, that reverse queries produce a bounded attribution set rather than exact attribution, and that shallow-confinement supports approximate rather than exact unlearning), this article adopts that same characterization and claims nothing stronger.

References to PathAI, AISight, the AIM biomarker line, Bristol Myers Squibb, Poplar Healthcare, named scanner vendors, and other third parties, and any characterization of the digital-pathology market, FDA/CE regulatory pathways, or competitive landscape, are external context drawn from public sources for comparison only. They are not claims of United States Patent Application 19/647,395, are not affiliated with or endorsed by those companies, and each named product's current feature set and regulatory status should be verified against that company's own disclosures. No trademark of any third party is claimed.