LAB / BUILDING MY DIGITAL SELF
My long-term goal is to build my own digital self, starting with peptide research and gradually connecting knowledge, reasoning and action. These are records of real work, reusable capabilities, validation and open questions.
LAB / BUILD LOG
Research capabilities and knowledge, built day by day.
Research capabilities · methods & execution
30 days · CommitsYear-to-date dates scroll horizontally; the selected date is shown below.
Knowledge & experience · sources & memory
30 days · CommitsYear-to-date dates scroll horizontally; the selected date is shown below.
agentic_ai_platform94commits
knowledge_wiki0commits
Updated: · UTC · — unavailable
A TREE IN THE MAKING
Each maple leaf is a method. New capabilities take root through evidenced work.
Rooted in real work.
Follow a leaf. See where it leads.
63 methods · counted once across stages
Leaves name methods, not experimentally validated results. Hit validation, hit-to-lead and lead optimization have distinct boundaries; experimental feedback can return to any round.
target or complex structure and explicit chain mapping → contact, SASA and epitope hotspot assessments.
Biological advice, contact hotspots, geometric feasibility and trajectory evidence remain distinct; no causal target validation.
Research role · Composed tasks
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
model-specific DNA arrays or variants, track identities and selected assets → model-bound track predictions or variant effects.
Enformer, Borzoi and Flashzoi are distinct native operations; genomic signal is not protein binding or causal validation.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
registered reference FASTA, chromosome sizes, peaks and selected assets → native profile and track predictions.
Reference and peak identity must match; predicted accessibility is not a causal or protein-binding result.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
DNA sequences, layer or generation controls and explicit assets → DNA generations, likelihoods or layer embeddings.
This worker does not score proteins; generation and embeddings are not phenotype or binding measurements.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
registered primary hg38 reference, variants and selected model assets → native gpn star vep variant effects.
The caller retains only the declared primary hg38 surface; other models or embeddings are not qualified here.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
registered sequence or chemical search space and query → matched records with source identity.
Retrieval matches are not confirmed binding hits.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
pre/post count tables, selection paths and controls → enrichment tables, QC and candidate evidence.
Enrichment needs background and orthogonal binding validation.
Research role · Composed tasks
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
Registry-resolved compiled case and hash-bound checker contract, explicit monomer/binder/motif task, count, seed and batch size → Durable asynchronous compute JobHandle, native operation manifest, generated complex CIFs and candidate identity mappings.
Canonical L-proteins only; native qualification pending; cyclic and noncanonical chemistry refused.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
target structures and explicit generation constraints → generated structures, sequences and provenance.
Software source and controls do not establish native cyclic support or experimental affinity.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
target structure and explicit design and topology contract → designed sequences, structures and attempt provenance.
Source and extended request support do not qualify the native recipe; no silent v1 fallback.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
motif, scaffold and cyclic topology constraints → candidate scaffold structures and constraint evidence.
Closure geometry is not synthesis or binding proof.
Research role · Composed tasks
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
RFD3 specifications or RF3 named molecular records and selected assets → generated RFD3 designs or RF3 predicted structures with sample provenance.
RFD3 generation and RoseTTAFold3 RF3 prediction are distinct operations; cyclic chemistry requires specific admission.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
target and sequence conditions with explicit cyclic flow recipe → generated structures and source-bound evidence.
The software candidate remains natively unqualified; cyclic generation is not a binding measurement.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
target structure, sequence logits and gradient objective → designed sequences and predicted structures.
Design feedback is not independent refolding or experimental affinity; topology and restart conditions remain explicit.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
target PDB, chain roles, binder length and explicit gradient schedule → designed sequence and final predicted structure.
This is a bounded gradient operation, not the BindCraft campaign driver; confidence is not experimental affinity.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
target or motif coordinates, contigs and explicit diffusion controls → generated protein backbones and sample provenance.
Backbone generation needs downstream sequence design and independent evaluation; native constraint support is conditional.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
selected source config, conditioning inputs, seeds and model assets → native generated candidates and source-bound outputs.
Native search, sampler and conditioning stay upstream; software readiness is not native or scientific qualification.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
target and sequence design inputs, seed and explicit recipe → designed sequences, structures and completion evidence.
Scientific qualification pending; completed-unit reuse is not mid-job recovery.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
native design input, selected assets and dtype controls → generated structures and native output provenance.
The selected design dependency is required; software transfer does not establish affinity or native qualification.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
explicit receptor and peptide chains, topology, cycles and predictor settings → authenticated per-cycle sequence, prediction and selection evidence.
MPNN packing plus AF3 or ESMFold2 forward feedback supplies no sequence gradient; parent workflow owns authorization.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
validated design inputs and explicit cyclic overlay conditions → design invocation and environment contract.
Research stage; production generator not admitted; licence and native-output bindings remain unresolved.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
target protein sequence, peptide length, count and explicit assets → generated peptide sequences and selected-source provenance.
Selected-source licence and native qualification remain unresolved; generated sequences are not confirmed binders.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
backbone, fixed positions and sampling protocol → sampled sequences and optional relaxed structures.
Backbone-conditioned sampling alone is not a complete binder generator.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
sequence and model-specific conditioning → scores with model and residue-index provenance.
Likelihood and embeddings do not measure affinity.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
protein backbones, chain roles, fixed positions and sampling protocol → sampled or scored sequences with provenance.
Backbone-conditioned sampling alone is not a complete binder generator; scores are not affinity.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
protein sequences, selected ESMC operation and explicit assets → sequence representations or operation-specific scores.
Offline representation does not submit work or alter resident 6B admission; likelihood and embeddings are not affinity.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
optional sequence prefix, model identity and sample protocol → generated sequences or bidirectional averaged likelihood.
Unconditional or prefix generation is not target-conditioned binder design; likelihood is not binding energy.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
backbone coordinates, chain and sequence conditions → backbone-conditioned sequences or likelihood.
A supplied backbone is required; likelihood is not affinity. Native evidence is specific to the source, checkpoint, operation, device and fixture; it does not establish predictive accuracy.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
registered structures, selected model assets and design request → structure-conditioned protein sequences.
Structure-conditioned design does not establish binder activity or native qualification.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
protein sequence operations and host-bound selected assets → operation-specific scores, embeddings or predictions.
Built with Profluent-E1; software transport does not qualify accuracy or experimental affinity.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
candidates, molecular identity and supplied evidence → normalized evidence and explicit qualification gaps.
Computational confidence cannot establish experimental binding.
Research role · Composed tasks
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
molecular sequences or seed structures and explicit predictor or sampler mode → predicted poses, consistency or sampling evidence.
Consistency mode does not sample poses; confidence is not affinity.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
AF3 JSON entities and explicit seeds and sample census → native structure output census and protocol identity.
This is an input/output protocol boundary; native chemistry and model qualification remain separate.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
selected Protenix request entities, templates, seeds and assets → validated native request and complete structural output set.
Template hits and chemical atom validity require native featurisation evidence.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
selected OpenBind entities, assets and explicit mode recipe → bound native request and complete output census.
OpenBind is distinct from preview2; chemistry, numerical mode and weights need qualification.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
protein sequences, selected assets and folding protocol → validated folding payload and model identity.
Forward folding is distinct from gradient design; confidence is not affinity.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
molecular entities, bonds and explicit sampling protocol → predicted structures with atom and sampling provenance.
Cyclic software overlays do not establish native chemistry or binding accuracy.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
sequence, selected features, optional initial coordinates and folding protocol → AF2 forward structural prediction and provenance.
Coordinate-seeded forward prediction is distinct from ColabDesign gradient design; qualification is separate.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
explicit folding task, inputs, seeds, samples and bound assets → complete native predicted output set.
Transport and output accounting do not qualify scientific accuracy, numerical equivalence or recovery.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
molecular entities, explicit seeds, samples and selected assets → native structures with complete entity and sample accounting.
Source and transport checks do not establish chemistry accuracy, affinity or checkpoint recovery.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
molecular input, seeds, samples and bound selected assets → native structures and complete unit accounting.
Migrated bf16 fast_layernorm differs from upstream fp32; scientific qualification and native long-job recovery remain separate.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
Declared structures, sites and restraints → pose cross-checks with ADCP, HADDOCK3 or LightDock.
Docking provides pose evidence, not affinity ranking.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
molecular system, force field and declared physical protocol → protocol-bound energy estimates and diagnostics.
ddG_fold is not ddG_bind; distinct protocols cannot be fused implicitly.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
Explicit topology, temperature ladder and protocol → replica-exchange simulation and analysis.
Job recovery and trajectory evidence require separate acceptance; submission is not a scientific conclusion.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
Explicit topology, structures and parameters → relaxed poses and protocol-bound interface quantities.
Crosslinks and noncanonical residues require retained parameters and closure identity; refinement scores are not affinity.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
registered assay evidence, goals and host-authored design cards → durable projects, rounds and design panels.
Project tools retain evidence and prepare method handoffs; they do not reason, predict, launch computation or order synthesis.
Research role · Composed tasks
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
parent complex, explicit operations and evaluation contract → candidate pools, provenance and evaluation evidence.
Parent optimization is distinct from de novo generation.
Research role · Composed tasks
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
SMILES, registered support data and fixed model recipe → endpoint predictions, refusal states and provenance.
Endpoint recipes have explicit provenance; peptide and noncanonical transfer are unvalidated; deployment is checked separately.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
assay observations with source identity, units and censoring → SAR relationships with measured and inferred evidence separated.
Cross-assay values are not interchangeable; suggestions remain unmeasured.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
Explicit sites, context and scorers → a complete mutation matrix, coverage and rankings.
Binding and folding quantities remain distinct; fusion is explicit and native qualification is method-specific.
Research role · Composed tasks
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
Measured SAR, structure evidence and hypotheses → traceable proposals and molecular identities.
Proposals remain unevaluated; structural explanations need independent grounding.
Research role · Composed tasks
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
complex geometry and candidate bridge chemistry → ranked bridge pairs and feasibility observations.
Site choice is distinct from bond-closure validation, synthesis feasibility and binding evidence.
Research role · Scientific operations
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
verified claims, figures and evidence references → reports with claim-to-figure provenance.
Rendered output cannot close unfinished upstream scientific gates.
Research role · Execution & learning support
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
source-bound hypothesis and finding cards → durable scientific context.
A recorded hypothesis is not a verified conclusion.
Research role · Execution & learning support
The knowledge base retains sourced experience and methods for the next investigation.
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
authorized contract, task state and evidence → durable execution and acceptance records.
Workflow completion does not independently qualify scientific outcomes.
Research role · Execution & learning support
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
owner-bound workload, measured capacity and budget → execution descriptors, results and recovery state.
Process exit success does not establish scientific validity.
Research role · Execution & learning support
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
Registered logical model name → resolved model and weight identity.
Registration does not establish installed runtime or inference success.
Research role · Shared contracts
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
Authorized service discovery → declared service routes and runtime evidence.
A reachable endpoint is not a qualified scientific prediction.
Research role · Execution & learning support
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
question and maintained knowledge sources → ranked evidence references.
Retrieval, publication, adoption and measured effect are distinct states.
Research role · Execution & learning support
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
logical dataset name and access context → registered data identity and access results.
Registration alone is not data QC or permission to access restricted data.
Research role · Shared contracts
Registered data, benchmarks and provenance support comparable, reproducible research.
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
molecular graph, chains, residues and covalent declarations → identity mappings and validated topology records.
Graph validity does not prove chemical synthesis feasibility.
Research role · Shared contracts
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
explicit round identity, parent lineage, content-bound evidence and root control references → validated persistent round records and explicit historical migration exports.
Scientific rounds do not create root iterations or backtracks; records grant no compute or scientific authority.
Research role · Execution & learning support
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
scalar values and explicit frozen or training-pool extrema → normalized or inverse-transformed protocol values.
Normalization is not a scientific model; extrema and affine-v2 clipping protocol must be explicit and historical v1 is read-only.
Research role · Shared contracts
The source declaration or implementation was inspected; registration does not establish a native run.
Source checked ·
No public execution evidence attached.
That leaf is still out of sight. Try a broader search.
Rooted in sourced knowledge · traceable data · Methods & work records ↗Open the original record: Public method notebook ↗Open the original record: Method notebook ↗Open the original record: Work and validation records ↗Open the original record: Full work records ↗Open the original record: Retained historical snapshot ↗Open the original record: Protein Language Model ↗Open the original record: Protein Language Model ↗Open the original record: Structure prediction ↗Open the original record: Structure prediction ↗Open the original record: Biomolecular docking ↗Open the original record: Biomolecular docking ↗Open the original record: Inverse folding ↗Open the original record: Inverse folding ↗Open the original record: Generative backbone design ↗Open the original record: Generative backbone design ↗Open the original record: Partial diffusion ↗Open the original record: Partial diffusion ↗Open the original record: Interface redesign ↗Open the original record: Interface redesign ↗Open the original record: Rational design ↗Open the original record: Rational design ↗Open the original record: Computational mutation scanning ↗Open the original record: Computational mutation scanning ↗Open the original record: Molecular dynamics ↗Open the original record: Molecular dynamics ↗Open the original record: Experimental mutation scanning ↗Open the original record: Experimental mutation scanning ↗Open the original record: mRNA Display Analysis ↗Open the original record: mRNA Display Analysis ↗Open the original record: CI worker and workflow maintenance ↗Open the original record: Declare a regression-check baseline dependency ↗Open the original record: Version benchmark definitions and views ↗Open the original record: Separate batch and resident compute interfaces ↗Open the original record: Plan evaluation-record views and interfaces ↗Open the original record: Plan layered evaluation coverage ↗Open the original record: Integrate a design-method adapter ↗Open the original record: Plan a bounded data-catalog feasibility check ↗Open the original record: Review monitoring storage and recovery ↗Open the original record: Make a planning reference portable ↗Open the original record: Plan durable request evidence and recovery ↗Open the original record: Plan recovery for expired service resources ↗Open the original record: Plan controlled continuation after publication ↗Open the original record: Plan evidence checks for predictive-method research ↗Open the original record: Track workflow dependency provenance ↗Open the original record: Clarify check-planning boundaries ↗Open the original record: Plan method qualification across repeated runs ↗Open the original record: Plan recovery of missing method assets ↗Open the original record: Plan recovery after an interrupted installation ↗Open the original record: Daily evidence retention ↗Open the original record: Selfcheck evidence checks ↗Open the original record: Controlled interrupted-request recovery ↗Open the original record: Disk checks and workspace housekeeping ↗Open the original record: Command launcher maintenance ↗Open the original record: Disk reports and tool documentation ↗Open the original record: Knowledge note synchronization ↗Open the original record: Review record continuity ↗Open the original record: Preserve original review and request records ↗Open the original record: Note metadata compatibility ↗Open the original record: Integration regression repair proposal ↗Open the original record: Build and recovery checks ↗Open the original record: Delivery workflow integration ↗Open the original record: Interactive registry shape validation ↗Open the original record: Identity checks before message sending ↗Open the original record: Abort records for unsubmitted tasks ↗Open the original record: Reminder reading and calendar coverage ↗Open the original record: Metadata pagination and coverage ↗Open the original record: Capability-linked engineering checks ↗Open the original record: Offline message intake ↗Open the original record: Tool guidance and wiring maintenance ↗Open the original record: Optimization method routing ↗Open the original record: Lead optimization integration ↗Open the original record: Research workflow budget history ↗Open the original record: Capability contracts and validation mechanisms ↗Open the original record: Improve code reuse checks ↗Open the original record: Repair validation entry resolution ↗Open the original record: Calendar operations and readback ↗Open the original record: Agent cancellation and receipts ↗Open the original record: Recovery after interrupted installation ↗Open the original record: Property prediction method research ↗Open the original record: Test collection and replay qualification ↗Open the original record: Keep engineering checks relevant ↗Open the original record: Reduce false expiry during waits ↗Open the original record: Hand completed results over reliably ↗Open the original record: A real launch before readiness ↗Open the original record: A checklist with evidence ↗Open the original record: Diagnostics from field shape ↗Open the original record: Clarifying research tool responsibilities ↗Open the original record: Making overnight work more reliable ↗Open the original record: Defining data collection before use ↗Open the original record: Preserving validation history during migration ↗Open the original record: Making request recovery traceable ↗Open the original record: Keep two calendar boundaries explicit ↗Open the original record: Separate comparison evidence from presentation ↗
New leaves follow reviewed, evidenced capabilities. Source changes, checks and publication are recorded separately and reflected in daily snapshots. The breeze is decorative.
Snapshot needs an update; the last checked content is retained.
GROWING / QUESTIONS INTO CAPABILITIES
63 sourced methods and supporting capabilities are mapped. New qualifications, behavior changes and limitations appear here after review.
Explore the maple capability tree ↗Capability sources checked · 2026-10-27 · Snapshot needs an update
RECENT WORK / 2026-10-28
Records through 2026-10-28.
Selected records · through 2026-10-28
Records through 2026-10-28 · Selected records
Records through 2026-10-28 · Selected records
COGNITIVE SYSTEM / METHODS & HARNESS EVOLUTION
Experience, distillation, diagnosis and evaluation produce improvements that can be tested. This loop is still being built.
Retain outputs, evidence, failures and applicability boundaries.
Convert reusable experience into sourced knowledge and methods.
Compare task needs with method and capability gaps.
Record capabilities and limits to support future task selection.
Propose versioned changes to skills, tools or workflows.
Independently compare candidates with a baseline; evaluate the evaluator too.
CAPABILITIES / PUBLIC EVIDENCE
Up to three recent public records: what changed, its validation and the next step. Development activity and capability gains are kept distinct.