CORD — Contextual Operator Response Dynamics is a computational-biology project for learning and testing how biological interventions transform cellular state as a function of context, time, and measurement regime.
The project succeeds the earlier Constraint Geometry Theory (CGT) research program. CGT generated useful empirical findings and analysis machinery, but its strongest data support a narrower conclusion than a universal constraint theory: perturbational response structure is recurrent and functionally informative, yet materially context-conditioned. CORD makes that qualification part of the model rather than treating it as an exception.
Alpha milestone 0.1.0a4 provides the rebuilt CORD-002 software candidate: a metadata-first,
plan-locked, package-owned pipeline with a twelve-model held-out-study tournament, negative controls,
synthetic recovery systems, three deterministic canvases, and immutable bundle closure. The shipped
synthetic execution validates software behavior only; it is not a biological result. Empirical runs
may validly close at context_not_identifiable before expression access. CORD-001 remains an
inconclusive historical fixed-representation feasibility experiment and is not CORD validation.
The final CGT consolidation run, CGT_CORE_001 (2026-07-15), indexes 838 artifacts from 31 historical run directories and freezes 81 claims. Its strongest inherited observations are:
| Finding | Frozen evidence | Interpretation boundary |
|---|---|---|
| Held-out-gene fitness prediction | pooled Spearman rho = 0.613; pooled R2 = 0.353; mean fold-wise regression rho = 0.622; mean fold-wise ROC AUC = 0.919; endpoint-shuffle empirical p = 0.0099 | Retrospective gene-level prediction; not causal mechanism or intervention prediction |
| Recurrent response structure | 248 modes, 43 datasets, 9 contexts, 16 families; about 53% of raw family-coordinate variance retained after context residualization; same-label residual-transfer cosine about 0.56 versus controls near zero | Transfer attenuates under same-study-proxy exclusion and leave-context-out tests |
| Biological coherence | Three headline and six supporting signed axes had curated external pathway support | Labels are candidate interpretations and include manually curated, partially redundant axes |
| TCGA projection | 15 axes projected into approximately 9,300 primary tumors; mean absolute axis correlation fell from 0.753 raw to 0.265 after leave-one-axis-out global residualization | Bulk lineage, purity, RNA content, and microenvironment confound interpretation; supporting evidence only |
These results motivate CORD, but do not establish universal causal constraint laws. See the independent freeze interpretation.
Two compact CGT-derived layers coexist in the repository and must not be conflated:
| Layer | Scope | CORD disposition |
|---|---|---|
| Frozen CGT signed-coordinate bridge | 2,720 response rows from 39 datasets in one historical 16-coordinate representation | Preserved as a derived, transductive sensitivity/software bridge; it is not the authoritative CORD-001 cohort |
Audited CGT_CACHE_002_v1 response geometries |
54 registered datasets, 49 downloaded and attempted, 37 current schema-valid caches | Authoritative CORD-001 input cohort; every dataset remains in its own PCA basis and is compared only with basis-invariant geometry or train-only alignment |
The 37 valid caches contain 2,534 response rows representing 2,088 exact target labels. They are an inherited selected/capped perturbation subset, not every intervention in the source studies; shared normalized labels do not establish matched modality, dose, or time. Study-family and context annotations remain provisional and explicitly require human review. Large cache arrays stay outside Git; the repository stores the cohort registry, validation logic, input identities, and run records.
CORD treats a perturbation as a context-indexed operator acting on a measured or latent cellular state:
[ \Delta y_{\tau} = \mathcal{O}_{u,c,\tau}(y_0) + \varepsilon, ]
where (u) is an intervention, (c) is biological and technical context, (\tau) is elapsed time, and (y) is an observation linked explicitly to—but not conflated with—a latent state. Viability and other phenotypes are downstream endpoints to predict or control, not assumed definitions of the representation.
Initial falsifiable targets are cross-study operator transfer, context-conditioned response prediction, uncertainty calibration, and external endpoint prediction under grouped holdouts.
CORD-001 addresses only a deliberately narrower precursor: whether shared-target response geometry and train-only cross-dataset maps show enough signal to justify a measurement-complete operator study. It does not observe baseline state, dose, or time and therefore cannot estimate the full (\mathcal{O}_{u,c,\tau}(y_0)). Its design was locked after development inspection of the four eligible cache pairs, not registered before data inspection.
Phase 1 is limited to exactly three scientific notebooks:
CORD_001_contextual_operator_transfer.ipynb— audited fixed-representation geometry and held-out-label transfer feasibility.CORD_002_integrated_operator_validation.ipynb— raw-cell measurement of baseline, context, intervention, and time, followed by grouped external predictive validation and calibration.CORD_003_confirmatory_synthesis.ipynb— confirmatory synthesis, robustness, claims, and final Phase-1 figures.
Reusable acquisition, validation, modeling, provenance, and figure code belongs in src/cord, not
in additional cumulative notebooks. The historical CGT recovery notebook is a provenance utility
and is not counted as a Phase-1 CORD scientific notebook.
archive/cgt/ Immutable CGT lineage, freeze artifacts, and migration ledger
data/ Dataset registry conventions; raw public data are never committed
docs/ Architecture, scientific status, governance, and decisions
notebooks/ Curated entrypoints only; exploratory history remains in the CGT archive
provenance/ Run-record templates and provenance documentation
schemas/ Machine-readable contracts for runs and claims
src/cord/ Reusable CORD package code
tests/ Unit and evidence-freeze validation tests
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e '.[dev]'
pytest
cord-provenance validate archive/cgt/releases/core-001Every evidentiary CORD run must link:
immutable input identities -> code commit + configuration -> run record
-> hashed artifacts -> typed claims -> figures/manuscript statements
An output directory or notebook execution is not, by itself, provenance. Dataset versions, content digests, code identity, environment, random seeds, exclusions, null models, and validation partitions must be recorded. Full requirements are in docs/reproducibility.md and AGENTS.md.
The canonical CORD-001 notebook installs an exact scientific stack from
requirements/cord001-colab.txt, fixes numerical thread counts, and records the full installed
distribution inventory, NumPy/BLAS configuration, and active thread pools in its run bundle.
Large public datasets and caches are not stored in Git. The repository will contain acquisition metadata, checksums, small source-data tables, and reproducible materialization instructions. Larger immutable releases should be deposited in an archival object store or DOI-bearing repository and referenced by content digest.
No open-source license has yet been selected. Until one is added, the repository does not grant reuse rights. This is an explicit owner decision rather than an accidental omission.