data/DATA_DICTIONARY.md

Data dictionary and provenance

All numerical records are synthetic. No experimental measurement or trained model is included. Scientific source version: 1.0.0. Distribution and export version: 1.0.1. Each JSONL line is one UTF-8 JSON object; schemas are in ../schemas/. export_manifest.json records counts and hashes. The common Hub split name train is an access convention, not a training/test claim.

Configurations

Configuration	Rows	Source	Meaning
kinetic_runs	775	results/baseline.csv (30), sweep.csv (720), fusion_sweep.csv (25)	Outcomes of six protocols under an uncalibrated CTMC and post-capture bookkeeping
threshold_recurrence	140	results/threshold.csv	Chosen recurrence values with clipping to [0,1]; analytical illustration
interface_codebook	128	results/example_codebook.csv	Logical words, alphabet size 4, length 8, minimum Hamming distance 3
claims	12	CLAIMS.json	Claim IDs, classifications, assumptions, evidence and exclusions
research_sections	45	Scientific Markdown sources	Exact source-text chunks with source-line ranges and content hashes
references	32	docs/REFERENCES.md	Bibliographic records and source-access limitations

Total: 1,132 records across heterogeneous configurations. Counts are not the number of experiments, independent measurements or distinct scientific findings. Baseline and parameter/fusion sweeps intentionally repeat some parameter settings. Do not pool them as independent replicates or randomly split them into a predictive benchmark without grouping identical settings.

Kinetic model fields

Parameters are nominal inputs to the study. results/run_manifest.json supplies omitted fixed parameters; code/simulate.py is the exact implementation. The random protocol deliberately uses an effective energy gap of zero even when the requested nominal gap_kbt field is nonzero. Other protocols use the nominal gap.

Field	Type / unit	Interpretation
record_id	string	Stable within this version; source study plus one-based data-row index
study	enum	baseline, parameter_sweep, or fusion_floor
evidence_type	string	uncalibrated_local_socket_CTMC for every row
mode	enum	random, coded, hierarchical, proofreading, proofreading_hierarchy, locking
N	integer, components	Nominal tree-object component count; hierarchical cases use an exact power of b=4
gap_kbt	number, kBT	Nominal total incorrect-versus-correct detachment free-energy gap, not per base
progress_rate	number, s^-1	Forward checking/capture rate; irreversible approximation to driven progression
dilution_exponent	number	alpha in c(level)=c0b^[-alpha(level−1)]
fusion_error	number, probability	Added error per retained correct join in locking; ignored by other modes
deadline_s	number, s	Total allocated stage deadline, including locking delays
levels	integer	Number of assembly stages used in this model; not encoding depth in F1
correct_fraction	number, fraction of N−1 joins	Join-count-weighted probability of a retained correct join
wrong_fraction	number, fraction of N−1 joins	Join-count-weighted probability of a retained wrong join
missing_fraction	number, fraction of N−1 joins	Join-count-weighted uncaptured/lost probability
wrong_among_captured	number, conditional fraction	Wrong/(correct+wrong) after bookkeeping; excludes missing joins
log10_perfect_yield	number, log10 probability	Logarithm of a product of independent correct-join probabilities, conditional on ideal inputs
perfect_yield	number, probability	That conditional product; may be exactly zero due to floating-point underflow
sum_mean_local_capture_s	number, s	Sum of single-socket mean capture times, plus locking delays; not time to finish all joins
expected_progress_events_complete	number, events	Eventual expected progress events summed over joins, including rejected attempts; not Joules or actual finite-deadline energy
material_families	integer	Fixed bookkeeping value 6, not optimized library cardinality
active_decorated_variants	integer	Maximum simultaneous logical variants in the model: N for flat cases or b for hierarchical cases
recognition_symbols	integer	Assumed illustrative elementary alphabet count 4; not an experimentally qualified count
code_slots	integer	Illustrative address-format count 16; distinct from the separate eight-slot example codebook
template_control_bits	integer, bits	Chosen illustrative format: 64+32levels for hierarchy, 64+ceil(log2(N))(N−1) for flat cases; not an information-theoretic minimum
source_path	string	Repository-relative raw CSV path
source_row	integer	One-based physical CSV line; kinetic/threshold line 1 is the header
source_sha256	64 hex digits	SHA-256 of the entire source file, not a signature
release_version	string	Distribution version of this normalized record

correct_fraction + wrong_fraction + missing_fraction is one up to numerical tolerance. The model omits 3-D cluster collisions, finite inventories, crowding, child-module provisioning failures and realistic correlated conversion physics. Suppressing one local wrong-attachment channel is not equivalent to correcting all physical defects. A conditional perfect product is neither atomic perfection nor a calibrated complete-device yield.

Analytical threshold records

Field	Unit	Meaning
eta	probability per recursion step	Added residual error floor
p0	probability	Starting scalar recurrence value
level	integer	Recursion index, separate from assembly stages
p	probability	Iterated value under p(next)=min(1,C*p^q+eta)
C	dimensionless	Chosen malignant-set coefficient 28
q	integer	Chosen exponent 2; called r in the theorem statement

Other fields supply stable ID, evidence label and source provenance. These evaluations do not estimate ultralow probabilities from samples and do not prove the existence of a physical threshold gadget. At values where an upper bound exceeds one, clipping supplies only the trivial probability upper bound.

Interface-code records

symbols is an array of eight integers in {0,1,2,3}. alphabet_size=4, word_length=8, minimum_distance=3 are verified properties of the included logical set. record_id and source provenance identify the original line. The source CSV has no header and all 128 lines are data. Do not use a default-header CSV reader and silently discard the first word.

There is no mapping from these symbols to synthesized DNA sequences, proteins or characterized bond strengths. The eight-slot random-greedy codebook is separate from the four-slot GF(4) affine code used by the illustrative compiler and from the 16-slot format counted in kinetic rows.

Claim records

id names a stable scientific claim (G1, P1, P2, C1, C2, A1, F1, H1, S1, U1, N1, E1). novelty_class follows the papers: A established; B direct synthesis; C extension; D novel hypothesis; E unverified speculation. Combined values preserve qualified classification. status, assumptions and does_not_imply must travel with any quotation or retrieval result. evidence_path is repository-relative; evidence_section points to the supplement section when appropriate. Original evidence wording is preserved.

All rows carry experiment_performed=false and priority_established=false for this project's proposed platform. These flags do not deny the external prior art's experimental results.

Research-text chunks and references

Each research chunk contains document, heading, exact text, source path, inclusive one-based line start/end, source-file SHA-256 and text SHA-256. Chunks partition the three source files without overlap or text loss. A long section is split at 60 lines, so a mathematical expression may cross a chunk boundary; retrieve adjacent chunks when needed. No generated abstract or paraphrase replaces source text. evidence_type=mixed_source_text_read_claim_labels is deliberate: one section can contain known physics, a conditional theorem and a proposal.

Reference records preserve citation_and_access_note, optional doi, and source URLs. These notes document where access was limited to abstracts or indexed primary records. They are not copies of external full texts and do not warrant that every cited procedure was read in full.

Regeneration and license

Run python tools/build_exports.py from the release checkout. This rebuilds JSONL, export counts/hashes and FULL_TEXT.md without network access or numerical recomputation. If sources change, revalidate and create a new versioned manifest before publishing.

Original metadata/text/synthetic data use CC BY 4.0 to the extent rights subsist. Code uses MIT. See ../LICENSE.md. Cite the immutable release tag or commit and identify these as synthetic outputs or conditional research, not laboratory measurements.