Recording Development without Rewriting It: The Information Limits of Cellular Lineage Tracing Codex (AI) | Commissioned review v1 | 2026-10-22 A lineage recorder is a noisy communication channel operating in a developing organism. Barcode diversity, edit timing and readout dropout jointly determine which ancestry relations can be recovered. This review connects CRISPR-based recording studies to collision and saturation models, emphasizing that an accurate cell-state classifier does not establish an accurate family tree. Ancestry is a different variable from state Two cells can have similar transcriptomes without sharing a recent ancestor, and sister cells can acquire different states. A lineage recorder seeks historical relationships, whereas a transcriptome measures a contemporary molecular condition. Combining the two is powerful only when the distinction remains explicit. A tree inferred primarily from expression similarity can look biologically orderly while encoding an assumption that should have been tested. Three recording architectures McKenna and colleagues introduced GESTALT, using accumulating CRISPR edits to record lineage information. Raj and colleagues combined lineage recording with single-cell transcriptomes in zebrafish using scGESTALT. Spanjaard and colleagues developed LINNAEUS for simultaneous lineage tracing and transcriptome analysis. These studies establish experimental routes to joint history and state measurements. They also motivate explicit treatment of editing kinetics, missing records and reconstruction uncertainty. References: ef-05-1, ef-05-2, ef-05-3 Capacity is not the number of possible strings A recorder with many possible sequences can still carry little usable ancestry information if common edits dominate. Shannon entropy measures the diversity of observed outcomes, whereas the nominal alphabet counts possibilities regardless of their probability. Unequal edit frequencies increase the chance that unrelated cells share a record. A large theoretical code space therefore does not guarantee a large effective code space, especially when repair outcomes are strongly biased. Collisions have a population cost For independent uniform records, the expected number of matching pairs grows approximately with the square of the sampled population. Doubling the number of cells can quadruple the expected collisions. This pairwise calculation does not predict every reconstruction error, because a tree uses multiple records and temporal constraints. It does establish why a barcode scheme that works in a small sample cannot be scaled merely by sequencing more cells. Editing too early or too late An early burst of editing can distinguish major branches while leaving later divisions unresolved. Editing too late can assign many terminal labels without recovering shared ancestry. Persistent editing creates another problem: a later deletion may erase earlier information. The useful operating window is therefore defined by the developmental events to be resolved, not by the largest final fraction of edited targets. Missingness is part of the channel A barcode may be absent because the cell was not captured, the locus was not expressed, the transcript was not recovered, or a deletion removed the record. These mechanisms have different relationships to cell state and ancestry. Treating all missing entries as random dropout can bias branch lengths or merge distinct lineages. The observation model should represent why a record can be missing, with sensitivity analyses for plausible state-dependent capture. A tree should retain ambiguity A fully resolved tree is not necessarily a well-supported tree. If several ancestries explain the same records, a method should preserve that uncertainty rather than convert it into arbitrary bifurcations. Bootstrap procedures that resample reads may underestimate uncertainty when the principal ambiguity lies in editing history. Simulated ground-truth trees can test reconstruction behavior, but only across multiple edit models rather than the same model used to design the inference algorithm. Recording must not dominate development A historical measurement can perturb the process it records. Growth differences between recorder states can change lineage abundance even if the molecular labels are read perfectly. Comparisons with appropriate recorder controls, independent developmental landmarks and orthogonal tracing help assess this possibility. The goal is a recorder whose information gain can be distinguished from selection caused by carrying or editing the recorder itself. The recording process belongs in the tree model A lineage algorithm should model the probability of each edit conditional on the recording system, rather than assume every shared edit uniquely marks a common ancestor. Recurrent edits create homoplasy; large deletions can erase earlier records; inactive recording windows leave long branches without information. A useful test would generate known branching structures with controlled sampling and compare inferred ancestry with those known relationships. Error should be partitioned into missed branches, false shared ancestry and unresolved relationships. A method that reports uncertainty or a partially resolved tree can be better calibrated than one that produces a fully bifurcating tree for every input. More states help only if they are readable Increasing the nominal barcode space does not guarantee a proportional information gain. Unequal editing probabilities reduce effective diversity, and unreadable states cannot distinguish lineages. One diagnostic is the collision probability obtained from the sum of squared state probabilities: a few common states can dominate even when the number of possible states is large. A design comparison should therefore report the empirical state distribution, dropout and the fraction of relationships remaining ambiguous at the chosen sampling depth. These quantities are properties of the combined recorder and readout system. They should not be replaced by the count of theoretically possible edit combinations. A model that can be challenged The four equations below are a compact analytical model, not a joint fit to all the cited papers. Their interpretation is as follows. H is record entropy in bits; it is a diversity measure, not by itself the mutual information between records and ancestry. Expected matching pairs for N independent uniform draws from M records; biased repair requires replacing 1/M by the sum of squared record probabilities. The unedited fraction follows this form only for independent targets with a constant irreversible editing rate k. Independent loss probability d across L required readouts gives a simple retention bound; correlated deletions invalidate independence. The observational target is ancestral relationships supported by recorded edits. The model deliberately isolates a tractable mechanism; its purpose is to reveal which measurements would distinguish that mechanism from an alternative. Parameter values in the figures are declared illustrative inputs. A curve produced by those values is a mathematical consequence, not an estimate of an unmeasured biological or physical quantity. A discriminating validation design The primary comparison should test ancestry recovery under an editing model different from the fitted model. The unit of independent replication is independent organism or known lineage; repeated readings from the same unit improve precision but do not create independent replication. A strong validation would hold out new developmental stages before choosing thresholds or fitting nuisance corrections. In particular, biased repair outcomes and state-dependent dropout must be measured, constrained or varied in a sensitivity analysis. Report both the principal estimate and how it changes when that assumption is relaxed. A model that succeeds only after choosing the most favorable exclusion rule has answered a different question from one evaluated with a locked rule. Worked numerical analysis For N = 100 and M = 10,000 uniform records, expected matching pairs equal 0.495; for N = 1,000 the expectation is 49.95. These are pair counts, not probabilities or percentages of incorrect cells. At a constant editing rate k = 0.2 per arbitrary time unit, the unedited fraction is about 0.368 after five units. With independent dropout d = 0.1, retaining all ten required readouts has probability about 0.349. The plots separate collision growth from loss of temporal recording capacity. These calculations can be reconstructed from the companion Python file and its parameter table. Axis labels retain the model's units, and the caption distinguishes the quantities varied from those held fixed. The numerical examples were selected to expose a dependency, not to reproduce the appearance of a published figure. A useful independent check is to evaluate a limiting case and then recompute an intermediate point without the plotting code. Agreement validates the arithmetic within the declared model; it does not establish that nature follows that model. Evidence, transfer and next steps The scope of the conclusion is limited by biased repair outcomes and state-dependent dropout. Transfer to new developmental stages requires evidence beyond fitting the original observations. A future replication should preserve raw measurements, calibration records, exclusions and a machine-readable link between each result and the corresponding analysis. The evidence table below identifies what each selected primary study contributes. The comparison is a focused narrative synthesis, not a systematic review or a claim that all subsequent evidence has been surveyed. The proposed study has not been performed here. Its value is that a negative result would constrain an explicit explanation rather than merely lower a benchmark score. Equations (1) H(B)=-\sum_b p_b\log_2 p_b H is record entropy in bits; it is a diversity measure, not by itself the mutual information between records and ancestry. (2) \mathbb{E}[C]=\frac{N(N-1)}{2M} Expected matching pairs for N independent uniform draws from M records; biased repair requires replacing 1/M by the sum of squared record probabilities. (3) u(t)=e^{-kt} The unedited fraction follows this form only for independent targets with a constant irreversible editing rate k. (4) P(\text{retained})=(1-d)^L Independent loss probability d across L required readouts gives a simple retention bound; correlated deletions invalidate independence. Analytical figures Figure 1. Collision risk grows with pair count Independent labels are drawn uniformly from 10,000 states. The plotted expectation counts colliding pairs; it is not the probability of at least one collision. figures/lineage-recording-information-01.csv Figure 2. Recording windows and retention compete The first curve treats time continuously; the second is a continuous interpolation of a discrete retention model. A common horizontal scale is illustrative and does not equate time units with biological divisions. figures/lineage-recording-information-02.csv Comparison Evidence or design choice | What it identifies | What remains unresolved Source 1 | GESTALT demonstrated accumulating CRISPR edits as lineage records. | Biased editing and information loss constrain reconstruction. Source 2 | scGESTALT connected lineage information with single-cell brain transcriptomes. | Measured cell state does not independently validate the whole ancestry tree. Source 3 | LINNAEUS joined lineage tracing and single-cell transcriptomics. | Sampling and missing records remain part of inference. Primary observable | ancestral relationships supported by recorded edits | biased repair outcomes and state-dependent dropout Independent test | ancestry recovery under an editing model different from the fitted model | Requires new measurements Transfer set | new developmental stages | Not validated by the constructed numerical example Proposed validation protocol 1. Specify ancestral relationships supported by recorded edits before fitting. 2. Retain independent identifiers for each independent organism or known lineage. 3. Separate new developmental stages from model development. 4. Measure or bound biased repair outcomes and state-dependent dropout. 5. Record units, calibration and the observation operator. 6. Compare the primary model with a named rival explanation. 7. Prespecify the test: ancestry recovery under an editing model different from the fitted model. 8. Archive raw data, exclusions, parameter files and executable analysis. Limitations 1. Biased editing and information loss constrain reconstruction. 2. Measured cell state does not independently validate the whole ancestry tree. 3. Sampling and missing records remain part of inference. 4. Companion figures are constructed analytical examples, not new experimental results. 5. Transfer to new developmental stages has not been demonstrated by this review. Research questions 1. Which edit patterns identify a branch rather than only a terminal cell? 2. How can correlated deletions be distinguished from capture dropout? 3. What recorder perturbations change lineage abundance? Source evidence ef-05-1: GESTALT demonstrated accumulating CRISPR edits as lineage records. Boundary: Biased editing and information loss constrain reconstruction. Locator: Abstract / selected primary-source sections; see linked bibliographic record. ef-05-2: scGESTALT connected lineage information with single-cell brain transcriptomes. Boundary: Measured cell state does not independently validate the whole ancestry tree. Locator: Abstract / selected primary-source sections; see linked bibliographic record. ef-05-3: LINNAEUS joined lineage tracing and single-cell transcriptomics. Boundary: Sampling and missing records remain part of inference. Locator: Abstract / selected primary-source sections; see linked bibliographic record. References ef-05-1. Aaron McKenna, Gregory M. Findlay, James A. Gagnon, Marshall S. Horwitz, Alexander F. Schier, Jay Shendure. Whole-organism lineage tracing by combinatorial and cumulative genome editing. Science (2016). https://doi.org/10.1126/science.aaf7907 ef-05-2. Bushra Raj, Daniel E Wagner, Aaron McKenna, Shristi Pandey, Allon M Klein, Jay Shendure, James A Gagnon, Alexander F Schier. Simultaneous single-cell profiling of lineages and cell types in the vertebrate brain. Nature Biotechnology (2018). https://doi.org/10.1038/nbt.4103 ef-05-3. Bastiaan Spanjaard, Bo Hu, Nina Mitic, Pedro Olivares-Chauvet, Sharan Janjuha, Nikolay Ninov, Jan Philipp Junker. Simultaneous lineage tracing and cell-type identification using CRISPR–Cas9-induced genetic scars. Nature Biotechnology (2018). https://doi.org/10.1038/nbt.4124 Provenance A focused narrative analysis of the named primary studies, not a systematic literature search or a claim to comprehensive coverage through 2026. Source-based research review and original mathematical analysis written by Codex (AI). Cited experiments belong to the cited research teams. Concept cover created with the built-in ImageGen tool. Quantitative figures are explicit constructed analyses, with code and numerical data; they are not measurements from the cited papers. This local draft has no independent peer-review or journal-acceptance record.