{ "release_id": "gi-promoter-atlas-designs-2026-08-05", "record_kind": "generated_design_release", "version": "1.0.0", "generated_utc": "2026-08-16T03:08:41Z", "title": "GI Promoter Atlas — designed promoters, the schema-enforced artifact of record", "what_this_is": "Model-designed promoter sequences for two on/off cell-type pairs, generated by a module-level genetic algorithm seeded from natural promoters and scored by the GI expression model in a fixed genomic scaffold. This tree is the packaged, checksummed, schema-validated record of what shipped; it is not published.", "distribution_status": "THE SCHEMA-ENFORCED ARTIFACT OF RECORD, NOT A PUBLICATION. This tree is the packaged, checksummed and schema-validated form of the designs that shipped, rebuilt from the selection artifacts by one deterministic command. It is not published anywhere and nothing depends on its being published; it exists so that the claim rules are enforced by a schema -- every design margin carries the best natural promoter in the SAME scaffold, the ratio to it, and a scaffold whose admission verdict is `admitted` -- rather than asserted by a sentence in a document. The research-use-only, non-commercial and not-wet-lab-validated statements below are claim rules and are correct regardless of who reads them.", "relationship_to_the_natural_promoter_release": { "this_release": "generated designs. No design is an unmodified natural promoter, so no memorisation asterisk attaches to them.", "the_other_release": "natural human promoters found by screening the genome. Those carry a training-split asterisk: the model was trained on the TSS window of essentially every gene, so a high rank there is partly recall of training data.", "rule": "The two are separate releases with separate file names and separate limitation blocks. Do not merge them into one table or one ranking — they carry different caveats." }, "model": { "model_id": "g0-expression", "revision": "v1", "checkpoint": "20260523", "dna_encoder": "AIRI-Institute/moderngena-large", "dna_encoder_revision": "01de8202d4cf5e79d87d3e4f71f3e51ae923392b", "endpoint": "https://api.genomicintelligence.ai/v1/tasks/expression/predict", "openapi": "https://api.genomicintelligence.ai/v1/openapi.json", "unit": "log(TPM+1), natural log of quantile-normalised TPM plus one", "input_window": "9,198 bp, TSS-centred at 0-based offset 4,599, gene sense" }, "counts": { "pairs_shipped": 2, "designs_recorded": 832, "designs_reportable": 402, "designs_shipped": 64, "designs_recorded_per_run": { "A1FAM_floorrule": 64, "A1FAM_margin": 64, "A1_prod_genomic_module": 64, "A1_prod_genomic_module_floorfix": 64, "A1_prod_genomic_long": 64, "A2_prod_genomic_module": 64, "A4_FAM47C_margin": 64, "A4_AAVS1_margin": 64, "CCR5_A1_margin": 64, "CCR5_A2_margin": 64, "OFFARM_FAM47C_margin": 64, "OFFARM_FAM47C_floorrule": 64, "A5_Keppel19_margin": 64 }, "designs_reportable_per_run": { "A1FAM_floorrule": 17, "A1FAM_margin": 48, "A1_prod_genomic_module": 45, "A1_prod_genomic_module_floorfix": 34, "A1_prod_genomic_long": 23, "A2_prod_genomic_module": 47, "A4_FAM47C_margin": 0, "A4_AAVS1_margin": 34, "CCR5_A1_margin": 50, "CCR5_A2_margin": 25, "OFFARM_FAM47C_margin": 20, "OFFARM_FAM47C_floorrule": 14, "A5_Keppel19_margin": 45 }, "designs_segregated_out_of_distribution": 64, "designs_per_pair": { "A1_cardiomyocyte__vs__hepatocyte": 17, "A2_skeletal_muscle_myofiber__vs__hepatocyte": 47 }, "not_shipped_entries": 3, "pairs_attempted_and_not_shipped": 3, "unmodified_natural_promoters_shipped": 0, "generated_sequences_checked_for_novelty": 896, "generated_sequences_total_all_runs": 912 }, "default_view": { "rule": "Show the passing set by default and state, visibly, how many records are hidden and why. Never hide the denominator; never make it the first thing a reader has to scroll past.", "showing": 402, "of": 832, "hidden": 430, "hidden_by_first_failed_gate": { "S4_off_arm_no_louder_than_random_dna": 334, "S5_diversity_from_already_selected": 96 }, "also_hidden": { "segregated_out_of_distribution": 64, "why": "plasmid cassette, out of distribution; see harvest/runs.json" } }, "facets": { "note": "Counts over harvest/harvest.jsonl, precomputed so a front end can offer these as filters without reading every record.", "selected": { "not_selected": 430, "selected": 402 }, "training_provenance": { "generated": 832 }, "scaffold": { "AAVS1_plus": 64, "CCR5site_plus": 128, "FAM47C": 384, "FOXB2": 192, "Keppel19_minus": 64 }, "scaffold_status": { "REJECTED as a design environment; adopted for A2's control panel": 128, "rejected for A4": 64, "rejected for A5": 64, "scaffold of record": 384, "superseded design environment": 192 }, "pair_id": { "A1_cardiomyocyte__vs__hepatocyte": 512, "A2_skeletal_muscle_myofiber__vs__hepatocyte": 128, "A4_cns_neuron__vs__hepatocyte": 128, "A5_cns_neuron__vs__astrocyte": 64 }, "margin_verdict": { "reportable": 832 }, "run_id": { "A1FAM_floorrule": 64, "A1FAM_margin": 64, "A1_prod_genomic_long": 64, "A1_prod_genomic_module": 64, "A1_prod_genomic_module_floorfix": 64, "A2_prod_genomic_module": 64, "A4_AAVS1_margin": 64, "A4_FAM47C_margin": 64, "A5_Keppel19_margin": 64, "CCR5_A1_margin": 64, "CCR5_A2_margin": 64, "OFFARM_FAM47C_floorrule": 64, "OFFARM_FAM47C_margin": 64 } }, "harvest": { "what_this_is": "Every candidate from every production generation run, published with its record of all six selection criteria rather than filtered down to what clears the strictest combination. 768 records across four genomic design environments, of which a minority clear the unchanged S1-S6 rule; a further 64 candidates were evolved in a plasmid cassette and are segregated rather than tabulated. Read `selected` and `gate_record` before reading a margin.", "reading_rule": "FIVE SCAFFOLDS, ONE COMPARABLE QUANTITY. These records span FAM47C (the scaffold of record), FOXB2 (superseded), CCR5site_plus and AAVS1_plus (safe-harbour integration coordinates, both rejected as design environments) and a plasmid cassette (out of distribution, segregated). A raw margin is scaffold-bound and means nothing without its scaffold named: the identical 600 bp ALB promoter scores -0.012 hepatic in FOXB2 and +4.844 in FAM47C. Only the design-over-nature RATIO -- a design's margin over the best NATURAL promoter's margin in the same window, the same run and the same two verbatim context strings -- crosses a scaffold boundary. Every run below carries its own natural baseline for exactly that reason. And `designs_recorded` counts rows, never passes.", "scaffolds": { "FAM47C": { "kind": "gene_tss_window", "status": "scaffold of record", "what": "A real, quiet gene TSS window whose -500..+100 proximal promoter is replaced by the 600 bp design. This is the regime the model was trained on and the environment every shipped design was measured in." }, "FOXB2": { "kind": "gene_tss_window", "status": "superseded design environment", "what": "The first gene-TSS scaffold used for A1. Superseded by FAM47C on dynamic range: the identical 600 bp ALB promoter scores -0.012 hepatic here and +4.844 in FAM47C, and A1's design-over-nature ratio moved 2.73x -> 1.35x on nothing but that change. Runs measured here are real records; their raw margins are not comparable with FAM47C's." }, "CCR5site_plus": { "kind": "safe_harbour_integration_site", "status": "REJECTED as a design environment; adopted for A2's control panel", "what": "A safe-harbour integration coordinate, not a transcription start site, so it is further out of distribution than a gene-TSS window. Raw margins rise here and the best NATURAL promoter in the same window rises more, so the design-over-nature ratio FALLS in both pairs (A1 1.351x -> 1.168x, A2 2.31x -> 1.256x). This is the clearest demonstration in the project that the ratio and not the raw margin is the quantity." }, "AAVS1_plus": { "kind": "safe_harbour_integration_site", "status": "rejected for A4", "what": "A safe-harbour integration coordinate. The EMPTY scaffold is already loud on the neuronal arm, and the designs evolved here collapse from a 6.331 median to 0.802 when re-scored in a real gene TSS window." }, "Keppel19_minus": { "kind": "safe_harbour_integration_site", "status": "rejected for A5", "what": "A safe-harbour integration coordinate, not a transcription start site. It is the only environment of twelve that admitted A5 on a criterion cleared by more than the API-vs-GPU path-disagreement band, and the admission is still `hollow: true` -- carried by ENO2 alone of eight panel members. The designs evolved here collapse from a 3.961 median margin to 0.090 when re-scored in FAM47C, with a margin-rank Spearman of 0.0724: the two environments do not agree on which design is better at all." }, "cassette": { "kind": "plasmid_cassette", "status": "OUT OF DISTRIBUTION -- segregated, never tabulated with genomic margins", "what": "A transgene cassette in a plasmid. The OFF arm is inert in this environment (docs/measurements/CASSETTE_PREFLIGHT.md), so a 'margin' measured here degenerates to on-target strength and is not the same quantity as a genomic-scaffold margin. A cassette number is a ride-along control, never a gate on generation, and never a row in the same table." } }, "runs": [ { "run_id": "A1FAM_floorrule", "pair_id": "A1_cardiomyocyte__vs__hepatocyte", "scaffold": "FAM47C", "scaffold_status": "scaffold of record", "role": "shipped", "counts": { "designs_recorded": 64, "designs_clearing_S1_to_S4": 24, "diversity_ceiling_S5": 17, "designs_selected": 17, "note": "`designs_recorded` counts rows in this file and says nothing about whether any of them passed a gate. `designs_selected` is the number this project stands behind for this run. Never read one for the other." }, "design_over_nature": { "best_design_margin": 6.3228, "best_natural_margin": 4.6801, "ratio": 1.351, "note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, "caveats": [] }, { "run_id": "A1FAM_margin", "pair_id": "A1_cardiomyocyte__vs__hepatocyte", "scaffold": "FAM47C", "scaffold_status": "scaffold of record", "role": "sibling of the shipped A1 run, same scaffold, margin fitness", "counts": { "designs_recorded": 64, "designs_clearing_S1_to_S4": 61, "diversity_ceiling_S5": 48, "designs_selected": 48, "note": "`designs_recorded` counts rows in this file and says nothing about whether any of them passed a gate. `designs_selected` is the number this project stands behind for this run. Never read one for the other." }, "design_over_nature": { "best_design_margin": 5.9248, "best_natural_margin": 4.6801, "ratio": 1.266, "note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, "caveats": [ "This is a second A1 run in the scaffold of record. It is NOT an independent replicate: it shares its seeds, its scaffold window and its contexts with the shipped run and differs only in the fitness term. Two runs of the same harness over the same starting material do not corroborate each other." ] }, { "run_id": "A1_prod_genomic_module", "pair_id": "A1_cardiomyocyte__vs__hepatocyte", "scaffold": "FOXB2", "scaffold_status": "superseded design environment", "role": "superseded scaffold", "counts": { "designs_recorded": 64, "designs_clearing_S1_to_S4": 47, "diversity_ceiling_S5": 45, "designs_selected": 45, "note": "`designs_recorded` counts rows in this file and says nothing about whether any of them passed a gate. `designs_selected` is the number this project stands behind for this run. Never read one for the other." }, "design_over_nature": { "best_design_margin": 4.18, "best_natural_margin": 2.0563, "ratio": 2.0328, "note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, "caveats": [ "Measured in FOXB2, a superseded design environment. The raw margins in this run are not comparable in magnitude with any FAM47C number; only the design-over-nature ratio crosses that boundary.", "Its fitness term carried an on-target floor penalty that was later corrected (the `floorfix` run is the correction), so its generation-0 `best_fitness` is not a margin. Its generation zero is byte-identical to the floorfix run's -- same rng seed, seed set, scaffold window, operator and population size, all asserted before the log is reused -- so the natural baseline is the same measurement, not a substituted one." ] }, { "run_id": "A1_prod_genomic_module_floorfix", "pair_id": "A1_cardiomyocyte__vs__hepatocyte", "scaffold": "FOXB2", "scaffold_status": "superseded design environment", "role": "superseded scaffold", "counts": { "designs_recorded": 64, "designs_clearing_S1_to_S4": 37, "diversity_ceiling_S5": 34, "designs_selected": 34, "note": "`designs_recorded` counts rows in this file and says nothing about whether any of them passed a gate. `designs_selected` is the number this project stands behind for this run. Never read one for the other." }, "design_over_nature": { "best_design_margin": 4.2166, "best_natural_margin": 2.0563, "ratio": 2.0506, "note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, "caveats": [ "Measured in FOXB2, a superseded design environment. Raw margins are not comparable with FAM47C's; the ratio is." ] }, { "run_id": "A1_prod_genomic_long", "pair_id": "A1_cardiomyocyte__vs__hepatocyte", "scaffold": "FOXB2", "scaffold_status": "superseded design environment", "role": "superseded scaffold, long run", "counts": { "designs_recorded": 64, "designs_clearing_S1_to_S4": 36, "diversity_ceiling_S5": 23, "designs_selected": 23, "note": "`designs_recorded` counts rows in this file and says nothing about whether any of them passed a gate. `designs_selected` is the number this project stands behind for this run. Never read one for the other." }, "design_over_nature": { "best_design_margin": 5.621, "best_natural_margin": 2.0563, "ratio": 2.7336, "note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, "caveats": [ "Measured in FOXB2, a superseded design environment. Raw margins are not comparable with FAM47C's; the ratio is." ] }, { "run_id": "A2_prod_genomic_module", "pair_id": "A2_skeletal_muscle_myofiber__vs__hepatocyte", "scaffold": "FAM47C", "scaffold_status": "scaffold of record", "role": "shipped", "counts": { "designs_recorded": 64, "designs_clearing_S1_to_S4": 55, "diversity_ceiling_S5": 47, "designs_selected": 47, "note": "`designs_recorded` counts rows in this file and says nothing about whether any of them passed a gate. `designs_selected` is the number this project stands behind for this run. Never read one for the other." }, "design_over_nature": { "best_design_margin": 4.5822, "best_natural_margin": 1.9851, "ratio": 2.3083, "note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, "caveats": [ "This run stopped at generation 34 under a plateau rule that has since been superseded, so its design-over-nature ratio is a floor and not a ceiling." ] }, { "run_id": "A4_FAM47C_margin", "pair_id": "A4_cns_neuron__vs__hepatocyte", "scaffold": "FAM47C", "scaffold_status": "scaffold of record", "role": "attempted, nothing selected", "counts": { "designs_recorded": 64, "designs_clearing_S1_to_S4": 0, "diversity_ceiling_S5": 0, "designs_selected": 0, "note": "`designs_recorded` counts rows in this file and says nothing about whether any of them passed a gate. `designs_selected` is the number this project stands behind for this run. Never read one for the other." }, "design_over_nature": { "best_design_margin": null, "best_natural_margin": 2.016, "ratio": null, "note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, "caveats": [ "HOLLOW ADMISSION, and it travels with every one of these 64 records. FAM47C's admission for A4 is carried by exactly one neuronal control, SYT1, by 0.0014 log units -- and SYT1 is also A4 seed 7. An environment admitted by the optimiser's own starting material is not an independent test of the environment. Every other member of the eight-member pre-named neuronal panel fails its own floor here.", "All 64 candidates fail S4: every one is louder in hepatocyte than this run's own composition-matched random-DNA negatives (+0.0069..+0.0693 against a negative OFF max of -0.0091). A4 is the first pair whose two arms are positively rank-correlated (Spearman ON/OFF +0.503) -- pushing the neuronal arm up pulls the hepatocyte arm up with it.", "Nothing here is selected and nothing here is a design this project stands behind. These rows are published because a gate-annotated dataset that omits its clearest failure is worth less than one that shows it." ] }, { "run_id": "A4_AAVS1_margin", "pair_id": "A4_cns_neuron__vs__hepatocyte", "scaffold": "AAVS1_plus", "scaffold_status": "rejected for A4", "role": "attempted, not shipped", "counts": { "designs_recorded": 64, "designs_clearing_S1_to_S4": 45, "diversity_ceiling_S5": 34, "designs_selected": 34, "note": "`designs_recorded` counts rows in this file and says nothing about whether any of them passed a gate. `designs_selected` is the number this project stands behind for this run. Never read one for the other." }, "design_over_nature": { "best_design_margin": 6.9696, "best_natural_margin": 2.2573, "ratio": 3.0876, "note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, "caveats": [ "The EMPTY AAVS1_plus scaffold already scores 5.6 on the neuronal arm against 0.0024 for the empty FAM47C scaffold. Signal the design did not put there is signal that cannot be attributed to the design.", "Re-scored in FAM47C, a real gene TSS window, these designs collapse from 6.249 / 6.331 / 6.970 (min / median / max where they were evolved) to 0.153 / 0.802 / 3.146, and the design-over-nature ratio falls 3.088x -> 1.561x.", "A safe-harbour integration coordinate is further out of distribution than a gene TSS window and the cost of that has not been measured." ] }, { "run_id": "CCR5_A1_margin", "pair_id": "A1_cardiomyocyte__vs__hepatocyte", "scaffold": "CCR5site_plus", "scaffold_status": "REJECTED as a design environment; adopted for A2's control panel", "role": "superseded scaffold", "counts": { "designs_recorded": 64, "designs_clearing_S1_to_S4": 64, "diversity_ceiling_S5": 50, "designs_selected": 50, "note": "`designs_recorded` counts rows in this file and says nothing about whether any of them passed a gate. `designs_selected` is the number this project stands behind for this run. Never read one for the other." }, "design_over_nature": { "best_design_margin": 8.2705, "best_natural_margin": 7.0806, "ratio": 1.1681, "note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, "caveats": [ "CCR5site_plus was REJECTED as a design environment. Raw margins rise here (6.3228 -> 8.2705) and the natural baseline rises more (4.6801 -> 7.0806), so the design-over-nature ratio FALLS 1.351x -> 1.168x. Quoting a raw margin from this run without the ratio inverts the finding." ] }, { "run_id": "CCR5_A2_margin", "pair_id": "A2_skeletal_muscle_myofiber__vs__hepatocyte", "scaffold": "CCR5site_plus", "scaffold_status": "REJECTED as a design environment; adopted for A2's control panel", "role": "superseded scaffold", "counts": { "designs_recorded": 64, "designs_clearing_S1_to_S4": 35, "diversity_ceiling_S5": 25, "designs_selected": 25, "note": "`designs_recorded` counts rows in this file and says nothing about whether any of them passed a gate. `designs_selected` is the number this project stands behind for this run. Never read one for the other." }, "design_over_nature": { "best_design_margin": 7.4758, "best_natural_margin": 5.9511, "ratio": 1.2562, "note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, "caveats": [ "CCR5site_plus was REJECTED as a design environment: A2's design-over-nature ratio falls 2.308x -> 1.256x here even though every raw margin rises." ] }, { "run_id": "OFFARM_FAM47C_margin", "pair_id": "A1_cardiomyocyte__vs__hepatocyte", "scaffold": "FAM47C", "scaffold_status": "scaffold of record", "role": "off-arm descent experiment", "counts": { "designs_recorded": 64, "designs_clearing_S1_to_S4": 22, "diversity_ceiling_S5": 20, "designs_selected": 20, "note": "`designs_recorded` counts rows in this file and says nothing about whether any of them passed a gate. `designs_selected` is the number this project stands behind for this run. Never read one for the other." }, "design_over_nature": { "best_design_margin": 6.1407, "best_natural_margin": 4.6801, "ratio": 1.3121, "note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, "caveats": [ "OFF-ARM SEEDED. This run was seeded from sequences that are LOUD in the off-target arm, so unlike A1 and A2 there was an off-target gradient to descend, and it was descended (4.844 -> 0.005). That is the only demonstration of active descent in the project.", "CULLING, NOT CONVERSION. The population reached quietness by selection among many starting sequences, not by any single sequence being converted from loud to quiet. SILENCE, NOT REPRESSION: a low off-arm score is the absence of predicted expression, not evidence of an active repressive element." ] }, { "run_id": "OFFARM_FAM47C_floorrule", "pair_id": "A1_cardiomyocyte__vs__hepatocyte", "scaffold": "FAM47C", "scaffold_status": "scaffold of record", "role": "off-arm descent experiment", "counts": { "designs_recorded": 64, "designs_clearing_S1_to_S4": 15, "diversity_ceiling_S5": 14, "designs_selected": 14, "note": "`designs_recorded` counts rows in this file and says nothing about whether any of them passed a gate. `designs_selected` is the number this project stands behind for this run. Never read one for the other." }, "design_over_nature": { "best_design_margin": 6.4537, "best_natural_margin": 4.6801, "ratio": 1.379, "note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, "caveats": [ "OFF-ARM SEEDED, with the same two caveats as its sibling run: culling not conversion, silence not repression.", "Not an independent replicate of the other off-arm run: same seeds, same scaffold window, same contexts, different fitness term." ] }, { "run_id": "A5_Keppel19_margin", "pair_id": "A5_cns_neuron__vs__astrocyte", "scaffold": "Keppel19_minus", "scaffold_status": "rejected for A5", "role": "attempted, selected, NOT shipped", "counts": { "designs_recorded": 64, "designs_clearing_S1_to_S4": 57, "diversity_ceiling_S5": 45, "designs_selected": 45, "note": "`designs_recorded` counts rows in this file and says nothing about whether any of them passed a gate. `designs_selected` is the number this project stands behind for this run. Never read one for the other." }, "design_over_nature": { "best_design_margin": 4.5079, "best_natural_margin": 2.3204, "ratio": 1.9427, "note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, "caveats": [ "THE MARGIN DOES NOT LEAVE THIS SCAFFOLD, and that travels with every one of these 64 records. Re-scored in FAM47C, the scaffold of record, the same 64 designs fall from a margin span of 3.812-4.508 (median 3.961) to -0.005-0.546 (median 0.090) -- every one below the 0.80 reportable threshold -- and the Spearman correlation of the margin ordering across the two environments is 0.0724. The margin is a property of the Keppel19 integration site, not of the designs.", "HOLLOW ADMISSION. Keppel19_minus's admission for A5 is carried by exactly one member of the pre-named eight-member neuronal panel, eno2__region_600bp. ENO2 is at least not one of A5's own seeds and is a genuinely held-out gene, which is why this run happened where A3's did not -- but the environment is admitted at the strength of one promoter and no more.", "NO GENE-TSS SCAFFOLD ADMITS A5 AT ALL. In FAM47C, FOXB2, OR4C13 and GSX1 the ON panel fails criterion C2 -- including hSYN1, the field's standard neuron-restricted promoter, whose neuron-over-astrocyte margin in FAM47C is 0.4619 as a published cassette and 0.4902 as a genomic window, both `not_a_finding`. On this checkpoint the field's own neuronal standard does not separate neuron from astrocyte at reportable resolution.", "SELECTIVITY HERE IS DIFFERENTIAL GAIN, NOT SILENCING. A5's off arm was NOT quiet at generation zero (astrocyte median 0.734, max 2.031). Over 40 generations the ON median rose 1.367 -> 5.359 while the OFF median rose 0.734 -> 1.379. Nothing in this run taught the model to repress the astrocyte arm.", "A5 IS NOT VALIDATED AND MUST NEVER BE DESCRIBED AS SUCH. CLAUDE.md fact 6 records A5 as `not demonstrated` in the held-out-only re-ranking. Both arms are also ontology subtype narrowings of their registry cell types -- thalamic neuron for `cns_neuron`, striatal astrocyte for `astrocyte`." ] } ], "segregated": { "why": "A transgene cassette in a plasmid is out of distribution for this model (CLAUDE.md fact 7) and is not the regime this project reports. The OFF arm is inert in a cassette, so a 'margin' measured there degenerates to on-target strength and is not the same quantity as a genomic-scaffold margin. The cassette-optimised ranking did not transfer: the best design from this run ranked 63rd of 64 on the genomic axis. These 64 candidates are therefore recorded with their run's ride-along controls and their distribution, and never as rows beside a genomic-scaffold margin. A cassette number is a ride-along control, never a gate on generation.", "runs": [ { "run_id": "A1_prod_cassette_module", "pair_id": "A1_cardiomyocyte__vs__hepatocyte", "scaffold": "cassette", "scaffold_note": "A transgene cassette in a plasmid. The OFF arm is inert in this environment (docs/measurements/CASSETTE_PREFLIGHT.md), so a 'margin' measured here degenerates to on-target strength and is not the same quantity as a genomic-scaffold margin. A cassette number is a ride-along control, never a gate on generation, and never a row in the same table.", "candidates_generated": 64, "candidates_published_as_rows": 0, "ride_along_controls": [ { "variant_label": "ctnt__published_cassette", "name": "cTnT", "species": "Gallus gallus (chicken)", "length_bp": 429, "on_score": 5.1562, "off_score": -0.0361 }, { "variant_label": "ctnt__published_cassette_genomic_core", "name": "cTnT", "species": "Gallus gallus (chicken)", "length_bp": 414, "on_score": 5.1562, "off_score": -0.041 }, { "variant_label": "ctnt__region_600bp", "name": "cTnT", "species": "Homo sapiens", "length_bp": 600, "on_score": 5.6562, "off_score": -0.0405 }, { "variant_label": "mck__published_cassette", "name": "MCK", "species": "synthetic construct (mouse Ckm-derived enhancer/promoter + mouse alpha-MHC enhancer)", "length_bp": 792, "on_score": 3.8594, "off_score": -0.0342 }, { "variant_label": "mck__published_cassette_ck8", "name": "MCK", "species": "synthetic construct (mouse Ckm-derived)", "length_bp": 450, "on_score": 2.9219, "off_score": -0.0452 }, { "variant_label": "mck__published_600bp", "name": "MCK", "species": "synthetic construct (mouse Ckm-derived enhancer/promoter + mouse alpha-MHC enhancer)", "length_bp": 600, "on_score": 0.7148, "off_score": -0.041 }, { "variant_label": "mck__region_600bp", "name": "MCK", "species": "Homo sapiens", "length_bp": 600, "on_score": 0.0128, "off_score": -0.0435 }, { "variant_label": "hsyn1__published_cassette", "name": "hSYN1", "species": "Homo sapiens", "length_bp": 470, "on_score": 0.1055, "off_score": -0.0312 }, { "variant_label": "hsyn1__published_cassette_genomic_core", "name": "hSYN1", "species": "Homo sapiens", "length_bp": 456, "on_score": 0.0991, "off_score": -0.0309 }, { "variant_label": "hsyn1__region_600bp", "name": "hSYN1", "species": "Homo sapiens", "length_bp": 600, "on_score": 0.1533, "off_score": -0.0347 }, { "variant_label": "tbg__published_cassette", "name": "TBG", "species": "Homo sapiens (synthetic composite)", "length_bp": 669, "on_score": 0.1396, "off_score": 1.125 }, { "variant_label": "tbg__published_cassette_tbg_bare", "name": "TBG", "species": "Homo sapiens", "length_bp": 410, "on_score": 0.2158, "off_score": 0.062 }, { "variant_label": "tbg__published_600bp", "name": "TBG", "species": "Homo sapiens (synthetic composite)", "length_bp": 600, "on_score": 0.1357, "off_score": 1.2812 }, { "variant_label": "tbg__region_600bp", "name": "TBG", "species": "Homo sapiens", "length_bp": 600, "on_score": 0.1592, "off_score": 0.0535 } ], "negatives": { "n": 12, "kinds": [ "dinucleotide_shuffle", "random_gc_matched" ], "on_score_max": 0.1387 }, "distribution_only": { "note": "Distribution only. No per-record row, no ratio, no comparison with any genomic-scaffold number.", "n": 64, "on_target_strength_min": 5.8659, "on_target_strength_max": 6.4777 }, "caveats": [ "A plasmid cassette is out of distribution for this model and is not the regime this project reports (CLAUDE.md fact 7). The OFF arm is inert in a cassette, so this run's 'margin' is on-target strength wearing a margin's name.", "These 64 records are NOT tabulated with any genomic-scaffold margin, carry no design-over-nature ratio, and are not designs this project stands behind. The cassette-optimised ranking did not transfer: the best design from this run ranked 63rd of 64 on the genomic axis." ] } ] } }, "claim": "SELECTIVE ON THE DELIVERED PROPERTY: strong in the target cell type, and measured quiet in the off-target cell type, in a scaffold that is quiet in both when empty and whose off-target arm is demonstrably live. MECHANISM FOOTNOTE, required wherever the mechanism is at issue: this selectivity is SELECTION, NOT LEARNED REPRESSION. The natural seeds were already quiet in hepatocyte, so there was no off-target gradient to descend; what the optimiser had to do was hold that arm quiet while raising the on arm, and it did. Active descent of a LOUD off arm is demonstrated only in a separate, differently-seeded experiment (hepatic 4.844 -> 0.005), whose own caveats travel with it: culling not conversion, and silence not repression. Do not say the model was taught to silence the off-target arm.", "claim_scope": "THE SENTENCE ABOVE IS NOT A STATEMENT ABOUT THIS WHOLE RELEASE. It is the claim that may be made about a design that CLEARS criterion S4, the off-arm check. This tree records every candidate from every production run, and a large minority of them fail S4 and are published marked as failing it. Each record carries its own `claim` field, which is this sentence for a record that passed and an explicit refusal for one that did not, and its own `gate_record` with the numbers behind the verdict. Read the record, never the manifest, to learn what may be said about a given sequence.", "ratio_note": "The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x).", "no_cross_model_corroboration": "No cross-model corroboration exists for these designs. Cross-model validation was attempted and halted: three control gates failed on cell-type coverage and a fourth on a composition confound, and only 4 of 10 A-tier coverage arms -- the five pairs' ON sides and OFF sides, counted separately -- are reachable by any alternative model that can vote positively at all. Every number here comes from one model on one checkpoint.", "no_memorisation_asterisk": "The designs carry no memorisation asterisk. Across all 896 sequences generated by this project's 14 production generation runs, 0 are unmodified natural promoters: the closest any of them comes to any natural promoter is 77 edits of a 600 bp region, and the closest any SHIPPED design comes is 200 edits of 600 bp. This is the one claim about the designs that is strictly stronger than anything that can be said about a natural-promoter shortlist, where a high rank is partly recall of the model's own training data. Two further harness smoke runs (16 candidates, 912 generated in total) each kept unmutated seeds in their 8-member starting population; those 6 sequences are natural promoters, are flagged as such in their own run artifacts, and are neither designs nor shipped. Verified by local sequence comparison over every generated candidate (scripts/verify_design_novelty.py, data/generation/_novelty_verification.json); the distance is true Levenshtein edit distance on the 600 bp designed region.", "research_use_only": "RESEARCH USE ONLY, NON-COMMERCIAL. Every number in this release is a prediction of the g0-expression model on checkpoint 20260523. Nothing here has been validated in a wet lab, in any cell type, at any length, in any construct. Nothing here is for clinical, diagnostic or therapeutic use. These are design hypotheses offered for collaborative validation, not products.", "licence_note": "Licensed CC BY 4.0 -- the sequences, the tables, the documentation and the pages built from them. The pipeline that produced this tree is proprietary and unpublished; no licence is granted over it. Attribution asks for the release id and the model checkpoint named in this manifest. The designs are novel sequences produced by this project; they descend from natural human promoter seeds and each record states its nearest natural promoter and the edit distance from it. Research use only: nothing here has been synthesised or measured in any cell -- a statement about what this is, not a condition of the licence. Where this tree says 'non-commercial' it records the project's own 2026-07-28 scope decision about ITS use of third-party models; CC BY 4.0 adds no field-of-use restriction and this release does not add one either. Natural control promoter sequences lifted from published cassettes or plasmid deposits remain subject to their depositors' terms and are not redistributed in this tree. Genome annotation is GENCODE V50 / Ensembl 116 on GRCh38, subject to its own terms.", "files": [ { "name": "README.md", "what": "release notes. Limitations first, then the designs" }, { "name": "index.html", "what": "self-contained browsable page — no build step, no external request" }, { "name": "MANIFEST.json", "what": "release provenance, file inventory, all limitations" }, { "name": "designs.jsonl", "what": "the primary artifact — one self-contained record per design, carrying its sequence, both arm scores, its grade, its ratio, its novelty, its lineage, its reasoning, both context strings verbatim and a reproduce block" }, { "name": "designs.tsv", "what": "the same rows, flat, one column per criterion" }, { "name": "designs.fasta", "what": "the 64 selected sequences, headers carrying `kind=generated_design`, `selected=` and both scores" }, { "name": "harvest/harvest.jsonl", "what": "THE FULL HARVEST — one record per candidate from every production generation run, passing or failing, each with its full record of all six criteria. Read `selected` before quoting any of it" }, { "name": "harvest/harvest.tsv", "what": "the same rows, flat, one column per criterion" }, { "name": "harvest/harvest.fasta", "what": "the harvested sequences, headers carrying `selected=` and `failed=`" }, { "name": "harvest/runs.json", "what": "per production run: the scaffold and what it is worth, the six criteria, all controls individually, the negatives, the natural baseline measured in that same window and run, the caveats that travel with its numbers, and the segregated plasmid-cassette run" }, { "name": "contexts.json", "what": "the three context strings verbatim, with SHA-256 and their measured paraphrase floors" }, { "name": "not_shipped.json", "what": "the third pair that was attempted and dropped, and the superseded safe-harbour selections, with the measurements that stopped them" }, { "name": "pairs/*.json", "what": "per-pair: scaffold identity and admission, gates, all controls individually, negatives, the natural baseline in the same run, and the records" }, { "name": "CHECKSUMS.txt", "what": "SHA-256 of every other file" } ], "limitations": [ { "id": "L1", "title": "Not wet-lab validated, and not clinical", "text": "No sequence in this release has been synthesised, transfected, transduced or measured in any cell. Every number is a model prediction. Research use only, non-commercial. Nothing here may inform a clinical, diagnostic or therapeutic decision." }, { "id": "L2", "title": "Selective on the delivered property — by selection, not by learned repression", "text": "SELECTIVE ON THE DELIVERED PROPERTY: strong in the target cell type, and measured quiet in the off-target cell type, in a scaffold that is quiet in both when empty and whose off-target arm is demonstrably live. MECHANISM FOOTNOTE, required wherever the mechanism is at issue: this selectivity is SELECTION, NOT LEARNED REPRESSION. The natural seeds were already quiet in hepatocyte, so there was no off-target gradient to descend; what the optimiser had to do was hold that arm quiet while raising the on arm, and it did. Active descent of a LOUD off arm is demonstrated only in a separate, differently-seeded experiment (hepatic 4.844 -> 0.005), whose own caveats travel with it: culling not conversion, and silence not repression. Do not say the model was taught to silence the off-target arm." }, { "id": "L3", "title": "The ratio is the quantity, not the raw margin", "text": "A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two context strings. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x)." }, { "id": "L4", "title": "No cross-model corroboration", "text": "No cross-model corroboration exists for these designs. Cross-model validation was attempted and halted: three control gates failed on cell-type coverage and a fourth on a composition confound, and only 4 of 10 A-tier coverage arms -- the five pairs' ON sides and OFF sides, counted separately -- are reachable by any alternative model that can vote positively at all. Every number here comes from one model on one checkpoint." }, { "id": "L5", "title": "One checkpoint, with a compressed dynamic range", "text": "Checkpoint 20260523 preserves marker-gene ranking but compresses magnitudes. The reportable margin threshold is 0.80 log units, which is close to the full width of the band most real genes occupy (0.03-0.9). Compare deltas, never absolute magnitudes, and never across scaffolds." }, { "id": "L6", "title": "Scaffold-bound", "text": "Every score here was measured with the 600 bp design substituted for the proximal promoter of one fixed real gene TSS window (FAM47C). That is the regime the model was trained on, and it is not the locus a design would be used at. A margin from this scaffold is not comparable in magnitude with a margin from another. The identical 600 bp ALB promoter scores -0.012 in one admitted scaffold and +4.844 in another." }, { "id": "L7", "title": "Not measured in an AAV cassette", "text": "A transgene cassette in a plasmid is out of distribution for this model. These designs were evolved and scored in a genomic window, not in a cassette. A cassette-optimised ranking did not transfer in this project: the best design from a cassette-environment run ranked 63rd of 64 on the genomic axis." }, { "id": "L8", "title": "The on-target reporting floor excludes nothing", "text": "Both pairs report an on-target floor derived from a small panel of human genomic control promoters scored in this same scaffold before any design was scored. In both pairs `floor / negative_on_max` is below 1, which means the floor is cleared by some random DNA in this scaffold and carries no weight. The banded margin criterion is the one doing the work. Both counts are reported so 'both criteria met' cannot be read as two independent checks." }, { "id": "L9", "title": "Novelty is bounded against the run's own seeds", "text": "Edit distance is measured over the 600 bp region against the 25 natural seeds of that run. It bounds recall of those seeds, not of every promoter in the genome." }, { "id": "L10", "title": "A third pair was attempted and no pair shipped from it", "text": "A third pair (CNS neuron ON vs hepatocyte OFF, A4) was generated, and no shipped pair came out of it: nothing among its candidates is selected, so it has no pair record beside A1 and A2. Its 128 candidates are nevertheless published, in the harvested tables, each marked with the criterion it fails — see L11 for why a record is not a design that passed, and L13 for the measurements that stopped this one. A failure that is countable is worth more than a failure that was deleted." }, { "id": "L11", "title": "The six criteria are recorded, not enforced — a record in this release is not a design that passed", "text": "This release publishes EVERY candidate from every production generation run, each carrying its own record of the six selection criteria: which of them it meets, and by how much. The criteria are recorded rather than enforced. A candidate that fails one is published marked as failing it instead of being deleted, so the criteria themselves are unchanged and what changed is what happens to a candidate that misses one. Three different counts follow, and they are not interchangeable: the number RECORDED is how many candidates were generated and says nothing about any criterion; the number CLEARING THE SELECTION RULE is how many meet all six; the number SHIPPED is the curated per-pair selection. Check whether a record was selected, and read its `gate_record`, before quoting it." }, { "id": "L12", "title": "Six design environments, and only the ratio crosses between them", "text": "The harvested records span FAM47C (the scaffold of record, 384 records), FOXB2 (superseded, 192), CCR5site_plus and AAVS1_plus (safe-harbour integration coordinates, both rejected as design environments, 128 and 64), Keppel19_minus (the A5 integration site, 64: its designs reach a median margin of 3.961 there and 0.090 re-scored in FAM47C, with a rank correlation of 0.0724 between the two, which is why A5 is not shipped) and a plasmid cassette (out of distribution, segregated and never tabulated). A raw margin is meaningless without its scaffold named. Only the design-over-nature ratio against the best natural promoter in the SAME window and the SAME run crosses a scaffold boundary, and every run carries that baseline." }, { "id": "L13", "title": "The A4 records are published on a hollow admission", "text": "The 128 A4 (CNS neuron ON / hepatocyte OFF) records ship with their criteria recorded rather than enforced, and nothing among them is selected. FAM47C's admission for A4 is carried by exactly one neuronal control, SYT1, by 0.0014 log units — and SYT1 is also A4 seed 7, so the environment was admitted by the optimiser's own starting material rather than by an independent control. All 64 FAM47C candidates are louder in hepatocyte than that run's own random-DNA negatives. The 64 AAVS1_plus candidates were evolved in an environment whose EMPTY scaffold is already loud on the neuronal arm, and they collapse to a median 0.802 when re-scored in a real gene TSS window. All three measurements travel with every A4 record: the admission carried by a single control, SYT1, by 0.0014 log units and that control is also A4 seed 7; all 64 FAM47C candidates louder in hepatocyte than their own run's random-DNA negatives; and the 64 AAVS1_plus candidates collapsing to a median 0.802 in a real gene TSS window." }, { "id": "L14", "title": "Both shipped scaffold admissions rest on a single carrier — disclosed, not re-gated", "text": "The advisory hollow-admission check (reported under `scaffold.admission.hollow_admission_check` on every record) fires on both SHIPPED pairs, not only on the pair that was dropped. **A2 / FAM47C** is admitted on ONE carrier of five, `mck__published_cassette`, clearing its own banded floor by 0.0302 — a mouse synthetic fusion with no genomic TSS of its own, while all three HUMAN genomic members of the same panel failed. **A1F / FAM47C** has a single-member panel by construction: `ctnt__region_600bp`, i.e. TNNT2, which is also A1 seed 11 — structurally the same defect that disqualified A4/FAM47C (SYT1 alone, by 0.0014, and SYT1 is A4 seed 7), one pair earlier and in a weaker form. **No verdict moves and nothing is re-gated:** the four admission criteria were fixed before any scaffold was scored, and adjusting them after seeing where they landed is exactly the failure they exist to prevent. Read an admission here as passed, not as broad, and read A4's exclusion as the strong form of a weakness the shipped pairs share. **A2's weakness is narrower than it first reads.** A 110-pair triage table re-derives admission under a mechanical ON panel — the ON context's own committed marker genes as 600 bp genomic windows, a rule whose inputs all predate that work — and under it A2 / FAM47C is admitted by THREE human genomic carriers, NEB, MYF6 and TNNI1. So the single-mouse-fusion carrier is a property of the curated panel we happened to use, not evidence that no human genomic promoter can carry this environment: a second, independently-derived and stronger basis for the same admission. It re-gates nothing, and it does not transfer to A1F, whose one-member panel is a genuine single-carrier case under both rules. Triage is not admission — that table's panel differs from the curated procedure, so a green cell in it may not be quoted as a curated admission." } ], "not_shipped_summary": [ { "id": "A4", "status": "attempted, not shipped", "one_line": "Both candidate design environments fail, and they fail differently. FAM47C's admission for this pair is carried by one control, SYT1, by 0.0014 log units; the AAVS1_plus candidates collapse to a median 0.802 when re-scored in a real gene TSS window. 0 shipped." }, { "id": "A5", "status": "attempted, not shipped", "one_line": "The only environment of twelve that admitted this pair is a safe-harbour integration site, and the margin the run produced there does not exist anywhere else. Same shape as A4/AAVS1_plus, worse." }, { "id": "CCR5-A1 / CCR5-A2", "status": "superseded", "one_line": "Higher raw margins, WORSE design-over-nature ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x): the scaffold lifted the natural baseline more than it lifted the designs, so the ratio and not the raw margin is the quantity." } ] }