From a Desired Body to the Rules That Build It

Summary

Body measurements define a broad family of candidate rules.

The compiler's output is a developmental program: a compact seed and local rule that grow the requested form, maintain it while it acts, and restore its organization after perturbation. Ultimately, the target should describe a reliable developmental basin, so changed starts and moderate damage still converge on the requested anatomy. We therefore examine geometry, trajectories, function, and robustness as distinct parts of the specification.

Each Flow Lenia rule in this analysis has 40 parameters. We ran 5,000 distinct rules and described each stable result with 15 measurements covering body form, mass, energy, and movement. We then asked how much those measurements narrow down the rule that produced them.

Although each rule has 40 parameters, the sampled rules vary along about 25.5 independent directions. Their 15 measured outcomes vary along only about 5.9. The 19.6-direction gap is a coarse estimate of the rule variation available within a measured region of body space. A second calculation estimated a similar gap of 20.9 directions.

A neighbor analysis found the same breadth. Rules that produced nearby measured bodies were separated by 98.6% of the distance between randomly paired rules. The compiler can exploit these alternatives: propose several rules, run them, and choose according to the anatomy, stability, or behavior specified by the target.

Ultimate goal

Compile a requested anatomy into a compact seed and a local rule that can grow, maintain, and repair it.

Present finding

Rules producing nearby measured bodies remain almost as far apart as random pairs.

Current strategy

Generate several candidates, run each one, and choose with criteria suited to the target. The three approaches differ in what the target specifies and which variables can change.

5,000stable and distinct forms

19.6 Dglobal rule dimensions left after measuring the body

37.7 / 40mean local rule dimensions left by the descriptors

0.986neighbor ratio (shuffled = 1.0)

01 / THE INVERSE

Relating rules to bodies

Nearby body measurements can come from distant rules.

Phenotype space contains the measurements made on each body. We call the set of rules that produce similar measurements a fiber. A large fiber means that the body measurements constrain the generating rule only loosely.

Schematic of a phenotype base space with genotype fibers above several phenotypes. — 01A
Each point below represents a set of body measurements. The rules compatible with those measurements form the vertical set above it. The geometry is illustrative rather than measured.
reused schematic

Two different sections through genotype fibers selecting different rules over the same phenotype positions. — 01B
Because several rules can fit the same target, the system needs a policy for choosing among them. Different policies may select rules with different unmeasured dynamics.
reused schematic

A path through genotype space projects to a phenotype path until it enters a failure band where no stable phenotype is produced. — 01C
A smooth change in rule parameters can still pass through a region where no viable body develops. Search therefore has to preserve viability as well as reduce distance to the target.
reused schematic

Figures 01A–01C are conceptual diagrams. Figures 02–04 are generated from the frozen anatomical-compiler results.

A dimension bar decomposing the genotype into measured phenotype dimension and inverse-fiber dimension. — 02
The corpus occupies about 25.5 intrinsic rule dimensions and 5.9 intrinsic phenotype dimensions. Their 19.6-dimension difference is estimated from the geometry of the samples; it is not the raw count of 40 parameters minus 15 measurements.
5,000 forms

Observed inverse locality ratio with a narrow interval compared with the shuffled null. — 03
The ratio is 0.986, where 1.0 means that nearby measured bodies are no closer in rule space than shuffled pairs. The difference from 1.0 is detectable, but it is small.
cohort diagnostic

Heatmap of seven local Jacobian singular values across twelve genotypes. — 04
This local test perturbed 12 rules and followed seven descriptors. About five response directions crossed the threshold, but most sensitivity was concentrated in roughly 2.3 effective directions.
12 local probes

The 19.6 and 37.7 dimension estimates answer different questions. The first compares the global intrinsic dimensions of the 5,000-rule corpus and its 15 measured outcomes. The second subtracts the effective rank of a seven-descriptor local response from the full 40-parameter rule. Both indicate substantial ambiguity and should not be compared numerically.

02 / THE COMPILER

Three targets, three inverse problems

Three target types require three different searches.

All three searches use forward simulation. The target may be a handful of measurements, a complete spatial field, or an observed field from a known rule family. Each target defines a different success criterion.

Approach 01 / descriptors

Match trajectory measurements

Search for a rule whose trajectory has the requested mass, occupied area, variation, and spatial spread.

Target: Four summary measurements
Starting field: A fixed generic patch
Search variables: The full 40-parameter rule

See specimen 2876

Approach 02 / form

Maintain a supplied body

Use the requested spatial field as the initial condition and search for a rule that keeps it coherent and active.

Target: The complete multichannel field
Starting field: The target body itself
Search variables: The full 40-parameter rule

See the form demonstration

Approach 03 / parameters

Recover hidden rule values

Keep the rule family and starting cells fixed, then fit the unknown parameters to an observed field.

Target: Orbium at step 600
Starting field: Known native cells
Search variables: R, m, and s

See the Orbium recovery

Shared search loop

Specify the target

Choose measurements, a spatial field, or a known-family observation. That choice determines what a successful result means.

approach-specific

Choose starting rules

Retrieve nearby rules, sample a conditional model, or draw values from a declared parameter range.

proposal stage

Simulate and score

Run every candidate forward and compare its measurements or field with the target.

forward rollouts

Refine and replay

Concentrate the next search around the better candidates, then replay the selected rule beyond the scored interval.

validation

The results below test parts of this system. The conditional proposal check used eight samples across two targets. The refinement result used four targets. The MLX and Swift comparison covered one rule, one starting field, and six steps.

Small conditional proposal test

The conditional model produced stable bodies but missed the target measurements.

1A1B1C1D2A2B2C2D

All eight simulations remained stable. Their mean descriptor error was 2.79 standardized units, where zero is a perfect match, and an average of 3.6 parameters per sample had to be clamped back into range. This is a two-target mechanism check, not an accuracy estimate for the full target population.

Four dumbbells showing every target moving from a higher start cost to a lower refined cost. — 05
Lower is better on this standardized descriptor distance. Refinement improved all four targets, reducing the mean from 1.149 to 0.576, although one target still ended at 0.947.
four targets

Line chart of MLX versus Swift disagreement across steps two through six. — 06
For one genotype and initialization seed, MLX and Swift remained closely aligned through step six. This check covers the beginning of one trajectory.
implementation check

03A / DESCRIPTOR MATCHING

Approach one

The fitted rule matched four measurements and grew a different anatomy.

Specimen 2876 was selected from the 5,000-form corpus because its developed body is compact, tapered, and easy to recognize. The search receives four measurements from its 1,200-step trajectory. It does not receive the target images.

Observed phenotypeThe source rule grows from a generic patch into a tapered body with a bright central ridge. The video covers 1,200 simulation steps.

Fitted ruleThe selected rule grows a segmented body with a different spatial arrangement. Its mass and mass variation are close to the target. Occupied area and spatial spread remain farther away.

The fitted organism reproduces several bulk measurements, yet its anatomy is visibly different. The segmented structure shows how much body-plan variation remains when the target specifies only four measurements.

799.9 → 791.1mean mass · 1.1% relative error

0.0141 → 0.0141mass variation over the recorded trajectory

11.7% → 10.5%occupied area · 10.2% relative error

1005.0 → 1152.8spatial spread · 14.7% relative error

Leaf-like source organism midway through its development.

Observe the full trajectory

The body changes as it develops, so a single frame misses part of the phenotype. The target measurements cover the full trajectory.

The same source organism near the end of the measured trajectory.

Define the scoring measurements

Mass, mass variation, occupied area, and gyration, a measure of spatial spread, define the request. Orientation, segmentation, topology, and locomotion are absent from the target.

measurement	target	fitted
mass	799.9	791.1
occupancy	11.7%	10.5%
gyration	1005.0	1152.8

Body grown by the nearest retrieved rule before refinement.

Retrieve a starting rule

The target specimen itself was excluded from retrieval. The search finds twelve other corpus entries with nearby measurements, simulates them, and starts from the lowest-cost proposal. This is one leave-one-out example, not a cohort evaluation.

Body grown by the best rule after six refinement generations.

Refine through simulation

Six generations of 32 candidates reduced the MLX descriptor cost from 6.99 to 5.26, a 25% decrease. A separate Swift replay measured a standardized cost of 0.975. The two values come from different forward implementations and are reported separately.

07A

The largest improvement occurred in the first generation. Later generations changed the rule only slightly under the MLX objective.

6 × 32 rollouts

Initial retrieved organism. — retrieval · cost 6.99

Best organism after generation one. — generation 1

Best organism after generation two. — generation 2

Best organism after generation six. — generation 6 · cost 5.26

Radial kernel profiles, growth responses, and parameter values for the fitted three-kernel rule. — 07B
The fitted genotype contains one shared radius and three kernel blocks. Each block specifies a radial profile and a growth response. The second kernel has zero growth height in the selected rule, so two kernels drive the final dynamics.
40 parameters

03A.1 / WHAT THE TARGET OMITTED

Similar measurements, different anatomy

Four measurements preserved scale while anatomy changed.

The target and recovered trajectories have similar mass and occupied area, but their late bodies remain visibly different. Segmentation and pixelwise structure are absent from the objective, so they have no effect on the score.

Late target body with one tapered central form. — 08A
The source trajectory retains one dominant tapered body with peripheral material.
target

Late fitted body split into repeated segments. — 08B
The fitted rule distributes similar mass across a repeated segmented arrangement.
fitted

03B / FORM MAINTENANCE

Approach two

We searched for a rule that could maintain a supplied body.

In this experiment, the target is a body taken from the dataset, and its complete field is also the initial condition. Candidate rules are judged by whether that field remains coherent, active, and topologically similar during the recorded rollout.

Target multichannel field, the target reseeded under the fitted rule, and a coherent body grown from noise under the same rule. — 09
The center panel tests the actual objective: whether the fitted rule maintains the supplied field. The right panel asks a separate question. Starting the same rule from generic noise produces a coherent body, but not the requested anatomy.
one form demo

0.095 → 0.060best recorded form drift

96%largest-component fraction in the target

10recorded search iterations

This experiment solves maintenance: the fitted rule holds an existing anatomy together. The full compiler must also discover a compact starting pattern and the rule that develops it into the requested body.

03C / ORBIUM INVERSE

Approach three

We recovered three Orbium parameters from one mature field.

This experiment isolates within-family parameter recovery. We supplied Orbium's update law, kernel, growth function, and starting cells, then hid three numerical parameters: the interaction radius R, growth center m, and growth width s. The search evaluates values by comparing each simulated body with the reference at step 600.

Fix the model

The search uses the native qd24_additive_v1 update and Orbium's one-kernel geometry. Other update laws and kernel layouts are not considered.

known structure

Set a wrong starting point

The search starts from R=10, m=0.12, and s=0.024. These values disperse the initial cells into a ring and small fragments.

wrong start

Search the full ranges

The first stage evaluates 2,688 parameter combinations. Every candidate runs for 600 steps.

28 × 96 rollouts

Refine the best region

The second stage evaluates 1,152 candidates near the broad-search winner. The selected rule is then run to step 1,200.

12 × 96 rollouts

Orbium simulated at step 600 with the deliberately incorrect starting parameters. — **Initial proposal**The starting rule breaks the seed into a diffuse ring and small fragments.

Native Orbium target at step 600. — **Observed body**This single step-600 field is the image used by the final-frame search.

Compiled Orbium at step 600. — **Compiled result**The recovered parameters grow one localized body with the target's structure.

The initial parameters fail. The fitted parameters produce a close reconstruction. Each candidate is simulated for 600 steps, then its complete field is compared with the reference field.

Hidden native ruleThe target is Chakazul O2(a). The search is scored against its step-600 field. The rule family and starting cells are supplied separately.

Fitted ruleThe body stays coherent through step 1,200, although the search was scored only at step 600.

10 → 12.94interaction scale R · native 13

0.12 → 0.1526growth center m · native 0.150

0.024 → 0.0180growth width s · native 0.017

3,840candidate rollouts · broad search plus local refinement

10

The narrow second stage improved the field match from 0.878 to 0.969 and brought all three hidden values close to the native rule.

broad search → local refinement

step 100 · unseen

step 300 · unseen

step 600 · target

step 1,200 · unseen

native rule

Native Orbium at step 100. — early history

Native Orbium at step 300. — early history

Native Orbium at step 1200. — held-out future

compiled rule

Compiled Orbium at step 100. — not scored

Compiled Orbium at step 300. — not scored

Compiled Orbium at step 1200. — not scored

The fitted parameters also produced a coherent body at the unscored steps. This experiment isolates parameter recovery by supplying the update family, kernel geometry, growth function, starting cells, and elapsed time.

We repeated both stages with three frames

A control search used steps 100, 300, and 600. It returned R=12.80, m=0.1435, and s=0.0185. Its replay also remained coherent through step 1,200.

The single-frame run recovered closer values

Both runs used the same broad and refinement budgets. The three-frame run ended farther from the hidden values than the run scored only at step 600. This is one organism, so it does not establish that single-frame targets are generally better.

04 / FUNCTION

Predicting behavior from shape

Shape predicted activity far better than locomotion.

We measured six behaviors in 8,174 creatures and reduced them to principal axes. The first axis is dominated by speed, path length, displacement, and center velocity, so we call it locomotion. The second is dominated by energy and temporal variance, so we call it activity. Together they account for 87.2% of the behavioral variation.

Scatterplot of 8,174 creatures in the first two functional principal axes, colored by speed. — 11
Most creatures occupy a dense low-speed region, while fast locomotion forms a sparse tail. Any model trained on this distribution will see many more slow examples than fast ones.
8,174 creatures

ActivityR² 0.725

LocomotionR² 0.038

TortuosityR² 0.110

Occupancy and spread helped predict activity. The same shape measurements explained only 3.8% of the locomotion axis. Matching form measurements does not determine how the result will move.

Observed versus predicted functional scores from the shape-to-function regression. — 12
Activity predictions cluster near the diagonal. Locomotion predictions remain compressed near the bottom of the plot.
descriptive regression

05 / LIMITS

Development across initial conditions

Endpoint density was a poor guide to developmental motion.

We fitted a potential surface to the regions where trajectories ended, then compared its downhill direction with their observed motion. The alignment was weak. In a separate perturbation test, five of eight genotypes became more dispersed over time. These results give no evidence that a selected rule will correct perturbations or converge from nearby starts.

Eight inferred basins in a two-dimensional descriptor space, sized by weight, with aggregate downhill alignment 0.124. — 13
Across 3,600 trajectory points, observed drift aligned with the fitted downhill direction by 0.124. An alignment of 1 would follow the fitted landscape exactly; zero would show no preferred alignment.
descriptive

Slopegraph of early versus late phenotype spread for eight genotypes run from six initial conditions. — 14
Three rules narrowed the differences between their six initial conditions, while five widened them. The mean late-to-early spread ratio was 1.39, where values above 1 indicate divergence.
8 × 6 rollouts

Bulk shape was much more repeatable than motion across starting conditions. Across the same six initial conditions, mean variation was 2.5% for shape and 51% for motion. A rule can produce a broadly consistent form while leaving its trajectory highly contingent on the start.

Scatterplot of complexity versus drift for 80 coherent and random creatures. — 15
Across 80 creatures, complexity and drift had a correlation of r = 0.045. In this sample, visual complexity provided almost no information about whether a body would remain near its initial location.
80 creatures

Coverage map of the occupancy and gyration archive used for compiler warm starts. — 16
The warm-start archive occupied 195 of 576 cells (33.9%). The empty regions show body measurements for which this corpus offers no direct proposal.
proposal coverage

06 / SYNTHESIS

What works now

Three working searches define the path to the full compiler.

A full system will connect them: describe a target through geometry, function, and development; explore the many rules that can produce it; then search jointly for a compact seed and a rule whose development remains reliable after changed starts or damage.

What the experiments show

In the 5,000-form regime, many rules produce similar measured bodies.
The conditional model produced eight stable rules across two held-out targets; their mean descriptor error remained 2.79.
Simulation-guided refinement lowered descriptor cost for all four historical targets.
The organic descriptor case matched bulk measurements while producing a different anatomy.
In one form-maintenance demonstration, a fitted rule kept a supplied field coherent and active.
Inside the known additive family, the Orbium search recovered three nearby parameter values from one mature field.
The Orbium reconstruction remained coherent for 600 steps beyond its search horizon.
Shape predicted the activity axis much better than the locomotion axis.

What the full compiler needs next

Measure descriptor-matching success across a held-out target cohort.
Test form maintenance across many supplied fields.
Search jointly for a compact seed and its developmental rule.
Select or infer the rule family before continuous parameter search.
Determine when a short trajectory adds useful information beyond a mature final frame.
Test whether MLX and another implementation remain equivalent over long trajectories.
Add locomotion and other behaviors directly to the target.

07 / RECEIPTS

Files and digests

Each chart points back to a specific file.

The ledger records the source for every number, specimen, replay, and reused schematic. SHA-256 identifies the exact bytes. The June files do not contain the revision that produced those historical results; the September case-study entries include their exact configs and media.

Show source ledger

Kind	Source	Bytes	SHA-256 prefix
frozen artifact	`stage1_fresh_fiber.json`	1,424	`1fcae1ea8b1613f7…`
frozen artifact	`stage1_jacobian_fiber.json`	4,632	`9e5e975176e31401…`
frozen artifact	`stage2_cinn.json`	1,244	`d5ecd3b9f31ca5b6…`
frozen artifact	`stage3_refine.json`	551	`be159631b5f0097f…`
frozen artifact	`compiled/demo1/result.json`	1,930	`cf085fb32904ed5f…`
frozen artifact	`compiled/demo1/target.png`	18,649	`de6ebc2862bb8589…`
frozen artifact	`compiled/demo1/creature.png`	16,356	`53dcf3f8d11f758a…`
frozen artifact	`compiled/demo1_cinn/result.json`	1,889	`a9f5988e1fa3c754…`
frozen artifact	`compiled/demo1_cinn/creature.png`	15,957	`58e787e1bf2451f2…`
frozen artifact	`compiled/form_demo/result.json`	6,826	`044567fdf4ea014c…`
frozen artifact	`compiled/form_demo/form_compile.png`	66,658	`b3b03d35b8af0ba2…`
frozen artifact	`mlx_validation.json`	1,039	`79316f3cb993eef9…`
frozen artifact	`qd_archive/archive.json`	350,936	`614a1919941427fd…`
frozen artifact	`qd_archive/coverage.png`	2,321	`a6ed57c87e362506…`
frozen artifact	`waddington_flow.json`	2,141	`6537452489f02b52…`
frozen artifact	`canalization.json`	2,357	`c5a5924d1ae5c946…`
frozen artifact	`functional-morphospace/functional_morphospace.json`	1,685	`e8f6e4b61b319865…`
frozen artifact	`functional-morphospace/shape_function_coupling.json`	1,398	`b9923f08f942351a…`
frozen artifact	`functional-morphospace/functional_morphospace.png`	65,958	`52f66e8726ea151a…`
frozen artifact	`functional-morphospace/shape_predicts_function.png`	13,462	`101f3f6ed749b081…`
frozen artifact	`3c15/functional_map_mlx/functional_map.json`	9,209	`e69a9cec67faaa34…`
frozen artifact	`3c15/functional_map_mlx/bodies_montage.png`	139,541	`55b2a1b2466e1bab…`
reused explanatory schematic	`site/assets/blog/lenia-morphospace-report/fig-fiber-bundle.svg`	4,621	`43000b7e7db2b7c0…`
reused explanatory schematic	`site/assets/blog/lenia-morphospace-report/fig-sections.svg`	4,182	`569fec81d4c9dd71…`
reused explanatory schematic	`site/assets/blog/lenia-morphospace-report/fig-bridge-genotype-phenotype.svg`	3,035	`363d5905c657c2e8…`
2026-09-01 case-study rerun	`compiled/descriptor-organic-2876-final/result.json`	15,065	`b5076bb598eb5baa…`
2026-09-01 case-study rerun	`compiled/organic-2876-trial/result.json`	7,625	`c1193ac52130e294…`
2026-09-01 case-study rerun	`media/organic-2876/target-base.json`	1,890	`96712d27919bb1b9…`
2026-09-01 case-study rerun	`media/organic-2876/recovered-base.json`	2,145	`7e61c204d374891b…`
2026-09-01 case-study rerun	`media/organic-2876/search.json`	1,051	`8b5d8113be1a7d28…`
2026-09-01 case-study rerun	`media/organic-2876/target-body.mp4`	561,891	`e70259f567a5b091…`
2026-09-01 case-study rerun	`media/organic-2876/recovered-body.mp4`	957,390	`9b42766440cf423b…`
2026-09-01 case-study rerun	`media/organic-2876/target-run/anatomical-target-2876/frames_color/frame_000600.png`	108,733	`7396b4565167a62f…`
2026-09-01 case-study rerun	`media/organic-2876/target-run/anatomical-target-2876/frames_color/frame_001190.png`	122,830	`676f2be888b9ae58…`
2026-09-01 case-study rerun	`media/organic-2876/recovered-run/anatomical-recovered-2876/frames_color/frame_000600.png`	174,617	`b136b78fcb83bd47…`
2026-09-01 case-study rerun	`media/organic-2876/recovered-run/anatomical-recovered-2876/frames_color/frame_001190.png`	121,202	`da36d4748bdf8ecb…`
2026-09-01 case-study rerun	`compiled/descriptor-organic-2876-final/trace-00.png`	17,882	`fa03595c4170e155…`
2026-09-01 case-study rerun	`compiled/descriptor-organic-2876-final/trace-01.png`	15,293	`1c4156380fc3ec66…`
2026-09-01 case-study rerun	`compiled/descriptor-organic-2876-final/trace-02.png`	14,439	`8a31ce5df13ad3bb…`
2026-09-01 case-study rerun	`compiled/descriptor-organic-2876-final/trace-06.png`	14,818	`67ef9c151f42f916…`
2026-09-01 case-study rerun	`compiled/organic-2876-trial/target.png`	13,121	`47aac3a2842dedcf…`
2026-09-01 case-study rerun	`compiled/organic-2876-trial/creature_swift.png`	11,834	`2eb60b71e7951b21…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/experiment.json`	1,284	`e7d17cd2fca63593…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/still-es.json`	197,817	`c8262451e333bf09…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/still-run/best.json`	141	`2e99f52c9591bc8c…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/still-run/history.jsonl`	12,161	`540933e403833518…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/trajectory-es.json`	591,583	`dbb42ebb131bbec6…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/trajectory-run/best.json`	143	`97902a5fa38d9c4f…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/trajectory-run/history.jsonl`	12,205	`b3d53ac97bb5e77a…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/report-assets/orbium-initial-600.png`	255,120	`040ffb2271b6a55b…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/report-assets/orbium-inverse-summary.json`	1,271	`904916197598ed88…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/report-assets/orbium-rule-recovery.svg`	6,418	`a0125a9a26e7e89b…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/report-assets/orbium-still-100.png`	187,493	`88843f7bf4824041…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/report-assets/orbium-still-300.png`	186,697	`b7b91bc0d08bd62f…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/report-assets/orbium-still-600.png`	185,894	`e7eab62cced81915…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/report-assets/orbium-still-1200.png`	189,664	`4c4cdf26049b32a0…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/report-assets/orbium-still-zoom.mp4`	812,518	`37933bf5d5c5bb09…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/report-assets/orbium-target-100.png`	186,009	`2befab83732f67c4…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/report-assets/orbium-target-300.png`	187,779	`2aacf01203662f32…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/report-assets/orbium-target-600.png`	186,927	`e0ca77a833258f14…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/report-assets/orbium-target-1200.png`	186,735	`2d39c4a8ffcb94ed…`
2026-09-02 within-family Orbium inverse	`orbium-system-id-v5/report-assets/orbium-target-zoom.mp4`	834,738	`12c499769fe12ce6…`
2026-09-02 narrow Orbium refinement	`orbium-system-id-v6-refine/refine-es.json`	197,819	`7bcda98c84215935…`
2026-09-02 narrow Orbium refinement	`orbium-system-id-v6-refine/refine-run/best.json`	143	`43c2f254196e5526…`
2026-09-02 narrow Orbium refinement	`orbium-system-id-v6-refine/refine-run/history.jsonl`	5,165	`e28679376dddcbf6…`
2026-09-02 narrow Orbium refinement	`orbium-system-id-v6-refine/trajectory-refine-es.json`	591,585	`840f48b9aab5b4be…`
2026-09-02 narrow Orbium refinement	`orbium-system-id-v6-refine/trajectory-refine-run/best.json`	140	`93cfbf398087a085…`
2026-09-02 narrow Orbium refinement	`orbium-system-id-v6-refine/trajectory-refine-run/history.jsonl`	5,192	`7b2bca59b1f57f3b…`

The June and September evidence have different provenance. The frozen June artifact JSON files do not record their source revision, so artifactProducingRevision remains empty. The September organic rerun and within-family Orbium inverse are listed separately with exact configs, results, frames, and videos. The complete machine-readable ledger is source-manifest.json.

From a desired body to the rules that build it.