# Seed Model v0.3 closure: scoped findings and reference organism

Closure date: **2026-10-10** (America/Santiago). Status: **development stage closed**.

v0.3 closes with scoped positive, unfavorable, and inconclusive findings about
installed instinct and bounded acquired memory in the evaluated worlds. It does
not close with demonstrated learned hunting, general learned-policy superiority,
learned food recognition, or biological equivalence. Open research questions are
limitations of this closure, not unfinished execution requirements.

The [canonical design record](README.md), [architecture synthesis](ARCHITECTURE_AND_OBSERVED_BEHAVIOR.md),
[orientation contract](ORIENTATION_CONTRACT.md), and
[experience-memory-decision contract](EXPERIENCE_MEMORY_DECISION_CONTRACT.md)
retain their historical decisions. This closure does not amend frozen endpoints,
parameters, hypotheses, or protocols.

## What is installed, acquired, demonstrated, and unresolved

| Category | Scope at closure |
|---|---|
| Installed mechanisms | Local survival ROM; bounded recovery; ideal newborn-relative motion integration; candidate construction and associative selection rules; spatial return/local-search rules. Later experimental profiles also install visual reconstruction/tracking, short-horizon guidance, and relative-action alternatives/arbitration/exploration. |
| Experience-acquired state | Bounded candidate-local energy associations; one replaceable food-place anchor acquired after a positive lived energy consequence; adaptive relative-action values in the relevant experimental profiles. An acquired state or update is not itself a demonstrated behavioral benefit. |
| Demonstrated experimental effects | Utility of the spatial package in the tested static worlds and EXP-005 moving_slow reserve contrasts; unfavorable reserve effect of the tested EXP-006 guidance implementation; no demonstrated intended adaptive-minus-frozen reserve advantage in EXP-006. |
| Unresolved capabilities | Learned hunting, useful waiting, predictive interception, general policy superiority, food/non-food discrimination, unique failure-mechanism attribution, and biological benchmark alignment. |

The reference profile below excludes the later visual guidance and relative-action
selector. The installed rule, acquired state, and experimental package effect
must not be conflated. In particular, spatial storage and return steering were
not separately ablated.

## Development and evidence inventory

Counts below describe declared evaluation cohorts, excluding smoke and replay.
Earlier endpoints differ and are not pooled into one efficacy estimate.

| Experiment | Evaluation lives | Scoped finding |
|---|---:|---|
| [EXP-001](../../v0_3/experiments/exp_001_s0_memory_pilot/REPORT.md) | 64 | Sparse acquired-memory retrieval/action influence; restricted-lifetime enabled-minus-disabled +0.03125 steps [0, 0.09375], without demonstrated survival benefit. The later [smoke/evaluation overlap erratum](EXP_001_ERRATA.md) remains applicable. |
| [EXP-002](../../v0_3/experiments/exp_002_s0_representation_comparison/REPORT.md) | 192 | More active candidate-local reuse, but local-minus-scene reserve -0.0004463 [-0.0339850, +0.0316858]; no positive aggregate reserve advantage. |
| [EXP-003](../../v0_3/experiments/exp_003_s0_navigation/REPORT.md) | 256 | Spatial-minus-local reserve +0.0328677 [+0.0299307, +0.0362185] in abundant static uniform food, supporting the combined spatial package in that setting. |
| [EXP-004](../../v0_3/experiments/exp_004_s0_food_value_response/REPORT.md) | 576 | Robust repeated feeding under fixed-source yield assignments, but late yield-swap allocation Q -0.000916 [-0.002167, +0.000336], not demonstrated richer-source allocation. Q is feeding events/transition, not reserve. |
| EXP-005 | 384 | Local-plus-spatial moving-prey package reserve advantage; acquired storage versus installed steering unresolved. |
| EXP-006 | 512 | Intended relative-learning reserve advantage not demonstrated; tested guidance unfavorable in moving_slow. |
| EXP-007, supplementary | 384 | Delayed-versus-short package advantage not demonstrated; unfavorable point estimate and incomplete executable provenance retained. |

EXP-001–004 contain 1,088 evaluation lives. The central moving-prey closure
evidence is EXP-005 plus EXP-006: **896 lives**. EXP-007 contributes **384
supplementary lives**, separately qualified below. EXP-005–007 therefore contain
1,280 lives, not the total for v0.3; all seven declared cohorts contain 2,368.
No pooled capture/survival mixture is used as a reference-organism estimate.

## Central moving-prey findings

### Endpoint and paired interpretation

For EXP-005–007, full-horizon restricted reserve is the sum of post-transition
bodily energy across the fixed 1,024 transitions, divided by 1,024. After terminal
exhaustion the remaining transitions contribute zero. It measures finite-horizon
reserve maintenance, **not a survival count** or a conditional mean among survivors.

The declared analysis uses paired differences within each of 64 case bundles,
10,000 whole-case bootstrap resamples within the 32 original / 32 reflected
strata, and equally weighted stratum means. Intervals are descriptive 95%
paired-case bootstrap intervals, not pooled independent-sample inference or
claims of equivalence. The table preserves the original reports' displayed
precision; their linked analysis artifacts retain full numerical precision.

| Evidence in moving_slow | Declared contrast | Restricted-reserve estimate | Descriptive 95% interval | Positive / tied / negative |
|---|---|---:|---|---|
| EXP-005 primary | recovery_local_spatial minus recovery_rom | +0.23854 | [+0.15559, +0.31801] | 50 / 0 / 14 |
| EXP-005 secondary | recovery_local_spatial minus recovery_local | +0.26874 | [+0.19794, +0.33913] | 47 / 3 / 14 |
| EXP-006 primary | relative_adaptive minus relative_frozen | -0.009091 | [-0.067395, +0.052980] | 21 / 18 / 25 |
| EXP-006 secondary | guided_local_spatial minus legacy_local_spatial | -0.145416 | [-0.233300, -0.056783] | 22 / 0 / 42 |
| EXP-007 primary, supplementary | relative_delayed minus relative_short | -0.049811 | [-0.110324, +0.008844] | 18 / 17 / 29 |

Immutable evidence: [EXP-005 report](https://github.com/patriciomvera/seed-world-model/blob/f3c545ac46b71500f303c12f6e0004c2c8480836/v0_3/experiments/exp_005_visual_prey_pursuit/REPORT.md),
[EXP-005 analysis](https://github.com/patriciomvera/seed-world-model/blob/f3c545ac46b71500f303c12f6e0004c2c8480836/v0_3/experiments/exp_005_visual_prey_pursuit/results/analysis.json),
[EXP-006 report](https://github.com/patriciomvera/seed-world-model/blob/4d60120a9ad53855fb6aad90d55f1a69b708d485/v0_3/experiments/exp_006_relative_motion_action_learning/REPORT.md),
[EXP-006 analysis](https://github.com/patriciomvera/seed-world-model/blob/4d60120a9ad53855fb6aad90d55f1a69b708d485/v0_3/experiments/exp_006_relative_motion_action_learning/results/analysis.json),
and [supplementary EXP-007 report](https://github.com/patriciomvera/seed-world-model/blob/e3759d4caba6c98470da92661dc45b04a451be54/v0_3/experiments/exp_007_delayed_energy_return/REPORT.md).
Definitions remain in the original [EXP-005 protocol](https://github.com/patriciomvera/seed-world-model/blob/e090753800d40ac90189357daa967de6edc546fe/v0_3/experiments/exp_005_visual_prey_pursuit/PROTOCOL.md),
[EXP-006 protocol](https://github.com/patriciomvera/seed-world-model/blob/4d60120a9ad53855fb6aad90d55f1a69b708d485/v0_3/experiments/exp_006_relative_motion_action_learning/PROTOCOL.md),
and [archived EXP-007 protocol](https://github.com/patriciomvera/seed-world-model/blob/e3759d4caba6c98470da92661dc45b04a451be54/v0_3/experiments/exp_007_delayed_energy_return/PROTOCOL.md).

### Attribution and counterexamples

EXP-005's primary compares a **local-associative-plus-spatial package** against
inherited recovery ROM. Its entire effect cannot be assigned exclusively to the
anchor. The secondary is the more direct comparison for adding the spatial
package while retaining local association. That package combines an acquired
anchor with installed return/search behavior; it does not isolate storage from
steering or demonstrate learned routes. The positive aggregate result is not
uniform: the report's case 049 spatial-minus-ROM reserve difference is
-0.5783462524414062, and 14 primary pairs are negative.

EXP-006 **did not demonstrate the intended reserve advantage**. Its interval
spans unfavorable and favorable values; it does not prove no effect, establish
equivalence, or show that learning generally does not work. The secondary guidance
contrast is unfavorable for **this tested implementation in moving_slow**, not
for all motion perception or guidance. `relative_frozen` freezes relative-action
value learning only; legacy associative and spatial updates remain enabled.

| Reference profile | static_small horizon survivors | moving_slow horizon survivors | Combined |
|---|---:|---:|---:|
| EXP-005 recovery_local_spatial | 33/64 | 38/64 | 71/128 |
| EXP-006 legacy_local_spatial | 38/64 | 33/64 | 71/128 |

This repeats aggregate reference performance under new seeds and another
executable. EXP-006 did **not** repeat the spatial ablation. It is not a new
estimate of the isolated anchor effect. Absence of a uniform static-versus-moving
difference does not establish equivalence.

The inherited moving_slow ROM in EXP-005 captured in 59/64 lives. Capture
occurrence thus cannot establish learned hunting. Failure to sustain feeding is
observed, but a specific search/re-encounter bottleneck remains a hypothesis,
not an identified unique cause. The [v0.2 closeout](../../v0_2/README.md) figure
24.9% is avoidance-mode occupancy (1,912/7,680 transitions), not capture success
or a baseline for these reserve contrasts.

### Temporal and mechanism evidence

The declared EXP-006 adaptive-minus-frozen moving_slow halves are +0.007967
[-0.025574, +0.045380] and -0.026149 [-0.120241, +0.069029]. Its four 256-transition
windows are +0.007618, +0.008316, -0.019069, and -0.033228, each with an interval
including zero. These are predeclared secondary windows, not replacement
endpoints. The complete condition/arm summaries, captures/no-capture counts,
secondary comparisons, and supported planned figures remain in the original
reports and analysis artifacts.

EXP-006 records support approach-mode counts, relative reorientation, zero-speed
issued actions (a narrow operational waiting measure), selector eligibility,
exploration flags, and relative-credit updates. Both frozen and adaptive arms
explore; adaptive credit records distinguish updating from the frozen relative
mechanism. They do not record alternative scores, selected-value retrieval,
tracking state, or direct attribution of an action to acquired values.

Exploratory action selection must be distinguished from value-influenced choice.
A non-exploratory choice may still be neutral or tie-broken; an update does not
show retained retrieval or beneficial later use. The EXP-006 records
cannot establish which retrieval changed an action, whether waiting was useful,
whether a capture resulted from prediction, or how alternatives compared at
choice time. Successful capture or later improvement alone is not proof of
learning. EXP-007 has additional passive fields, but these do not resolve causal
credit or its executable-provenance limitation.

## Supplementary EXP-007 and archival status

One bounded inspection of existing source, configuration, protocol, manifests,
execution labels, and summaries was performed without running the experiment.
The compatible final identity is
`76a1ee93b924253260739c0973f4cd46683e3b9cf0e33b6b6e20d02874cfe46b`;
the existing final checkpoint filenames cover all 384 declared cells.

Its reported primary is **delayed minus short**, not adaptive minus frozen.
Return definition, temporal window, and retention change together. The negative
point estimate and second-window unfavorable result are retained, not discarded
because the interval includes zero. The report and full secondary results are
preserved; they do not demonstrate a delayed-package reserve advantage.

The payload's `executable_commit`,
`4d60120a9ad53855fb6aad90d55f1a69b708d485`, identifies EXP-006 and does not contain
EXP-007. Existing raw sources reproduce the recorded input fingerprint, which
verifies content compatibility. It does not independently prove a complete
pre-execution freeze or the exact implementation loaded at execution.

Archival commit **`e3759d4caba6c98470da92661dc45b04a451be54`** preserves original
experiment artifacts, results, and a byte-preserving archive of the 29 available
seed_model source files. It is an archival commit made now, **not a historical
executable freeze**. Historical provenance labels are unchanged; dirty production
files, local tests, and checkpoint caches are untouched. No missing implementation
was reconstructed. No planned figures or separate detailed mechanism summary
were found, and no replacements were generated. See the immutable
[provenance archive](https://github.com/patriciomvera/seed-world-model/blob/e3759d4caba6c98470da92661dc45b04a451be54/v0_3/experiments/exp_007_delayed_energy_return/PROVENANCE_ARCHIVE.md)
for hashes and precise limits. These limitations retain EXP-007's supplementary
placement and do not prevent closure.

## Reproducible reference organism

The reference is **EXP-005 `recovery_local_spatial`**, not current main or the
new closure-documentation commit.

| Reference identifier | Fixed value |
|---|---|
| Historical executable | `e090753800d40ac90189357daa967de6edc546fe` |
| Evaluation input identity | `ab45e4e42ac0d4a92d9cf42a7848def14025c632ee0e80bd4416acc6d230116f` |
| Root seed / evaluation horizon | 2026092105 / 1,024 transitions or terminal exhaustion |
| Associative capacity / trace / retention | 16 entries / 3 transitions, decay 0.5 / 12 transitions |
| Associative learning rate | 0.5 |

Verified against the [immutable freeze record](https://github.com/patriciomvera/seed-world-model/blob/42e0e26d0938791fab6fb06458c5f1f1962a11c2/v0_3/experiments/exp_005_visual_prey_pursuit/FREEZE_RECORD.md).
The available raw configuration, manifests, runner, and analysis hashes match
that record; canonical JSON hashes and raw-byte hashes use different conventions.

For reproduction, use the complete historical executable tree with its
[configuration](https://github.com/patriciomvera/seed-world-model/blob/e090753800d40ac90189357daa967de6edc546fe/v0_3/experiments/exp_005_visual_prey_pursuit/config.json),
[evaluation manifest](https://github.com/patriciomvera/seed-world-model/blob/e090753800d40ac90189357daa967de6edc546fe/v0_3/experiments/exp_005_visual_prey_pursuit/evaluation_manifest.json),
and [runner](https://github.com/patriciomvera/seed-world-model/blob/e090753800d40ac90189357daa967de6edc546fe/v0_3/experiments/exp_005_visual_prey_pursuit/run.py),
not a reconstruction from this prose or the dirty working tree. These immutable
references specify the geometry, sensing, capture, energy, ties, and seed rules.
No reproduction was executed for closure.

The organism combines inherited survival ROM and recovery; ideal newborn-relative
self-orientation; bounded candidate-local associations; one acquired food-place
anchor; and installed return/local search. Inspect its historical
[transition](https://github.com/patriciomvera/seed-world-model/blob/e090753800d40ac90189357daa967de6edc546fe/v0_3/seed_model/moving_prey.py),
[orientation](https://github.com/patriciomvera/seed-world-model/blob/e090753800d40ac90189357daa967de6edc546fe/v0_3/seed_model/orientation.py),
and [spatial memory](https://github.com/patriciomvera/seed-world-model/blob/e090753800d40ac90189357daa967de6edc546fe/v0_3/seed_model/spatial_memory.py).
This reference has **no visual guidance or relative-action selector**. It uses
ideal sensing/integration and a single anchor, not a learned food classifier,
general map, learned interception policy, or biologically validated organism.

## Dated scope resolution — 2026-10-10

The opening scope discussion included mortality, learned policy, and stone/food
discrimination. Mortality was not necessarily first introduced in v0.3: v0.2
already had energy-bounded lives and starvation terminals.

The canonical V03-H1–H4 document is a **provisional design record**, not a
version-wide requirement that adaptive learning outperform a frozen control.
The evaluated relative-learning comparisons did not demonstrate learned-policy
superiority; a positive acquired-spatial-package contrast is a narrower finding.
Learned food/non-food discrimination was not demonstrated and remains open.
Motion detection did not replace discrimination of food significance.

Broader capacity, cost, environmental-change, and instinct–memory interaction
questions remain open. Closure does not retroactively confirm all hypotheses,
rewrite earlier commitments, or amend experimental endpoints.

## Questions carried forward, without a new experimental design

- Food/non-food discrimination, including moving non-food distractors.
- Perceptual object tracking versus learned food significance.
- Sustained search and re-encounter.
- Spatial storage versus installed return behavior.
- Utility of motion guidance and alternative learning representations.
- Satiety, memory costs, sensory limitations, and biological benchmark alignment.

Existing records do not uniquely identify the failure mechanism or establish
where learning must be applied next. These are future questions, not authorization
for another experiment or v0.4 implementation.

## Closure handoff and checks

This handoff archives existing evidence and updates documentation/navigation only.
Quoted contrasts and reference identifiers were checked against reports, analysis
summaries, the freeze record, and historical Git objects. Changed local Markdown
links, archive contents/hashes, the focused diff, and staged whitespace were
checked. No full test suite or additional engineering fixtures were required.

No scientific lives, smoke/evaluation runs, simulation replay, sweeps, new
experiments, Colab workload, or notebook were run or created. No organism behavior,
learning rule, parameter, manifest, or historical protocol was changed. No public
release or Git tag was created. The closure commits are documentation/archive
history, distinct from the reference executable. Local caches and unrelated
work are preserved. Work stops at this closure and handoff.