# DrosoMem

*Sequence Memory in a Fly Connectome*

DrosoMem studies sequential recall in **computational models using actual Drosophila
connectome structure**. Pi is a trained-sequence benchmark, not evidence of a living
fly memorizing pi or a model predicting unseen pi digits.

[한국어 설명](docs/README-ko.md) · [Research status / audit](docs/research-status.md) ·
[Roadmap](docs/research-roadmap.md) · [Extension closeout](docs/additional-research-closeout.md)

## Current — internal scientific review and manuscript corrections, 2026-10-13

The [revised manuscript](docs/manuscript-draft.md) has completed an internal
[scientific review](docs/manuscript-scientific-review.md). Ten findings are
addressed: episode/rule counts, source/DAN selection, rate/sensitivity/noise
definitions, computational scope, figure captions and the direction of the
conn2res comparison. A thirteen-study panel/head/status table makes the actual
computational replication units explicit. Own-paper author placeholders are
removed as requested; literature attribution is retained.

[Revision checks](results/manuscript_review_v1/audit/checks.json) authenticate
selected quantitative claims, citation identities, canonical document bytes,
local links and the unchanged prior scientific/figure/audit records. No new
model fit or neural replay occurs. Registered outcomes, including M11's
acquisition-limited FAIL, stay unchanged. This internal review does not claim
external peer review or submission. Earlier dated handoffs are historical.

## Archived — original full manuscript and literature context, 2026-10-13

The [full manuscript draft](docs/manuscript-draft.md) develops P1/P2 and M1–M11
into introduction, explicit model/readout methods, results, discussion and
provenance/declarations. [Literature access notes](docs/literature-context.md)
and [BibTeX](docs/manuscript-references.bib) document nine primary references:
seven journal articles, one technical report and one versioned preprint.
Connectome-reservoir precedents and their different tasks are acknowledged.

[Document/provenance checks](results/manuscript_draft_v1/audit/checks.json)
verify all thirteen archived outcomes, selected count/percentage anchors,
citation identities, existing figures, local links and historical preservation.
No new model run, fit or historical neural replay is performed. M11's weak
acquisition and all earlier negative/undefined outcomes retain their meanings.
The deliverable is an internal author draft; byline/declarations and final
scientific review precede any separately authorized external submission.
Earlier dated handoffs below remain historical completion scopes.

## Archived — M1–M11 evidence synthesis, 2026-10-13

The [research synthesis](docs/research-synthesis.md) and
[claim/evidence/limitation matrix](docs/research-evidence-matrix.md) close the
bounded handoff after M11. They preserve13 canonical study outcomes (P1/P2 and
M1–M11), with each question, learned budget, protocol, raw archive, successful
independent checks and interpretation limit linked. No new model is fitted.

The core supported result is operational historical-input access in fixed
computational reservoirs. Tested material real-wiring advantages and M11's
acquisition/continuation gates remain failed; M1 stays undefined, M4 strict
matching INFEASIBLE, M8/M10 tool/task READY. These labels have different meanings.
[Pinned reproduction index](docs/research-reproduction-index.md) separates
producer/verifier revisions, parent/cache inputs and static CLI checks.
[Internal manuscript outline](docs/manuscript-outline.md) maps only supported
claims; no publication/DOI/external submission is claimed.

[Inventory](results/research_synthesis_v1/inventory/checks.json) authenticates1606
manifests and2371 recorded successful source references;19 incomplete/invalid
manifest records remain. All49539 prior result Git identities and10636
materialized protected files agree with the cutoff. Numerical replay was not
performed by this synthesis. [Final document audit](results/research_synthesis_v1/final_audit_retry2/checks.json).
The current programme and synthesis are closed; a full manuscript/literature
context task or a separately registered new research question remains future work.

## Closed M11 rule-continuation primary/secondary FAIL, 2026-10-13

The [M11 matched fitted-model comparison](docs/rule-transfer-results.md) is closed
with **primary/secondary FAIL**. All216 source/ten-null/shift fits and independent
SVD refits complete; all427680 teacher/autonomous state trajectories, predictions,
scores/counts and exact gates replay. Nine guards and smoke validation pass.

Excluding known period-2 rules prospectively, real held-out first accuracy is
13.23/13.94%,8-symbol completion0.51/1.01%, and real−shift completion0.51/1.01pp,
below registered25%/10%/5pp. Training teacher accuracy19.35/18.86% also misses50%,
so acquisition itself is limited under this fixed-reservoir/539-affine-scalar
budget. Real is not above every matched null; no wiring advantage is supported.
Fixed lag2 completes the110 period-2 cases and none of the330 primary cases.

[Protocol](docs/rule-transfer-protocol.md) ·
[Independent checks](results/rule_transfer_v1/main_validation/checks.json) ·
[Final audit](results/rule_transfer_v1/final_audit/audit.json).
M10 task READY, M9 and prior negative results remain unchanged. The bounded
M10/M11 comparison is complete. Next is responsible synthesis/archive and a
claim/evidence/limitation matrix; no automatic head/seed tuning or P3.

## Closed M10 rule-transfer task design READY, 2026-10-13

The [M10 finite task-design preflight](docs/rule-transfer-design-results.md) is
**READY**, independently checked: 990 mod-11 affine episodes separate training,
seen-rule new starts and held-out coefficients. Held-out prefixes/teacher windows/
rotated cycles have zero training overlap. History-3 completes all seen-rule
instances but none of 440 held-out coefficient continuations; the supplied-form
oracle completes all 440. All 63 control evaluations and 2970 total family
episodes reconstruct exactly; 7 safeguards pass. No neural model is fitted.

READY concerns the registered design checks. Some held-out rules have period 2,
so the future comparison also needs explicit lag-copy controls and per-rule
scores. M9's failed wiring advantage and earlier negatives remain sealed.
[Protocol](docs/rule-transfer-design-protocol.md) ·
[Independent checks](results/rule_transfer_design_v1/independent_validation/checks.json) ·
[Final audit](results/rule_transfer_design_v1/final_audit/audit.json).
Next: separately preregister a matched learned-model/shift-reference continuation
comparison, provisionally M11; not yet registered or executed.

## Closed M9 structural advantage FAIL, 2026-10-12

The [M9 fitted then frozen structural tracking study](docs/structural-tracking-results.md)
is closed with **primary/secondary material wiring advantage FAIL**. All87
cases have100% current/lag2 accuracy and64/64 unconditioned replacement pairs:
real, each10 partial/3 whole matched nulls and engineered delay are tied.
Real−null gap0pp misses the fixed3pp/strict-above-all rule. Current access
is valid; ceiling/basic tracking success does not override the failed hypothesis.

All174 full-state train/test hashes,174 independent head refits,5568 pre-states/
paired probes and22272 four-world window/final entries independently replay.
Eleven safeguards pass; no M9 technical failure. Per graph two490-parameter
affine heads learn on4000 train rows, then freeze before independent test;
recurrent/encoder/template learning0.45525 older result identities/failures
remain protected. [Protocol](docs/structural-tracking-protocol.md) ·
[Independent checks](results/structural_tracking_v1/main_validation/checks.json) ·
[Final audit](results/structural_tracking_v1/final_audit/audit.json).

The conditional bank shares many structures/current coefficients; the delay
has different dynamics and a source instantaneous front end. Ties at ceiling
do not prove all-wiring equivalence or biological learning. P2/M7 negatives
stay sealed. Next is proposed mathematical-sequence rule-transfer task design,
separating recall, prediction and unseen-rule application; not registered or
executed. No automatic harder-lag/noise/seed sweep or P3 follows this result.

## Closed M8 design preflight — 2026-10-12

The [M8 structural control construction preflight](docs/structural-specificity-design-results.md)
is **READY**, independently verified:10 partial/3 whole constructions preserve
node degrees/roles, both incoming/outgoing raw weight multisets, incoming-L1
factors and the exact current KC→MBON coefficients. Delayed edge locations
change22.34–23.76% partial /85.94–85.99% whole; all fixed budgets complete.
A separate608-active-coordinate engineered current/lag2 reference fits the
scalar state/48-observation/490-head budget; all18 numerical code ranks validate.

This is construction feasibility, **not measured memory or wiring superiority**.
The delay system changes anatomical identities/dynamics and retains the source
instantaneous front end. Source neural trajectories0/new decoder fits0.
All13 graphs regenerate exactly;9 valid safeguards pass. First smoke's root-ID
type error and its nested record are preserved; the value-preserving retry passes.
45245 old result identities remain protected. [Protocol](docs/structural-specificity-design-protocol.md) ·
[Independent checks](results/structural_specificity_design_v1/main_validation/checks.json) ·
[Final audit](results/structural_specificity_design_v1/final_audit/audit.json).

Next is a distinct preregistered structural **performance** comparison (working
M9, proposal only), with fair graph-specific fitting then frozen held-out heads.
The roadmap's priority2 remains open. P2/M7 FAIL and M4 INFEASIBLE stay sealed;
no P3 or graph/whole-brain/biological superiority claim is added.

## Closed M7 — 2026-10-12

The [M7 artificial observation-noise tracking study](docs/counterfactual-noise-results.md)
has registered **primary FAIL and paired secondary FAIL**. At the one fixed
sigma0.10 dose, joint lag2 tracking is255/384 (66.41%) partial and109/192 (56.77%)
whole secondary; every block misses the predefined90% rule. The same fresh
clean counterparts are100%. Old graph/encoder/scaler/head remain frozen,
with0 new trained parameters; noise is added only to endpoint observations.
No smaller dose, extra seed, new decoder or revised criterion follows failure.

All9 cases,576 noisy pairs,9 baseline state hashes/576 pre-states/2304
four-world windows and9 OLD-head SVD checks independently reproduce this
negative result. All23 safeguards pass; no neural run or independent replay failed.
All44,844 prior result identities and historical scientific files/failures
remain protected. Observation-only failure does not demonstrate internal
information erasure or biological noise robustness/learning.
[Protocol](docs/counterfactual-noise-protocol.md) ·
[Independent checks](results/counterfactual_noise_v1/main_validation/checks.json) ·
[Final audit retry](results/counterfactual_noise_v1/final_audit_retry1/audit.json).
The first closeout's source-record variable error is preserved; the audit-only
repair leaves all scientific evidence/settings unchanged.

Next is the roadmap's **structural specificity** priority: assess feasible
real/rewired/minimal-delay controls, then preregister a distinct bounded
comparison. It remains a proposal; prior P2 FAIL and M4 INFEASIBLE stay sealed.
Rule transfer and responsible release remain later proposals. P3 is excluded.

## Closed M6 — 2026-10-08

The [M6 frozen-readout counterfactual tracking study](docs/counterfactual-tracking-results.md)
passes its prospective lag2 paired-transition criterion:384/384 primary pairs
and192/192 paired whole secondary follow one ACTUAL past-symbol replacement
while prefix/future inputs remain identical. All64 anchors count in every
block; original failures were never filtered. Existing M5 heads/scalers stay
frozen, with0 new trained parameters. Unchanged current/frequency/reference
predictors have exact joint0, which is a paired-control property, not10% chance.

Independent reconstruction confirms9 baseline hashes,576 pre-state entries,
2304 four-world window entries and9 OLD-head SVD authentications. The29-test
valid retry passes; the prior LF/CRLF source-capture failure remains archived.
All44,571 older result identities/scientific source remain protected.
[Protocol](docs/counterfactual-tracking-protocol.md) ·
[Independent checks](results/counterfactual_tracking_v1/main_validation/checks.json) ·
[Final audit](results/counterfactual_tracking_v1/final_audit_retry1/audit.json).
The first artifact audit's CSV column-order failure is retained; its retry
aligns names with unchanged value checks and scientific settings.
This supports fixed computational historical-value tracking, not biological
learning, a storage site, graph superiority or formal capacity.

## Closed M5 — 2026-10-08

The [M5 finite-lag symbol-contrast sensitivity study](docs/temporal-functional-sensitivity-results.md)
passes its prospective lag5 attenuation criterion in all six primary blocks.
Mean normalized local derivative gain is0.66135% of immediate gain; paired
whole firstthree secondary also passes (0.01635%). These are model response
magnitudes, not percentages of memory lost. Fresh iid lag2 decoding is100% in
both cohorts despite attenuated responses, and structural support remains broad.

All9 cases,18 complete-state hashes,189 independent ridge refits,288 complex
probe windows and1512 full-coordinate FD checks validate. Finite valid symbol
replacements differ from local interpolation tangents. No M5 attempt failed;
31 safeguards pass,44,147 historical result identities and M4's invalid guard
snapshot remain preserved. [Protocol](docs/temporal-functional-sensitivity-protocol.md) ·
[Independent checks](results/functional_sensitivity_v1/main_validation/checks.json) ·
[Final audit](results/functional_sensitivity_v1/final_audit/audit.json).
No unique cycle, biological learning or formal capacity claim is added.

## Closed M4 — 2026-10-08

The [M4 exact cycle-control feasibility study](docs/cycle-attribution-feasibility-results.md)
is closed with registered **INFEASIBLE**. Partial labeled degrees/role totals
already force cycles; in both graphs, relevant positive-delay cycles prevent a
same-node DAG from retaining complete all-lag structural input-observation walk
support. All9 fixed mapping cases/90 symbol witnesses validate independently.
Whole degree-only feasibility remains UNRESOLVED. This is a strict design limit,
not a failed cycle-memory hypothesis or a universal ban on causal comparisons.

No new decoding run was added. The29 safeguards pass with authenticated retry;
the first guard's source-capture validation failure and exact snapshot are
preserved. All43,983 prior result identities and old scientific source remain
sealed. [Protocol](docs/cycle-attribution-feasibility-protocol.md) ·
[Independent checks](results/cycle_attribution_v1/main_validation/checks.json) ·
[Final audit](results/cycle_attribution_v1/final_audit/audit.json).

## Closed M3 — 2026-10-08

The [cycle-free feedforward sufficiency study M3](docs/temporal-cycles-results.md)
passes its prospective iid lag2 criterion in all six fresh partial blocks:
acyclic DAG with zero direct carry decodes11998/12000 symbols (99.9833%), with
minimum88.5pp excess above the strongest registered baseline. Whole firstthree
secondary also passes. Current decoding is100% in every arm.

Independent replay confirms54 complete-state hashes,567 ridge refits and9
finite-history certificate pairs. Common12/154-symbol suffixes erase differing
earlier histories exactly under the DAG's structural bounds. This supports
delayed feedforward sufficiency in the fixed computational scope; intact−DAG
differences also include deleted degrees/weights/paths and do not isolate cycles.
M1 stays invalid, M2/P1/P2 conclusions and43,495 prior result identities remain
sealed. [Protocol](docs/temporal-cycles-protocol.md) ·
[Independent checks](results/tdc_cycles_v1/main_validation/checks.json) ·
[Final audit](results/tdc_cycles_v1/final_audit/audit.json).

## Closed M2 — 2026-10-08

The [joint current/history decoding study M2](docs/temporal-tradeoff-results.md)
passes its precommitted fresh-input criterion: carry_only versus instantaneous
gains6.562pp mean lag1–5 decoding while losing15.025pp current accuracy, with
both directions in all six partial blocks. Secondary whole firstthree also
passes its own joint rule. All72 full-state hashes,756 independent refits and36
diagnostic panels validate. The lag focus and hypothesis were informed by M1;
no settings or criteria changed after new outcomes.

This is a specific intervention/readout comparison. Full and synaptic_only
retain100% current accuracy and strong historical decoding, so it establishes
no universal need to sacrifice current access for memory. M1 remains
assay-invalid; prior wiring/whole-brain negatives and42,717 old result-file Git
identities are preserved. [M2 protocol](docs/temporal-tradeoff-protocol.md) ·
[Independent verification](results/tdc_tradeoff_v1/main_validation/checks.json) ·
[Final audit](results/tdc_tradeoff_v1/final_audit/audit.json).

## Closed M1 — 2026-10-08

The [state-carry / previous-state synaptic-drive study](docs/temporal-mechanism-results.md)
is complete with registered outcome **assay-invalid**, not confirmed mechanism
advantage. Four fixed interventions retain current KC-to-MBON input access and
equal decoder budgets. Carry-only current-symbol decoding is84–86%, below the
prospective90% eligibility requirement; primary and secondary endpoints remain
undefined. Full−carry-only mean historical accuracy is descriptively+17.815 pp
partial and+15.970 pp whole, but those gaps do not rescue the failed assay.
All72 full-state trace hashes and756 independent refits validate exactly at
predictions/counts. The first verifier memory failure, technical indexing repair
and all42,015 pre-study result blobs are preserved. No settings or criteria were
tuned. [Protocol](docs/temporal-mechanism-protocol.md) ·
[Independent verification](results/tdc_mechanism_v1/main_validation_retry1/checks.json) ·
[Final audit](results/tdc_mechanism_v1/final_audit/audit.json).
P3 remains excluded; earlier P1/P2 conclusions below are unchanged.

## Closed v1/P1/P2 programme — 2026-10-06

[v1 closeout](docs/drosomem-v1-closeout.md) seals baseline `0cf0867` and its
negative results. The [Temporal Decodability Curve](docs/temporal-memory-curve-results.md)
uses independent iid K10 streams, fixed recurrent weights and equal-budget
ridge decoders of historical inputs. All six partial-graph blocks pass the
prospective lag 1–5 access criterion. Mean accuracy is83.11% at lag 2,43.60% at
lag 5 and 9.62% at lag 20; chance is10%. This supports decodable past information
in a source-derived computational reservoir, not sequence prediction or
biological learning.

The [matched null ensembles](docs/temporal-null-ensemble-results.md) contain20
partial and 10 whole-brain controls preserving each neuron's degrees, role blocks
and raw presynaptic weight multisets. The registered wiring-advantage criterion
fails in both analyses. Partial real−null mean is−2.285 pp raw accuracy; secondary
whole-brain real is above all 10 controls but only+0.985 pp, below the fixed2.7 pp
requirement. Preserve that small positive direction and the failed gate together.
All planned159 cases,318 complete state replays and 3,339 independent lag refits
are validated. Failed console-only generation and same-seed retry records remain.

[Full synthesis and Korean explanation](docs/temporal-memory-synthesis.md) ·
[P1 protocol](docs/temporal-memory-curve-protocol.md) ·
[P2 protocol](docs/temporal-null-ensemble-protocol.md) ·
[Final provenance audit](results/tdc_v2/final_audit/audit.json).
No general real-wiring/whole-brain superiority, storage site, physiological
dopamine learning, living-fly pi memory or formal memory capacity is established.
P3 spiking/plasticity/physiology is excluded from this completed work.

## Sealed v1 studies

The [length × alphabet × sequence-family follow-up](docs/followup-3-results.md)
completed 162 trained-sequence cases and 54 independent-stream probes. Its
registered lag-2 historical-input criterion passes for iid and Markov inputs;
periodic motif recall also reflects easy prediction. The
[whole-brain scale and pathway follow-up](docs/followup-4-results.md) finds no
`brain5` decoding advantage over the partial graph, while its matched
DAN→MBON cut criterion passes. A separate
[whole-brain rewiring control](docs/followup-4b-results.md) finds no registered
real-wiring advantage over one degree- and role-preserving graph. The
[physiology and local-learning follow-up](docs/followup-5-results.md) verifies
local synaptic changes, but its MBON response and fixed-decoder benefit criteria
fail. Each result retains its prospective protocol, raw cases and independent
verification. These are computational findings, not evidence of biological
memory storage or calibrated physiology.

The [DAN→MBON temporal-pathway study](docs/temporal-pathway-results.md) is
complete: 12 paired state panels, 96 new readout fits and 64 independently
replayed autonomous paths. Its preregistered current-path-dominance criterion
fails in both cohorts. A current-only cut improves refitted MBON delayed
decoding by 2.894/2.225 percentage points, while a persistent cut loses
8.408/8.186 points. The interaction does not identify an anatomical memory
storage site. [Protocol](docs/temporal-pathway-protocol.md) and
[independent audit](results/temporal_pathway_main_validation/checks.json).

The [fresh-model coordinate-SD structural confirmation](docs/fresh-coordinate-sd-results.md)
is complete: 96 new readout fits and 1,824 independently replayed autonomous
paths. The registered intact-over-degree criterion fails in both fresh cohorts.
Coordinate-SD versus own-median calibration changes the structural gap, but
does not establish an intact-wiring advantage. The prior bounded extension and
its negative results remain closed.

The [bounded observation-noise extension](docs/additional-research-closeout.md) is closed.
The final [coordinate-SD calibration](docs/coordinate-sd-noise-results.md) criterion is **not confirmed**:
96 reused partial heads, no new fits, all 2,688 paths independently replayed.
This does not establish biological memory, whole-brain superiority or physiological validity.

The bounded ACT I–V studies and the [three additional computational studies](docs/additional-research-results.md)
are complete within their registered scopes. Completion does not mean every hypothesis succeeded.

| Additional question | Outcome |
| --- | --- |
| Does observation-noise allocation help beyond pi? | Confirmed on random digits and shuffled pi, including fresh model/task seeds |
| Is intact partial wiring better than role-preserving controls? | Not confirmed; keep the negative result |
| Does allocation still help with synthetic spatially correlated observation noise? | Confirmed under the registered random/brain1 criterion |

All use 48 observed MBONs and a 482-parameter readout. These studies test noise allocation
and partial wiring; they establish no whole-brain superiority or internal learning. Strong-noise recall remains weak.
These three studies contain 8,640 certificate evaluations (6,336 distinct exposure settings), 1,144 actual evaluation
trajectories, 1,074 independent full replays and 100 new readout fits. Certificates stop at the first
error; they are not full autonomous sequences. Physiological validation remains open.

## Reproduce the studies

Use Python 3.12, a fresh environment and new output directories. Run from the repository root.
For the final coordinate-SD calibration, see the [exact commands](docs/coordinate-sd-noise-results.md#10-reproducibility).
The commands below reproduce the preceding extensions.

```powershell
py -3.12 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements-act1-lock.txt
$env:PYTHONPATH = 'src'
# Prepare the pinned graph cache only if absent:
.\.venv\Scripts\python.exe -m flying.training.phase6 prepare --cache outputs/act1-graphs
.\.venv\Scripts\python.exe scripts/research_suite.py run --stage sequence --out outputs/suite_sequence_new
.\.venv\Scripts\python.exe scripts/verify_research_suite.py outputs/suite_sequence_new --out outputs/suite_sequence_new_validation
.\.venv\Scripts\python.exe scripts/research_suite.py run --stage structure --parent outputs/suite_sequence_new --out outputs/suite_structure_new
.\.venv\Scripts\python.exe scripts/verify_research_suite.py outputs/suite_structure_new --out outputs/suite_structure_new_validation
.\.venv\Scripts\python.exe scripts/research_suite.py run --stage correlation --parent outputs/suite_sequence_new --structural-validation outputs/suite_structure_new_validation --out outputs/suite_correlation_new
.\.venv\Scripts\python.exe scripts/verify_research_suite.py outputs/suite_correlation_new --out outputs/suite_correlation_new_validation
```

Keep the committed parent artifacts. The [protocol](docs/additional-research-completion-plan.md)
defines seeds, budgets, stopping rules and replay subsets; individual reports contain raw tables,
checkpoints and plotting commands. Distinguish **representation**, **decoding** and **internal learning**.
The subsequent graph-relative and coordinate-SD calibrations are also complete.
The fresh-model coordinate-SD study and the subsequent DAN→MBON temporal-pathway
study are complete within their own fixed scopes. [Data and attribution](data/README.md) · [Primary-source review](docs/research.md) ·
[Historical reproduction](docs/historical-reproduction.md) · [License](LICENSE)


## Citation, data terms, and archival

For release and reuse metadata, see [`CITATION.cff`](CITATION.cff),
[Data sources and attribution](DATA_SOURCES.md),
[Limitations](LIMITATIONS.md), and the
[Zenodo release guide](docs/ZENODO.md).

DrosoMem source code is MIT-licensed. FlyWire-derived public data retain their
upstream CC BY-NC 4.0 terms and attribution requirements; the code license does
not relicense those data.