AlphaMorph Research β release archive (r3 latest; r0, r1 and r2 kept unchanged)
This repository is the independent release archive of AlphaMorph Research. Each release is a tarball of the exact source tree, plus its release manifest. The same trees are published as one git repository on Twigpine, with tags r0, r1, r2 and r3: https://node.twigpine.com/z6Mkq8iCmFCJCGdroHB5SFpwJzUPR8KBTNgEQV7Hbqus3nch/alphamorph-research.git (browse: https://explorer.twigpine.com/repos/z6Mkq8iCmFCJCGdroHB5SFpwJzUPR8KBTNgEQV7Hbqus3nch/alphamorph-research; publisher DID did:key:z6Mkq8iCmFCJCGdroHB5SFpwJzUPR8KBTNgEQV7Hbqus3nch).
The repository holds standard-library-only Python research software and eight reproducible experiments (E0βE7):
- tools that check the evaluation pipeline of rule-based paper-trading studies (r0, r1);
- from r2, the authentic brain-core mechanisms of the AlphaMorph research system, runnable end to end on synthetic observations: attention and signals, memory, inhibition, Go/No-Go, conflict and threat, evolution, evidence-based selection and lineage;
- from r3, the authentic engine (
degenbrain/brain.py) behind an inert model port, deterministic full-engine replay, and single-mechanism counterfactual interventions with complete decision traces.
No release shows profitable trading, successful evolutionary learning, or the full private trading system.
Releases
| r3 (latest): Reproducible Brain Interventions | r2: Open Brain Core | r1 | r0 | |
|---|---|---|---|---|
| Files here | r3/alphamorph-research-r3.tar.gz, r3/EXPORT-MANIFEST.json, r3/SHA256SUMS |
r2/β¦ (unchanged) |
r1/β¦ (unchanged) |
repository root (unchanged) |
| Git commit | c952f36668d4420868f11e09d446bd0d3d6dd088 (parent: r2) |
2ea7fcb58874aac683fb2f8fc4504c2a8331c099 |
80efd37a2edfbc43fb36086192c84b6a9c04b2d5 |
7e102f9b0cb3a7ae6e98483889c5f3eb44f66561 |
| Tag object (unsigned, annotated) | 5c4b34758b389de6bd8d49a475cd85c4d92483f8 |
ca7db008e592ab1853ddb8ebd0e426c8268177e3 |
fe30010642f20043fc8f615183a6fa2b3fb84c9f |
4fc9c6a55482c589f0005cdcca0eb0dd6a4dad02 |
| Tree SHA-256 | 875b45b80111af4ed4c309f4eefa9546c2579b602820a978f4c6009fc358caa1 |
e5a6e71f9a503e97de6cadbcf4567b1ba849480650c8d5b67c14a688a6a0b079 |
fae9d8ddf8ab97575f05043234c5e7682a220c2c0af1ed468ad0e461d70a41be |
0e7ee866313dca347b2c652cd70011cead62c9d9ffa2b5b6680ea4f3f86fcf05 |
| Tarball SHA-256 | 0638b5f3f554db315b96b6029e9bc92129a536b8f3fe151e892860b4e180b07a |
518fb19e71a31796403a5cb525ad86dc35342259cd001d6b34128429f1378033 |
9313a187bf8f882d1f38e9dc9b07fd4175b8e66cd826219f4b565f662e0eb650 |
9fd34078a5bc708cd1ef580f991879af3c6864253a92cabb48ddc8a115b63a44 |
| Contents | r2 plus the engine export, replay, interventions, E7 and docs/INTERVENTIONS.md |
r1 plus the brain core, E6, the decision-trace tool and docs/BRAIN-CORE.md |
r0 plus E4, E5 and the technical report | E0βE3 |
The releases are not cryptographically signed. Hashes and matching copies show that the files are intact and identical across channels. They do not establish who published them (see RELEASE-KEYS.md in the tree).
What r3 adds: Reproducible Brain Interventions
- The authentic engine.
degenbrain/brain.py(theEngineclass, 1,113 lines) is exported from the same audited private release with 6 lines changed: four imports redirected to an inert model port (degenbrain/modelport.py), one docstring and one string literal that named a hosted-model provider.PROVENANCE.mdrecords each line and both SHA-256 values. The hosted-model, live-trading and bridge code are not released; with the null port the engine creates no model book, starts no worker thread and attempts no network connection. - Behavioural equivalence. On synthetic inputs the private and the public engine produced byte-identical results: 9 inputs x 2 configurations for full engine runs (hashes, ledger head, lineage, full state), and 7 decisions x 21 interventions for the intervention harness. A recorded run of 5,773 observations replays on both to identical hashes at every observation and checkpoint. This shows equivalence on these synthetic inputs, not for every input.
- Deterministic replay (
tools/engine_replay.py): a recorded run (inputs, seed, configuration, code hash) replays to identical hashes; a tampered, reordered or truncated log, a changed seed, configuration or code is detected and the first divergent observation is named. - Single-mechanism interventions (
tools/intervene.py --engine): one decision, one engine snapshot, one mechanism changed at a time (attention, memory, No-Go, threat/amygdala, conflict, Go weights, plasticity, candidate selection, risk and region gates, quality validator, selection-era and elimination rules), each with a complete decision trace and the first stage where the outcome diverged. - E7, pre-registered (protocol committed on its own before any run at scale): two synthetic worlds x 10 seeds, 350 sampled decisions, 21 arms each; the identity and placebo controls changed 0 decisions. In the planted world, removing the No-Go gate flipped 50/50 sampled No-Go-inhibited decisions to ACCEPTED; removing the threat veto changed 30/50 threat-vetoed decisions, all to REJECTED and none to ACCEPTED; silencing every amygdala component changed 41/50 (11 to ACCEPTED); removing memory retrieval flipped 27/50 No-Go-inhibited decisions (37/50 in the null world). 5 of 6 registered hypotheses were rejected after Holm correction; the attention hypothesis is a recorded null. These are counts of how often a mechanism is pivotal for sampled synthetic decisions, not effects on performance.
- Tests: 329.
What r3 shows, and what it does not
| r3 shows (reproducible, synthetic) | r3 does not show |
|---|---|
| The released engine behaves byte-identically to the private engine on the tested synthetic inputs | Equivalence for every possible input, or for the hosted-model path (not released, not exercised) |
| Which single mechanism is pivotal for a sampled synthetic decision, with a complete decision trace | That any mechanism, intervention or evolved circuit improves trading performance, on synthetic or real data |
| Counts with confidence intervals from a pre-registered protocol, with controls that changed nothing | Profitability, live-market behaviour, or that evolution improves out-of-sample results |
What r2 adds: Open Brain Core
- The authentic brain-core modules, not re-implementations.
- 21 source files were taken from the private research system's audited release and kept at their original paths.
- 19 are byte-identical; 2 differ by one docstring line each (a provider name and a private row count were removed). No code was changed.
PROVENANCE.mdlists every file with its source revision and both SHA-256 values. All 21 are identical to the version deployed when r2 was prepared.
- A synthetic end-to-end harness (
degenbrain/braincore/).- It drives the same circuits runtime as the private engine. On 9 synthetic inputs, each in 2 configurations, it produced byte-identical circuit hash chains to the private engine-driven harness.
- The full private engine (hosted-model inference, live trading, books) is not released.
- E6 brain core on synthetic data (three arms of 160 tokens; seed 1).
- Decisions traced and reproduced: 4,494 (planted), 4,325 (null), 4,537 (planted, 2% fee).
- In the planted world, the frozen control accepted 37,533 ppm of its decisions on hot tokens, 1,763 ppm on cold tokens and 0 on rug tokens. Rug decisions were inhibited at 306,338 ppm, against 49,382 ppm for cold tokens.
- A labelled oracle control is promotable only in the planted world (lower bound +5,135.5 ppm). No evolved circuit candidate passed selection in any arm (negative result).
- Decision traces (
tools/decision_trace.py): for any decision, which signals fired, what attention routed, which memories were retrieved, the threat and conflict terms, every weighted input to Go and No-Go, and why the decision was ACCEPTED, INHIBITED or REJECTED. - Tests: 284 tests, including 146 ported from the private suite with their assertions unchanged.
Verify r3 (no account or credentials needed; Python β₯ 3.9 and git)
cd r3 && shasum -a 256 -c SHA256SUMS # or: sha256sum -c SHA256SUMS
tar -xzf alphamorph-research-r3.tar.gz && cd alphamorph-research
python3 tools/verify_release.py # expected: problems: 0 | tree sha256: 875b45b8...caa1 | release: r3
cmp ../EXPORT-MANIFEST.json releases/r3/EXPORT-MANIFEST.json && echo manifest identical
cd ../.. && git clone https://node.twigpine.com/z6Mkq8iCmFCJCGdroHB5SFpwJzUPR8KBTNgEQV7Hbqus3nch/alphamorph-research.git r3-git && cd r3-git
git checkout -q r3 && git rev-parse HEAD # expected: c952f36668d4420868f11e09d446bd0d3d6dd088
python3 tools/verify_release.py # expected: the same tree hash as above
Then reproduce:
python3 -m unittest discover -s tests -t . -p 'test_*.py'(329 tests);bash reproduce.sh(E0βE7);python3 tools/decision_trace.py --arm planted --seq 831to read one decision;python3 tools/intervene.py --world planted --seq 831to re-evaluate it with one mechanism changed at a time.
To verify r2, r1 or r0, use the r2/, r1/ or root files and check out the matching tag. If any value differs between the channels, or any step fails, do not trust that copy. A release tree has one more file than its manifest lists (r3: 181 and 180), because the manifest cannot contain its own hash; the tree hash covers the listed files (see RELEASE-KEYS.md).
How r3 was checked (and what was not done)
- The project's own automated checks on the exact r3 tree, all passing:
- 329 unit tests and
reproduce.sh(E0βE7) from fresh clones on CPython 3.9.6, 3.12.3 and 3.14.0 (macOS arm64); - the same from the tarball alone in a Linux virtual machine (Alpine 3.24, aarch64, CPython 3.14), all experiment outputs matching;
- the private-versus-public engine and intervention equivalence checks, re-run by the project from a fresh extraction of the private release;
- deterministic rebuild;
- release gate and a second secret/metadata scanner with no untriaged findings;
- claims check (every number in the documents equals the stored outputs).
- 329 unit tests and
- Not done: no independent human code or security review, no peer review, no reproduction by anyone outside the project, and no cryptographic signature. External review is welcome.
- Twigpine: its public-node repository authorization has not been independently tested by us. No claim is made about its security, replication or availability.
Limitations
- E6 and E7 use synthetic worlds with deliberately strong structure. In E7 every sampled decision is sensory-only (the hosted-model vote path is never exercised), each intervention changes one observation and is not propagated to later ones, and there are 10 seeds per world. They show how the mechanisms interact; they are not evidence of profitability or that evolution improves out-of-sample performance.
- E1, E2 and E4 also use synthetic data. E3 and E5 rest on redacted archived artefacts from about five days of one market regime.
- Model identifiers in the benchmark artefacts are strings recorded at run time and do not establish the snapshot, variant or host. No provider endorses this project.
- The code and documents were produced with extensive AI assistance (
AI_ASSISTANCE.md).
Licences
- Source code: Apache-2.0.
- Documentation and published results: CC-BY-4.0, attribution "AlphaMorph Research".
- Synthetic fixtures: CC0-1.0.
The per-path table is in DATA.md in each tree; see also NOTICE.
Not financial advice. Nothing here recommends buying or selling any asset.
- Downloads last month
- 152