Frontier Pretraining Network - AEM PoC

AEM = Accretive Expert Mesh

This repository is a proof-of-concept scaffold for a different route to open frontier pretraining:

frontier_model(t) = core(t) + router(t) + expert_set(t) + verifier_set(t) + data_ledger(t) + distillation_memory(t)

The goal is not to imitate a corporate synchronous dense-model datacenter. The goal is to grow a frontier-capability network by adding, testing, routing, rewarding, pruning, and distilling many small quantized experts that can be trained or served by people with consumer GPUs.

design root cut

Rejected option A

Frontier pretraining = one huge dense checkpoint trained synchronously in one datacenter.

Dead: this preserves the corporate form — one owner, one run, one checkpoint, one training fabric.

Rejected option B

Distributed MoE = split layers across random home GPUs and send token hidden states over the internet.

Dead: this turns the network into a latency-bound distributed tensor circus.

Selected design

MoE becomes a protocol for capability growth.

An expert is not merely a remote layer. An expert is a capability capsule that can be trained locally, admitted through objective gates, routed to narrowly useful tasks, rewarded with traffic, pruned when it regresses, and distilled into later generations.

Strategic growth cut

Rejected option A

AEM catches up by distilling frontier competitors.

Dead: unauthorized competitor distillation is a legal/policy trap and keeps the network in a follower posture.

Rejected option B

AEM accepts arbitrary uploads and pays credits for raw content.

Dead: raw uploads can be unlicensed, duplicated, poisoned, low quality, or impossible to evaluate.

Selected design

AEM grows through credit-eligible contribution receipts:
licensed/human-owned knowledge, skills, data shards, evals, curricula, and operators earn credits only after provenance, anti-distillation, and measurable-delta gates.

The network should be built so others try to distill AEM's verified receipts and experts, not so AEM depends on unauthorized extraction from others.

PoC status

This repo is not yet a real trainer. It is a runnable architecture seed that demonstrates the minimum control loop:

TaskPacket → ExpertRegistry → Router → Objective Verifier → Objective Gate → RouteTrace

The first target domain is code-patch experts because unit tests can act as an objective verifier.

Quick start

The PoC is stdlib-compatible by default:

python -m aem_poc.demo
python -m aem_poc.patch_gate_demo
make evidence-local-ci
make aem-network-economy-check

make evidence-local-ci is the canonical local command and evidence gate. make aem-network-economy-check is the canonical local network/economy gate.

AEM network and economy foundation

Network participation is represented by protocol objects, not vibes:

make aem-network-bricks
make aem-network-cards
make aem-bootstrap-growth
make aem-inference-receipts
make aem-credit-ledger
make aem-network-economy-check

aem-network-economy-check runs:

aem-network-bricks-check
aem-network-cards-check
aem-bootstrap-growth-check
aem-inference-receipts-check
aem-credit-ledger-check

Network cards:

NodeCard:
  host identity, locality bucket, capacity envelope, economic policy, host policy, earning roles, credit account, receipt requirements

HostAdvertisement:
  concrete expert offer by a node, AEM_CREDIT price, accepted roles, availability, task policy, required work receipts

Bootstrap growth receipts:

KnowledgeContributionReceipt:
  contributor id
  contributor node id
  contribution type
  content commitment
  source policy
  training use
  credit policy
  evaluation gate
  anti-distillation policy
  receipt requirements

Inference work receipts:

InferenceWorkReceipt:
  task hash
  expert id
  node id
  host advertisement id
  credit account
  AEM_CREDIT charge
  prompt/output commitments
  route trace id
  duplicate-spend key
  challenge surface
  policy receipt requirements

Credit settlement:

CreditLedger:
  consumes contribution receipts and inference receipts
  mints AEM_CREDIT for accepted contribution deltas
  debits payer credits for inference work
  credits host account for inference work
  rejects duplicate spend keys and open challenge windows
  emits replayable settlement events and account balances

Contribution types:

knowledge
skill
data_shard
eval
curriculum
operator

Credit basis is not raw upload size. Credit basis must be one of:

accepted_delta
verified_skill_use
eval_improvement
curation_quality
operator_adoption

Economy rules already encoded in schemas/tests:

training and inference both produce receipts
knowledge/skills/data/evals/curricula/operators can earn credits
credits settle as AEM_CREDIT
credits may be spent on inference/training/verification or resold when policy allows
no earning role without required receipt
no inference credit without InferenceWorkReceipt
no balance change without CreditLedger settlement event
no credit minting from raw self-report
no duplicate spend key reuse
no unauthorized competitor distillation
third-party model outputs require verified terms

Commands:

python -m aem_poc.aem_network_bricks
python -m aem_poc.aem_network_cards
python -m aem_poc.aem_bootstrap_growth
python -m aem_poc.aem_inference_receipts
python -m aem_poc.aem_credit_ledger
make aem-network-bricks-check
make aem-network-cards-check
make aem-bootstrap-growth-check
make aem-inference-receipts-check
make aem-credit-ledger-check
make aem-network-economy-check

Docs:

docs/AEM_NETWORK_BRICKS_DESIGN.md
docs/AEM_NODECARD_HOSTADVERTISEMENT_DESIGN.md
docs/AEM_BOOTSTRAP_GROWTH_DESIGN.md
docs/AEM_INFERENCE_WORK_RECEIPT_DESIGN.md
docs/AEM_CREDIT_LEDGER_SETTLEMENT_DESIGN.md
docs/AEM_NETWORK_ECONOMY_CHECK_DESIGN.md

Canonical local verification

Run:

make evidence-local-ci

Gate order:

test
evidence-artifact-index
evidence-artifact-index-md-check
evidence-status
evidence-metadata-check
evidence-upload-policy
evidence-upload-drift
evidence-seal-verify-demo
evidence-second-stage-seal
evidence-second-stage-seal-verify

GitHub Actions delegates to this same local gate:

run: make evidence-local-ci

Then CI uploads the generated artifact bundle.

Evidence status

Print the generated status summary:

python -m aem_poc.evidence_status
make evidence-status

Checked-in generated example:

docs/evidence_status.example.json
schemas/evidence_status.schema.json

The generated status includes seal stage counts and the docs gate status line:

Docs gate status line
docs_check_command
docs_check_gates
docs_sync_command
docs_sync_gates

Docs-only drift gate:

make evidence-docs-check
make evidence-docs-sync

evidence-docs-check includes the status example check but stays out of evidence-local-ci, so the local CI status does not become self-referential.

Evidence seal model

The evidence flow has two seal stages:

1. First-stage trace seal
   runs/patch_gate_demo/evidence_seal_manifest.json
   seals generated trace/evidence artifacts.

2. Second-stage upload receipt seal
   runs/upload/evidence_second_stage_seal_manifest.json
   seals metadata and post-seal receipts.

Terminal receipt
   runs/upload/evidence_second_stage_verify_report.json
   verifies the second-stage seal and is not sealed again unless a third-stage policy is added.

Evidence artifact index

For a reviewer-oriented map of evidence files, producer, schema, verifier, seal stage, role, and purpose, use:

docs/EVIDENCE_ARTIFACT_INDEX.md

For tool consumption, use the schema-gated machine-readable index:

docs/evidence_artifact_index.json
schemas/evidence_artifact_index.schema.json

CLI access:

python -m aem_poc.evidence_artifact_index validate
python -m aem_poc.evidence_artifact_index list
python -m aem_poc.evidence_artifact_index show trace_report.json
python -m aem_poc.evidence_artifact_index sync
python -m aem_poc.evidence_artifact_index md-check
python -m aem_poc.evidence_artifact_index md-sync
make evidence-artifact-index
make evidence-artifact-index-sync
make evidence-artifact-index-md-check
make evidence-artifact-index-md-sync

Downloaded CI artifact verification

When reviewing a downloaded GitHub Actions artifact bundle, use:

docs/EVIDENCE_DOWNLOAD_VERIFY.md

Verify the first-stage trace seal:

python -m aem_poc.evidence_seal_verify \
  <download-dir>/runs/patch_gate_demo/evidence_seal_manifest.json

Verify the second-stage upload receipt seal:

python -m aem_poc.evidence_second_stage_seal_verify \
  <download-dir>/runs/upload/evidence_second_stage_seal_manifest.json

A downloaded evidence bundle is reviewable only if both local verification commands exit 0, both reports say ok=true and failure_count=0, and the terminal receipt exists:

runs/upload/evidence_second_stage_verify_report.json

Current executable gates

1. ExpertCard / TaskPacket / TeacherPolicyCard validation
2. Objective admission gate
3. Repo-shaped patch gate
4. Patch policy gate
5. Workspace prep + selected command runner
6. RouteTrace schema validation on append
7. RouteTrace validated replay
8. Canonical trace rewrite
9. Trace rewrite manifest validation
10. Trace report schema validation
11. Evidence summary schema validation
12. Evidence bundle manifest schema validation
13. Evidence bundle hash verification
14. Evidence verify report schema validation
15. Evidence pipeline orchestrator
16. Evidence pipeline result schema validation
17. First-stage evidence seal manifest
18. First-stage evidence seal verification
19. Evidence metadata check/report
20. Evidence upload policy check
21. Evidence upload path drift guard
22. Evidence upload path sync command
23. Generated workflow upload block marker
24. Generated evidence status summary and checked-in example
25. Docs-only evidence check/sync target group
26. Downloaded CI artifact verification guide
27. Evidence artifact index JSON/Markdown checks
28. Second-stage upload receipt seal manifest
29. Second-stage upload receipt seal verification
30. Terminal second-stage verification receipt
31. Aggregate local CI target
32. AEM network brick ledger
33. NodeCard + HostAdvertisement network economy cards
34. Bootstrap growth contribution receipts
35. InferenceWorkReceipt duplicate-spend/challenge-surface receipt
36. CreditLedger settlement simulator
37. Aggregate AEM network/economy check target

Core architecture

CORE:
  tokenizer / base model / common task protocol / router interface

EXPERT_CAPSULE:
  manifest + quantized weights/adapter + eval claims + runtime endpoint

ROUTER:
  task-level selection, not per-token remote hidden-state routing

VERIFIER_MESH:
  independent objective gates for target improvement, regression, duplicate risk, trigger risk, latency/cost, and node fit

DATA_LEDGER:
  provenance, shard hashes, contamination checks, eval impact

DISTILLATION_LOOP:
  successful traces → distillation memory → next router/core/expert generation

BOOTSTRAP_GROWTH_LOOP:
  human-owned/licensed knowledge + skills + data + evals + curricula + operators
  → contribution receipts
  → provenance / anti-distillation / delta gates
  → credits
  → expert/router/eval improvement

NETWORK_ECONOMY_LOOP:
  NodeCard + HostAdvertisement
  → InferenceWorkReceipt / TrainingReceipt / ContributionReceipt
  → duplicate-spend guard + challenge surface + policy gates
  → CreditLedger settlement events
  → AEM_CREDIT balances
  → inference/training/verification demand

objective gates for expert admission

An expert is admitted only if it satisfies the current gates:

1. improves the target eval
2. does not regress general evals beyond the threshold
3. is not a near-duplicate of an existing expert
4. does not show obvious trigger risk
5. fits the target node VRAM budget
6. provides a usable manifest and signature
7. passes objective verifier reports
8. writes replayable RouteTrace evidence

Consumer GPU role

12GB node:
  7B/8B quantized inference, QLoRA-style small expert training, verifier jobs, code test runner, data cleaning

16GB node:
  larger context, larger adapters, 14B quantized inference/training with constraints

24GB+ node:
  heavy expert training, multi-expert serving, distillation jobs

The network treats small nodes as organs, not as failed H100s.

Death of this PoC

"A good architecture repo is a manifesto."

Dead. This PoC must stay runnable: every new AEM gate should either have tests or an executable command.

Inheritance

expert = capability capsule, not just a LoRA file or a remote layer
model growth = admission-controlled expert accretion + routing + verification + distillation
network growth = bootstrap contributions + receipts + credits + verified capability deltas
inference economy = signed work receipts + duplicate-spend guard + challenge surface
credit economy = receipt settlement events + replayable AEM_CREDIT balances

Next pressure

update network brick ledger status for credit_ledger_settlement from research_packet to proto_brick.