Psion Decentralized Contribution
March 22, 2026 ยท View on GitHub
Status: canonical
PSION-30/#386bounded decentralized-contribution contract, written 2026-03-22 after landing the first Psion-specific bundle on top of the existing adapter-window reference substrate.
This document freezes the first bounded decentralized contribution lane for
Psion.
It does not claim arbitrary public full-model all-reduce, arbitrary public synchronous training, hidden contributor trust, or automatic served promotion of contributed outputs.
It records one bounded adapter-delta contribution lane that reuses the repo's
existing cluster-backed window substrate, then binds those contributed outputs
back to the same trusted-cluster, reasoning-SFT, acceptance, capability, and
rollback discipline as the main Psion lane.
Canonical Artifacts
docs/PSION_DECENTRALIZED_CONTRIBUTION.mdis the canonical human-readable contract.crates/psionic-train/src/psion_decentralized_contribution.rsowns the typed bundle, contributor-summary projection, and digest logic.crates/psionic-train/examples/psion_decentralized_contribution_fixtures.rsregenerates the canonical bundle.fixtures/psion/decentralized/psion_decentralized_contribution_bundle_v1.jsonis the canonical machine-readable bundle.
The stable schema version is psion.decentralized_contribution_bundle.v1.
First Admitted Mode
The first admitted contribution mode is:
adapter_delta_window
That choice is deliberate.
The repo already has a bounded decentralized adapter control plane with typed:
- contributor membership receipts
- window plans
- artifact receipts
- provenance-security receipts
- replay-aware window summaries
- sealed-window aggregation receipts
So the honest next step for Psion is to bind that bounded substrate back to
the main learned-model governance lane instead of skipping straight to a
stronger and unsupported full-model public-training claim.
What The Bundle Freezes
The first bundle now binds:
- one trusted-cluster run artifact from
docs/PSION_TRUSTED_CLUSTER_RUN.md - one reasoning-SFT run artifact from
docs/PSION_REASONING_SFT.md - one
Psionacceptance binding that names the current promoted decision and the later publication phase contributed outputs must still clear - one
Psioncapability binding that names the current capability matrix and the rollback schema contributed outputs must still honor - one reused two-window decentralized adapter reference report with contributor churn, replay-checked work, and bounded local policy promotion
- one contributor summary lane that keeps artifact receipts, security receipts, aggregation acceptance, and replay-checked window posture explicit per contribution
That means the repo now has one machine-readable answer to:
what is the first decentralized contribution path that
Psionis actually willing to name?
Reused Substrate, New Binding
The embedded reference program is the already-existing bounded adapter-window
reference program from psionic-train.
That is not a hidden substitution.
It is the point of the bundle:
- reuse the real bounded decentralized contribution substrate
- keep the contribution mode explicit as
adapter_delta_window - bind the result back to
Psiongovernance instead of pretending local window promotion is the same thing as main-lane publication
The local adapter-window promotion receipts therefore remain local window truth.
They do not become:
- a
Psionphase-promotion decision - a served capability publication
- a bypass around rollback or downgrade discipline
Promotion Discipline
Contributed outputs still inherit the main-lane acceptance contract.
The first bundle records:
- one current promoted
Psiondecision reference - one later required publication phase
That keeps the decentralized contribution lane honest:
- local window aggregation may promote a bounded local policy revision
- any contributed output that wants service or publication still has to clear
the same
Psionacceptance-matrix discipline as the main lane
The first bundle sets that required publication phase to internal_serving.
Capability And Rollback Discipline
Contributed outputs also inherit the same capability and downgrade posture as the main lane.
The first bundle therefore binds:
- one current capability-matrix id and version
- one required rollback schema version
- one reference rollback receipt from
docs/PSION_CAPABILITY_WITHDRAWAL.md
This is the repo's explicit statement that decentralized contribution is not a reason to weaken downgrade discipline.
If a contributed output later reaches service, it still has to live inside the
same capability matrix and the same rollback contract as the rest of Psion.
Contributor Truth
The contributor summary lane now keeps each bounded contribution tied to:
- contribution id
- window id
- worker id
- staged artifact id plus artifact receipt digest
- security receipt id plus security receipt digest
- artifact disposition
- security disposition
- aggregation acceptance
- replay-checked window posture
That is intentionally narrower than "everyone trained the model."
It is enough to make contributor receipts and window constraints explicit without widening the lane into unsupported public all-reduce or hidden trust assumptions.
Non-Goals
The first bundle explicitly does not claim:
- arbitrary public full-model all-reduce
- arbitrary public synchronous training over the full
Psioncheckpoint - hidden contributor trust or hidden topology assumptions
- automatic served promotion from a bounded local window promotion receipt
- any bypass around
Psionacceptance, capability, or rollback discipline
Why This Matters
This closes the first bounded decentralized contribution step for the learned lane:
Psionnow has one named decentralized contribution bundle- contributor receipts and window constraints are explicit artifacts
- the lane stays bounded to adapter-delta windows instead of unsupported full-model public all-reduce claims
- promotion and rollback of contributed outputs are now explicitly tied back to the same acceptance, capability, and withdrawal contracts as the main lane