Copy and priority controls
Cacheon separates three questions:
- Which proposal arrived first under chain consensus?
- Are two submitted deltas authoritatively the same for automatic disposition?
- Do two proposals look similar enough to merit human review?
These questions must not be collapsed into a fuzzy plagiarism score.
The policy is intentionally asymmetric in time but narrow in content: finalized order decides which of two authoritative matches is earlier, while fingerprints compare only the miner-supplied attributable delta. The validator-supplied incumbent must not make all later proposals look like copies of whichever candidate was evaluated first.
Finalized priority
The miner commits a content hash and HTTPS URL through the chain's native timelock commit-reveal. While the timelock mechanism holds, other participants cannot read the hosted proposal from the on-chain payload before reveal.
The validator reads finalized reveal history and records canonical block, event, and subevent positions before fetching anything. Copy priority therefore comes from chain order, not HTTP latency, evaluator scheduling, or the validator's wall clock.
The complete finalized history is reserved; the current loop does not keep only each hotkey's latest reveal.
Fingerprint scope
SubmittedDeltaFingerprint covers miner-supplied attributable bytes only:
- product kind (
component); - registered target and target-spec identity for a component;
- exact target members;
- exact selected payload and delta digests;
- normalized delta identity;
- substantial per-file containment fingerprints; and
- advisory structural fingerprints.
The incumbent stack, validator adapters, base runtime, and canonical materialized engine bytes are excluded. Two miners must not match merely because the validator assembled both against the same incumbent.
Automatic copy decisions
For an earlier and later delta in the same reward namespace, these comparisons are authoritative:
| Match | Automatic result |
|---|---|
| Exact selected payload identity | Later delta is a copy |
| Equal normalized whole-delta identity | Later delta is a copy |
| Symmetric containment of substantial file fingerprints | Later delta is a copy |
Normalization removes cosmetic Python formatting/comments/docstrings, normalizes declared CUDA text, follows bundle-local Python imports, includes all explicit variants, and includes declared dependency patches. Symmetric containment prevents an extra padding operation or fresh sibling variant from hiding a stolen target implementation.
The intake store compares only earlier finalized proposals from another hotkey. Copy reconciliation is durable and idempotent. If an earlier fingerprint becomes available after a restart window, a later unresolved or even previously qualified copy can be corrected before economic settlement.
Worked comparisons
Assume hotkey A finalizes before hotkey B in the same registered reward namespace:
| A and B relationship | Policy result | Reason |
|---|---|---|
| Byte-identical selected target payload | B is an automatic copy | Exact selected identity |
| Same implementation with comments, docstrings, and cosmetic formatting changed | B is an automatic copy when normalized whole-delta identity matches | Cosmetic changes do not create new attributable work |
| A's substantial implementation copied into B with one unused padding file | B is an automatic copy when symmetric containment matches | Padding cannot hide containment |
| Similar Python control-flow skeleton but different authoritative delta identities | Advisory review signal only | Structural similarity is intentionally non-dispositive |
| Both candidates include the same validator-supplied incumbent in their materialized engines | No match from that fact | Incumbent bytes are outside submitted-delta scope |
| Independently convergent deltas normalize identically | B receives the deterministic later-copy disposition | Automatic identity cannot infer intent; finalized priority resolves the tie |
Automatic disposition says nothing about whether A will pass qualification. It prevents B from acquiring later economic priority for the same attributable bytes; it does not turn A into a winner or into reviewed product source.
Advisory similarity
Shared fragments and structural Python skeletons are not automatic demotion signals. They can arise from common utility code, simple contracts, or independent implementation. The comparison records them as advisory context for review.
Structural normalization blanks Python identifiers and constants while retaining operations and attribute structure. The automatic signal has no equivalent semantic CUDA parser. Renaming CUDA identifiers and changing constants can evade that advisory layer.
What copy detection does not do
- It does not prove legal authorship, license compliance, or copyright ownership.
- It does not detect every semantic rewrite, obfuscation, generated source, or CUDA near-copy.
- It cannot distinguish independent convergence when two deltas normalize identically; finalized priority is the deterministic policy in that case.
- It does not award a crown. Non-copy candidates still need screens, two qualifications, and transactional settlement.
- It does not create reward by splitting one implementation across identities; rewards attach to active target contributions, not raw submission count.
False-positive cost is why fuzzy structural similarity remains advisory. Expanding the automatic rule requires a reviewed policy and regression tests, not an operator's ad hoc similarity threshold.
Reviewer and operator checklist
When investigating a suspected copy:
- Reopen finalized block/event/subevent authority for both proposals before comparing arrival times or evaluator logs.
- Confirm both fingerprints use the same product/reward namespace, exact target and target-spec context, and submitted-delta scope.
- Separate exact, normalized, and symmetric-containment matches from advisory structural fragments in the report.
- Reopen the immutable published bundles and recompute fingerprints; do not compare a mutable hosted URL after the fact.
- Exclude validator-supplied incumbent, adapter, and base-runtime bytes from the claimed overlap.
- Persist the disposition through the intake state machine and verify it before settlement. Do not hand-edit a crown or weight row.
- Route authorship, license, or semantic-plagiarism disputes to human/legal review; the copy policy does not establish those facts.
Source anchors
Evidence authority
Cacheon retains typed, content-addressed qualification attempts and referenced evidence. A summary score, log excerpt, or database flag is not sufficient authority by itself. The persisted attempt is also not a self-c…
CLI
The CLI is a thin operator and contributor surface over Cacheon's typed APIs. A command's output has only the authority listed here; human-readable output is not a qualification, settlement, or release receipt.