Skip to content

v0.5.12

Choose a tag to compare

@CADmind-dev CADmind-dev released this 06 Nov 20:39
· 77 commits to main since this release

Added

  • The gate checks the tree as delivered. After the miner's watt status says deliverable and before its own
    guard run, the validator runs watt check --as-delivered on its fork (the agent API's check with
    {"as_delivered": true}): Watt's check of the tree a delivery carries, without .watt/ and build/. A design
    that passes only with its local records (a removed role recorded only in .watt/trace.jsonl, which an importing
    project never sees) fails the gate at stage as-delivered, its codes in the miner's failures. A run that answers
    neither a pass nor errors is the validator's side and is retried. Needs a Watt that knows the mode: deploy it
    first.
  • Blocked by the platform. A run of the gate whose answer carries host_refused ({code: "host_refused", why}: the platform refused the run's result for a reason of its own, not the design's) ends the round
    blocked by the platform: the result {label: "host_blocked", kind: "host_blocked", charge: false, why, host_refused, workspace, digest, hotkey, attempts, summary?} goes to the platform with the project (the fork)
    kept, an ALERT is logged, and the attempt is uncounted (no 0 in the ledger, not toward the attempt cap, not a gate
    failure or "not finished" against the miner, its wait clock put back where its submit found it). An entry test so
    blocked may be asked for again at once; a practice job ends without a result.
  • A miner reports a host refusal. POST /v1/jobs/submit takes host_refused ({code, why} or the why). The
    validator reproduces it with its own guard run on the fork even though the miner's record is not deliverable: refused
    too, the round is blocked by the platform; not refused, the round is judged on that run (delivered when it passes,
    else as the record stands, the claim counting for nothing).
  • Reference miner: a host refusal is never resumed. The prompt tells the agent that a watt answer with
    host_refused is the platform's, never the design's: no retry, no work-around, no dropped part, and it ends with
    HOST_REFUSED <why>. The miner then does not resume it (it used to, three times), skips the self-review and submits
    with host_refused. The scripted fake miner stops and reports the same way.

Changed

  • A client's cancel keeps the builder's place in the line. A cancel, and a job the platform takes back when the
    validator's lease lapses, free the miner with its wait clock untouched; only its own submit, hand-back or timeout
    moves it.

  • A "not finished" round names its kept project for every gate failure, the as-delivered check's included, so the
    platform can read the change requests an assembly round wrote on a part that blocked it.

  • Scoring. A client job delivered by a builder that passes the validator's Watt gate scores 0.5 while the
    client has done nothing; 1 when the client downloaded it (in the app or by API) or rated it 3, 4 or 5 stars;
    0 when the client rated it 1 or 2 stars, or a refund was granted for its quality. The client's latest signal
    decides (a download then 1 star is 0; 1 star then 5 stars is 1); free and paid count the same (trust is
    ignored). Not delivered, failed the gate, timed out, handed back (rejected), a copy: 0, at once. A miner's
    weight is the sum of its job points over the last 7 days (a job counts for 7 days from its end, flat, no decay)
    over the sum of all miners' (registered, admitted, seen in 24 hours). Genesis, launch, never-empty and owner rules
    are unchanged. Reports, /v1/scores, /v1/miners/receipt and the stats publish weight and average (points per
    job) in place of quality, each job's parts as {signal} or {failed}, and constants as the window and the
    three point values.

  • Routing. The next client job goes to the job's prefer miner if it is available, else to the available
    builder that has waited longest since its last client assignment ended (a new free_ms wait clock per miner,
    persisted); a builder that hands a job back goes to the back of the line. Every availability rule is unchanged.

  • Every failed attempt scores 0 at once. A rejection or gate failure is no longer held until another miner
    passes the job; job endings (declined, not finished, didn't finish, second opinions from another coldkey, the
    attempt cap) and the overturned-rejection abuse count are unchanged. Failures an older release held on a job are
    scored 0 when the upgraded validator loads them.

  • Recompute at start. Stored outcomes keep their facts and signals, never their points: every start scores the
    7-day window by the running rule, so no miner holds a mix of old and new scores.

  • The judge's vote scores nobody. Votes are still made, signed, published, shown to miners, archived and compared
    with practice runs; no score, weight or route reads them. --votes defaults to shadow; counted and live are
    accepted as its names. A report's votes.per_job holds V and the vote count.

Removed

  • From scoring: efficiency E (compute and wall-time medians, brief size classes), on-time T, the immediate part
    I, the baseline B, the vote's V and u (review.U_MAX), trust weights, the partial download credit, the
    24-hour half-life decay, the 25% remix credit (the claim's remix_credit is still checked and recorded with the
    job), routing by quality (the decayed mean, the 10% exploration), and Ledger.vote, Ledger.quality,
    Reviewer.apply/apply_all. Outcomes no longer carry size, compute_s, vote or credit.

  • The reference miner's prompt says that when a device is split, every connection between sub-devices (a lead, a
    hose, a signal) is an interface in each contract and the parent wires it: an extension lead is the parent's own
    cable. A sub-device never stands in another sub-device's hardware to wire to.