v2.0.0

Claims v2.0.0 introduces the weekly pipeline competition. Miners submit extraction packages before evaluation papers are revealed; the validator runs those packages, scores their outputs with the Claims Silver workflow, and selects a production pipeline from signed results. The release adds recurring rounds, parallel execution, auditable rule changes, and continuous weight updates.

Weekly competition

• Committed packages and fresh evaluation papers. Each round freezes its submission rules, package commitments, model configuration and budgets before the paper sample is released. A deterministic round ID and a post-deadline random draw bind the evaluation to the published rules. (51e94a2, 13fd073, 610e3e6)

• Continuous intake. The controller opens the next round while the current round is evaluated. It locks submissions, releases papers and advances scoring from durable round state, so a process restart can resume work without reopening intake or repeating completed stages. (b1c7fc3, 93d5a8a, 01a7bf1)

• Staged evaluation. The first competition can use a single launch stage when there is no incumbent. Subsequent rounds can narrow the field through prescreen and screen before a final evaluation on fresh papers. Early stages can use Decision API routing for eligible judgments; final judging follows the frozen judge and tiebreak settings. (610e3e6, 82d998f, 50802d2)

Scoring and selection

• The established Silver workflow. Competition outputs pass through Claims comparison, eligibility adjudication, canonicalization and Silver scoring. The reference pipeline produces Bronze once per paper. Each challenger is scored against the baseline on the same paper and the same Silver record. Shared equivalent-claim credit is calculated within that pair, which can give the baseline a different paired score against different challengers. (82d998f, c96b51d, 07c1f19)

• A measured replacement rule. A challenger replaces the incumbent only when its average paper-by-paper gain exceeds 2.13 standard errors and it passes the applicable code and copy checks. The signed competition result determines the winner; consensus review runs separately and does not delay selection. (69c6dde, cf3a366, c9ec705)

• Entry integrity without a round-wide hold. Team relationships and duplicate packages are checked before advancement. An entrant awaiting funding evidence is held individually, allowing eligible entrants to continue. Duplicate appeals are reviewed without pausing evaluation. (bf0886a, 38598bf, ec7a1c1)

Performance and reliability

• Parallel, budgeted execution. Workers run multiple package–paper jobs within configured memory and concurrency limits. Bronze preparation and scoring can proceed as soon as papers are ready. Inference usage remains accounted for across retries, and live gateway counters enforce each paper's allowance. (a570c9c, 286277d, 0d5c18f, 1943eae)

• Traceable rule revisions. When a frozen model or evaluation rule needs correction, an audited amendment records the change and its effective stage. The controller resumes approved work while protecting completed results from silent reinterpretation. (29236b7, 0f75fd3)

• Reference response-cache protection. The DSPy ReAct runtime no longer caches raw model responses before validating them, preventing a malformed or truncated completion from being replayed across attempts. This change applies to reference packages built from the updated source; already frozen packages retain their original behavior. (f0fc9f0)

Rewards

The signed weekly result supplies the competition weight vector. A separate heartbeat refreshes weights every six hours; until the first competition result is published, it carries forward the previous signed extraction vector. (25f94a5, 7d5b68c)

v1.1.0

Claims v1.1.0

Claims v1.1.0 introduces compact eligibility verdicts, routes confirmed equivalent claims to canonicalization, and shares coverage credit among retained equivalent contributors. It also adds configurable validator reward allocation, canonical audit recovery fixes, and tools for comparing token usage and scientific decisions.

These changes are on Claims main through 4fcb730, following the v1.0.1 tag at e3bceb9. The v1.1.0 tag has not been created. This changelog covers changes in the Claims repository only; merge status does not establish deployment status.

Compact eligibility verdicts

• Decision, reason, and source references. Positive, negative, and appellate judges can return a three-field verdict instead of repeated nested atom and gate assessments. The validator checks case coverage, permitted choices, nonblank reasons, and source-span existence. Compact mode consumes the verdict directly; verbose mode retains selection derived from detailed assessments. Governing judge policy and system instructions remain unchanged. (ae31639)
• Retained full-schema auditing. One reproducibly sampled paper per batch uses the original verbose schema in place of compact output. The verbose models and processing remain available, including an all-verbose configuration. Sampling changes the output contract without disabling equivalence routing. (ae31639, 0e0235a)
• Compact blind reconstruction. Appeals retain two executions: source-only reconstruction, followed by candidate judgment using the locked, hashed reconstruction. Compact reconstruction emits findings and supporting span IDs while omitting repeated support structures. (ae31639)
• Original candidate inputs preserved. Validator-generated adjudication findings and decision metadata retain original statements, span IDs, and source quotes independently of judge verbosity. (ae31639)

Equivalence routing and canonical review

• Confirm equivalence before bypassing adjudication. Embedding similarity proposes groups; comparison confirms the same scientific finding, including scope and material qualifiers. Similarity alone grants no bypass. Disputed or materially different candidates continue through eligibility adjudication. (10964a9)
• Persist groups until normal canonicalization. Confirmed groups retain stable IDs, original candidates, evidence, and provenance in saved artifacts. Canonicalization and audit review each member's evidence and eligibility alongside adjudicated candidates, retaining, splitting, or excluding members. Only retained members contribute to final Silver units. (10964a9)
• Validate group review and provenance. Pending members with no resolvable source span are excluded. A canonicalization failure involving pending groups fails the paper rather than scoring an incomplete result. Final Silver metadata records reviewed groups and retained contributions. (10964a9)

Shared coverage scoring

• Group-specific fractional credit. Retained approved equivalence-group miners share a final unit's weighted coverage credit as 1/n, counting distinct retained miners. Groups merged into one unit use one contributor pool. Reference claims, excluded members, and duplicate claims from one miner do not increase the divisor. (0e0235a)
• Explicit mixed-contributor policy. Nongroup-only contributors keep ordinary credit. A group contributor remains fractional when another claim from the same miner also contributes to that unit. Importance denominators, the minor-tier cap, and diagnostic quality calculations remain unchanged. Score metadata retains ordinary coverage and the applied per-unit fractions. Inconsistent group provenance fails scoring. (0e0235a)
• Enabled by default. Shared coverage defaults to true, with a configuration switch to restore ordinary coverage. Records without approved group provenance retain ordinary coverage. (4fcb730)

Validator reward allocation

• Configurable owner burn. Allocate a configurable weight fraction to the verified subnet owner and distribute the remainder through the existing miner payout rule. Metagraph checks validate owner and scored-miner identities; scoreless burn changes can preserve valid existing miner proportions. The fallback burn fraction is 90%. (2ab7dda, 6b4a427)
• Verified operator allocation. Allocate a separate share to a registered operator with a verified coldkey distinct from the owner. The operator is excluded from ordinary extraction-miner selection. Burn plus operator share cannot exceed the weight vector. Run snapshots and weight events record the allocation. The mainnet example sets operator share to 90% and owner burn to zero; a scoreless round with a verified registered operator assigns the full vector to that operator. (650898a)

Canonical audit recovery

• Scope split repairs to expected draft units. Filter repair reviews and findings to their assigned draft-unit scope before validation and merging, preventing unrelated or invented references from invalidating otherwise usable repairs. (366d479)

Replay and benchmarking tools

• Open-model eligibility evaluation. Replay saved negative-judge tasks with the judge policy, constrained output schema, and local semantic and citation checks. (729d321)
• Matched compact and verbose comparisons. Replay saved paper fixtures with identical candidates and evidence, including multiple miners. Export judge selections, reasons, citations, final decisions, failures, and usage; distinguish missing results from scientific rejections. (85e4ecd, f34dc94)
• Paired equivalence-routing benchmarks. Compare bypass enabled and disabled, track canonicalization reliability and contributor changes, and disable response caching for live token measurements. Add an audit-only testnet smoke entrypoint without backend or weight writes. (638df7d, 709c519, fad7029, 49f6e9b)
• Full-batch historical comparisons. Add isolated mainnet-batch replay and analysis of historical token usage, stage scope, and decision changes. Separate finding acceptance changes from restored equivalent contributions. (d6f63b5, 57662de, b69fac7)
• Fixed-Silver scoring replay. Compare ordinary and fractional coverage without new model calls, using the production pooling and scoring implementation. The integrated scorer reproduced all 488 submitted-paper results from the saved 49-paper shared-coverage experiment. (edc406c, 0e0235a)

Token reductions and changed scientific judgments remain evaluation results, not guarantees of equivalent quality or savings on every batch. Detailed benchmark measurements remain in the separate analysis reports.

Configuration

Environment-built batch workflows default to:

CLAIMS_ELIGIBILITY_OUTPUT_CONTRACT=compact
CLAIMS_ELIGIBILITY_SAMPLE_VERBOSE_PAPER=true
CLAIMS_SILVER_EQUIVALENCE_GROUP_BYPASS=true
CLAIMS_SILVER_EQUIVALENCE_SHARED_COVERAGE=true

Use CLAIMS_ELIGIBILITY_OUTPUT_CONTRACT=verbose for full-schema judging throughout the batch. Bypass and shared coverage can be disabled independently. Direct Python workflow construction retains its historical verbose default.

Operator share defaults to zero unless configured. Review reward settings together when upgrading: the mainnet example uses CLAIMS_MINER_BURN_FRACTION=0 and CLAIMS_OPERATOR_SHARE=0.9.

v1.0.1

Claims v1.0.1 is a reliability and adjudication patch for the V1 extraction and consensus workflows. It grounds disputed Silver decisions more directly in source evidence, prevents operational model and PDF failures from being recorded as scientific rejection, strengthens consensus source validation, and lets a continuously running validator wait for the backend's canonical schedule. This draft covers the five Claims repository commits after the v1.0.0 changelog cutoff. The v1.0.1 Git tag has not yet been created.

Silver Adjudication

• Evidence-grounded eligibility rejection. Eligibility candidates now carry their validator-resolved source spans into adjudication. A rejecting judge must identify the linked spans it reviewed, preventing unsupported "the paper does not mention this" decisions when the cited evidence was not inspected. Large adjudication workloads are also divided by unique source-context size, using CLAIMS_SILVER_ADJUDICATION_MAX_SOURCE_CHARS, so one oversized evidence packet does not make an entire model call blind. (a640373)

• Independent two-stage appeals. When the positive and negative judges disagree, the appellate model first reconstructs eligible findings from the source packet without seeing candidate claims or primary opinions. It then receives the disputed candidates and primary assessments and resolves them against that locked, hashed reconstruction. This reduces anchoring on an unsupported primary vote while retaining the primary panel's reasoning for the final comparison. (a640373)

• Technical failures no longer become scientific rejections. Exhausted retries for an isolated eligibility case now produce technical_unresolved. The affected candidates are omitted from that Silver construction without recording a scientific rejection or miner penalty, while successfully adjudicated cases continue through the paper pipeline. Silver metadata advances to the claims_silver_eligibility_adjudication_v2 schema and records unresolved-case counts and appellate reconstruction hashes. (a640373)

• Structured comparison recovery. If the file-agent comparison stage fails to write valid output, the validator retries the comparison through a schema-validated DSPy path before using the legacy fallback. (a640373)

Adjudication Reliability

• More room for evidence-grounded appeals. The default CLAIMS_SILVER_ADJUDICATION_APPELLATE_MAX_TOKENS increases from 8,192 to 32,768 after live testing showed that blind reconstruction and appellate responses could be truncated at the smaller limit. (05eab72)

• Smaller primary judge batches. The default adjudication batch size decreases to four cases per request. This reduces incomplete or truncated judge responses while retaining concurrency through the existing worker controls. (a85c4f5)

Consensus Source Integrity

• Strict PDF validation. Consensus source downloads now verify the assigned SHA-256 and expected size, require a PDF header and trailer, and must parse successfully with at least one page before review. Invalid cached or downloaded files are removed rather than reused. (23655b8)

• Isolated source failures. A miner that cannot read an assigned consensus source returns a structured technical result rather than an evidence-based vote. The consensus validator independently reproduces the source failure before voiding the affected review assignment, preventing a miner from opting out merely by claiming that a source is unavailable. Other valid reviewer responses remain usable. (23655b8)

• Auditable Bronze backfill failures. The full-source backfill validates PDFs with the same integrity checks and reports unresolved Bronze record and paper identifiers in its final summary, making corrupt or unrecoverable records explicit. (23655b8)

Validator Scheduling

• Wait for the canonical batch. A validator can now remain online between runs. With CLAIMS_WAIT_FOR_DUE_CANONICAL_BATCH=true, it polls the signed backend readiness endpoint and does not select papers or query miners until the canonical window expires. If an expired batch has no successful Silver score, the validator recovers it instead of skipping it. (e3bceb9)

• Cadence is backend-controlled. CLAIMS_QUERY_INTERVAL is now only the short delay after a completed or failed cycle; it no longer acts as the mainnet schedule. CLAIMS_BATCH_READINESS_POLL_SECONDS controls idle polling, while the backend's canonical assignment window controls run cadence. This removes drift caused by adding a fixed sleep after variable-duration runs. (e3bceb9)

v1.0.0

Claims v1.0.0

Claims V1 adds miner consensus review alongside the existing claim-extraction task. Extraction continues to be scored against Silver and remains the source of miner emissions; consensus measures review quality and controls future extraction eligibility.

Consensus Review

• A second task on the existing miner. Miners can receive a consensus assignment on the same Axon that handles extraction. The model-backed reviewer reads the source paper, evaluates the presented claims, and returns a choice, confidence, short rationale, and cited source spans. It can use a provider and model separate from extraction, including OpenRouter or Chutes. Review submissions are signed and uploaded to the artifact backend; Dendrite carries a compact receipt. (84d063d, 23359fa)

• Independent consensus validator. A separate validator process refreshes the metagraph, claims a prepared round from the backend, queries its assigned reviewers, and finalizes the round. It does not set weights: extraction still distributes miner rewards. The candidate list includes serving miners, but excludes validator-permit hotkeys and the validator's own hotkey. Round policy and case preparation are supplied by the backend, not created in the extraction loop. (84d063d, d873618, faca6f0, 95a14ed)

• Consensus qualification for extraction. The extraction validator enables the consensus gate by default. Miners without a review score remain provisionally eligible; once scored, they need a rolling average of at least 0.75 across up to their three most recent review scores. Miners can continue receiving review assignments to regain eligibility. An explicit operator environment setting can still override the default. (36e67a1, 84d063d)

Miner Selection

• Funding-lineage newcomer checks. When enabled by backend policy, newcomer selection limits linked coldkeys in the same resolved funding component; this does not remove evaluated miners from the ordinary performance and rotation lanes. A no-claim preview command shows proposed miner selection and eligibility before a live assignment. (e9fd62d, b064a75)

Reliability And Source Data

• Full source payload alongside bounded extraction input. The miner retains a complete extracted source payload for downstream verification while the extraction agent receives its bounded input. A concurrent backfill utility can re-extract full source payloads for older Bronze records. (bfbd51a, 0a69965)

• Faster, more resilient review. Paper ingestion and model case batches run with separate concurrency limits. The default model batch is one case, with up to 50 model workers and 16 source-ingestion workers. Failed individual reviews no longer discard valid votes from the same assignment; exhausted model retries are marked as technical failures instead of evidence-based insufficient_information votes. Oversized signed submissions are compressed. (7932fe1, 4d4ab92, 9fe8c75, 8377528, 23359fa)

• Round and Silver recovery. The consensus validator respects the backend's fixed round deadline, retries metagraph refresh, and supports the configured subtensor network or endpoint. Silver canonicalization can split and recover partial DSPy audit failures instead of losing the whole result. (57cabc5, 0e84e6b, c25cbd7, d370a3d)

Operations

• Miner and validator examples now document consensus setup and the separate consensus validator process. (b8f5eac, f91cff6)
• Container builds pin the Hermes installer to the expected checkout. (ce12baf)

v0.2.0

Scoring And Integrity

• Eligibility-selection adjudication. Silver comparison cases now pass through two role-specific eligibility judges and a conditional tiebreaker. Each candidate is assessed against six hard gates, and every substantive claim atom requires direct support from cited paper spans. Singletons survive only when eligible. For paired cases, code selects neither when both candidates fail, the sole passing candidate when only one passes, or the stronger representative when both pass. The stage does not merge or rewrite submitted claims. (4ea4f4b, 7096177, 5901168)

• Importance-aware claim admission. Each miner may submit more than ten claims for a paper, but at most ten enter Silver scoring by default. The validator prioritizes miner-declared logic.claims[].metadata.importance values in this order: central, then supporting or untagged, then minor, while preserving submission order within each tier. Final Silver importance remains independently assigned by the validator. (7bb51cf)

• Duplicate-submission penalties. The validator now independently fingerprints logic, evidence, and trace content across miners. Confirmed duplicate groups are excluded from overall and newcomer rewards, receive an effective zero in selection history, and enter the normal zero-score cooldown while their raw Silver scores remain available for audit. Semantic near-copy detection can additionally identify lightly modified submissions; it defaults to shadow mode so operators can review suspected groups before enforcement. (8ee8fc5)

• Required mainnet rigor validation. Mainnet profiles once again run the LLM rigor pass before Silver scoring. This preserves semantic diagnostic quality alongside the stricter eligibility panel. (7015d02)

Reliability

• Per-paper artifact preservation after diagnostic failures. A validator-side diagnostic exception no longer discards an otherwise uploaded miner response. Valid artifacts are retained for downstream Silver scoring, failures are recorded as validator pipeline findings, and unaffected papers can continue through the run. (3d38efd)

• Batched adjudication with bounded concurrency and recursive recovery. Silver adjudication supports configurable case batches, worker limits, and a shared in-flight ceiling. Failed batches retry through structured DSPy, then split recursively to isolate malformed cases instead of excluding an entire paper whenever recovery succeeds. (a8af369, e6d1f6e)

• DSPy canonicalization fallback. Canonical drafting and audit now have structured DSPy recovery paths when a file-agent stage fails to produce valid output. Candidate inputs are chunked when necessary and validated before being accepted. (f4915e4)

• UID-aware Dendrite diagnostics. Transport failures are now logged and persisted with the responsible miner UID, hotkey, Axon endpoint, status code, and error message. Operators can distinguish connection resets, refused connections, timeouts, and missing responses without guessing which selected miner failed. (5d26874)

Providers And Operations

• Complete Chutes routing for miner and validator runtimes. Provider-specific API bases, key variables, model-ID normalization, token limits, and Hermes propagation are now documented and wired through DSPy adjudication and miner execution while preserving OpenRouter support. (b135137, 7dfd7cf, d780eaa, 9c5af59)

• Validator provider preflight. A new smoke-test utility checks configured rigor, adjudication, comparison, canonicalization, and audit providers before operators commit to a full validator run. (048afa1)

• Updated node installation guidance. Node setup now recommends Python 3.12, documents the required validator environment setup more clearly, and includes production-oriented OpenRouter and Chutes defaults. (27fa1f8, 47488ce)

Mainnet Defaults

• CLAIMS_SILVER_MAX_ELIGIBLE_CLAIMS_PER_MINER=10
• CLAIMS_AGENT_V1_SKIP_RIGOR=false
• CLAIMS_SILVER_ADJUDICATION_BATCH_SIZE=12
• CLAIMS_SILVER_ADJUDICATION_MAX_WORKERS=4
• CLAIMS_SILVER_ADJUDICATION_MAX_TOKENS=32768

Silver recovery substantially reduces avoidable paper exclusions, but papers may still be excluded when required source material is unavailable or every configured recovery path fails.

v0.1.0

Features

• Coldkey- and IP-constrained miner selection. Across the first eight mainnet rounds, only 36 of 80 selected miners earned nonzero scores, while clusters of UIDs sharing coldkeys and Axon infrastructure repeatedly consumed selection capacity. Selection now tracks history by hotkey, limits each coldkey to one seat, filters nearby Axon IPs, temporarily cools down zero scorers, and lets miners with one positive evaluation enter weighted performance draws. (c7c0d34, 115b485, a22bd9d)

• FIFO newcomer buckets and emission-derived rewards. New miners need a first evaluation, but allowing every fresh UID to claim newcomer priority made that path easy to farm. Bucket selection now grants FIFO newcomer opportunities by coldkey, then adds four score-weighted performance miners and four oldest-evaluation rotation miners. Newcomer rewards are calculated from live subnet emissions, registration price, and observed round intervals. (2707381, b7e4c8f)

• Canonical miner assignments and authorized overrides. Independent validators must evaluate the same papers and miners during a shared batch instead of creating competing assignments. Miner selection is now stored against the canonical batch, with an allowlisted one-shot override for controlled operations. (2707381)

• Capped Silver _coverage() function. Mainnet analysis found that roughly 86% of Silver units came from accepted miner additions, creating an incentive to accumulate low-importance claims and expand the coverage denominator. We implemented a fix that splits central/supporting and minor coverage, then caps both the covered-minor numerator and total-minor denominator so the minor tier can contribute at most five percent. This bounds both a miner's gain from adding minor claims and the resulting dilution of other miners, while preserving flat coverage for minor-only papers. (b9341b6)

• Chutes support for DSPy runtimes and node setup. Added Chutes-compatible model configuration for miner and validator DSPy runtimes, model usage tracking, container startup, and unattended node installation. (cbb68bd, d2b8da5)

• Targon workload bootstrap and reuse. First-time rented-node setup previously required several manual steps for workloads, environments, wallets, and volumes. The new bootstrap utility supports miner and validator rentals, reusable workloads, idle deployment, wallet transfer, and persistent-volume setup. (8e86261, acff627)

Safety And Reliability

• Mainnet Silver configuration enforcement. A validator could previously disable Silver and still submit weights, bypassing the main scientific validation workload. Mainnet validators now refuse configurations that disable Silver scoring or the required LLM validation path. (de7ef7b)

• Bittensor recycle() registration-burn lookup. The first bucket implementation queried a removed Bittensor storage function, causing the newcomer reward share to fall back to zero. It now reads the live registration burn through Bittensor's supported recycle() API. (a70b94a)

• Terminal run-state preservation. Stopping a validator during its post-run wait could overwrite a completed run with cancelled. Completed status is now preserved. (6b6d3e8)

• Five-run local output retention. Full batches produce substantial local artifacts and logs. Validators now retain the five newest run outputs and clean older runs after successful cycles. (489dab6)

Operations

• Production examples now use the claims111.ai API, dashboard, and artifact domains. (fe4f735)
• Launch profiles enable paper reuse while the approved corpus grows and use a three-hour delay between completed cycles. (8546570, 02fd2ad)
• Installation now requires Python 3.11 or newer, supports explicit virtual-environment rebuilding, and configures the reference miner's Claims repository path automatically. (b7e4c8f)