nut - safety net protocol

benchmarks
safety net protocol
nut shows its chain of thought; verified staff under nda perform a safety pre-scan; then the public community reviews, discusses, and endorses or flags. a minimum public quorum is required for a chain to reach full trust status.

key principles

radical publication: chains are published (redacted where necessary) so the public can inspect reasoning.
safety-first pre-scan: nda staff prevent leaks of contractually-protected or dangerous material before public view.
public governance: public moderation follows transparent rules, edit histories, talk pages, and reputation-weighted consensus.
quorum + two-key trust: full trust badge requires nda clearance + x independent public endorsements.
auditability: signed audit hashes, immutable revision history, and exportable logs for partners/regulators.

pipeline — step by step (text flow)

generation
nut produces the raw chain for a task. raw internals remain internal.
automatic masking & sanitization
pII and obvious secrets are auto-detected and masked (e.g., [redacted: email], city-region). adversarial-sanitizer scans for instruction-like, illegal, or clearly dangerous fragments and flags for nda staff.
nda staff pre-scan (first gate)
verified staff operating under nda:
- confirm masking worked and remove contractual secrets
- redact or escalate if the chain contains regulated content (medical/legal/child-safety)
- optionally tag a chain as “partner-private” or route it to partner queues
  outcome: pass → public release queue; block/rollback → internal incident handling; escalate → compliance/legal.
public release (visible archive + talk page)
the redacted chain is published to the public archive with:
- full revision history (diffs, who edited what and when)
- an attached talk page for discussion
- metadata: decision id, nda-scan stamp, redaction log, signed audit hash, initial reviewer note
public moderation & endorsements
registered community moderators review the chain. actions they can take:
- endorse (upvote/endorse for trust)
- flag (safety/illegal/harassment/privacy)
- suggest edits (notation, rewording of steps, adding provenance)
- rollback to earlier revision
  endorsements and flags are visible; each endorsement contributes to quorum.
quorum check & trust badge assignment
if a chain reaches:
- nda cleared + fewer than x endorsements → provisional trust badge
- nda cleared + ≥ x independent endorsements → full trust badge
  the value x is public (default: 5), configurable per category or partner (higher for sensitive topics).
post-badge stewardship
full-trust chains remain editable but badge persists until revoked by transparent process (see revocation below). every edit produces new revision and may require re-endorsement depending on scope.

roles & permissions

anonymous visitors
- can view archives and talk pages
- can flag content (flags enter moderation queues)
registered contributors
- can comment and suggest edits on talk pages
- can flag; suggestions appear for moderators
trusted editors (community-elected or vetted)
- can edit public traces to improve clarity (edits are recorded and attributed)
- can add provenance or citations
- cannot remove reviewer metadata or change redaction placeholders
public moderators (verified community moderators)
- can endorse/endorse with comment, flag for escalation, rollback edits, lock a chain temporarily, and vote on dispute resolution
- their endorsements count toward quorum
- moderation actions are auditable and linked to moderator reputation
nda staff / verified staff
- operate the pre-scan gate, can tag, redact, or block prior to public release
- can add “partner-private” flags and route to partner queues
- do not perform community moderation once public (unless also registered moderator)
admins / stewards
- manage policy, elections, resolve arbitration, publish governance updates
- cannot silently remove data; all administrative actions are logged and transparent
partner reviewers (for partner-private chains)
- partner-appointed, enterprise-verified reviewers who participate in partner queues under NDA and SSO
- their endorsements may satisfy partner-specific rules for partner badges but do not substitute public quorum unless configured.

reputation, voting weight & quorum

each moderator has a reputation score derived from:
- tenure (how long they’ve been active)
- contribution quality (peer-reviewed edits)
- accuracy/history (how often their endorsements aligned with later arbitration outcomes)
- identity verification level (email, id, sso for partners)
endorsement weight = base weight × sqrt(reputation). this reduces dominance by a single high-rep user but rewards sustained positive activity.
quorum formula (default): need cumulative endorsement weight ≥ threshold (default threshold = 5 × base weight). this is equivalent to ~5 independent verified moderators. thresholds are publicly visible and configurable by category.
conflict-of-interest rule: endorsements from moderators with declared conflict (e.g., employed by a partner mentioned in the chain) are discounted or require additional independent endorsements.

moderation mechanics

revision history & diffs
every published chain is an editable document. every change is stored as a revision with author, timestamp, and signed diff.
talk pages
each chain has a talk page for structured discussion: concerns, evidence, and provenance requests are handled there.
rollbacks
moderators can rollback a chain to a prior revision if an edit introduces unsafe content. rollbacks are recorded and must include a reason.
locking
admins can temporarily lock a chain for edits (e.g., ongoing legal review) — locks are rare and must be justified publicly.
dispute resolution & arbitration
if endorsements and flags conflict, a dispute is opened. a small arbitration panel (mixed community + independent expert) reviews evidence and issues a binding decision. arbitration decisions are logged and published.
appeals
authors/partners can appeal decisions. appeals go to the arbitration panel and are processed under transparent timelines and logged scopes.

anti-abuse & safety measures

rate limits & feerolls
prevent mass endorsement attacks by rate-limiting endorsements per account and applying weight decay for burst activity.
honeypots & honeytokens
detect those attempting to extract secret patterns by inserting canary tokens in internal test data; access triggers alerts.
identity verification tiers
higher moderation power requires stronger verification (e.g., verified email + id + reputation threshold)
adversarial sanitizer
automated pre-publish checks for instruction-like or potentially harmful steps; flagged chains require higher quorum or specialist review.
audit trails & signed records
every moderator action is signed and recorded; audit logs are immutable and exportable to partners on request (under nda if necessary).
spam & sockpuppet detection
machine learning and heuristics detect coordinated fake accounts; suspected accounts are quarantined pending review.
human only
moderation actions that affect badge state, endorsement quorum, or public trust require a cryptographically-signed human attestation. ai agents may propose suggestions but are never permitted to submit or sign moderation actions. actions from machine identities are blocked and audited.

nda staff integration & partner flows

partner-priority queues
partners can request priority nda pre-scans for specific chains or projects. priority requests are logged and visible to partners.
partner-verified badges
partners can apply for a partner-verified badge for their project streams. requirements:
- nda staff clearance + partner reviewer signature(s) + public quorum (unless partner requests partner-only visibility, in which case an enterprise badge is used and visible only on partner pages).
on-prem / private-cloud option
for sensitive enterprise projects, the whole pipeline (generation, masking, nda review, partner moderation) can run on partner infrastructure; public release still follows the same public moderation rules unless the partner opts for partner-only archives under contract.
legal & compliance hooks
partner workflows allow adding legal-review tags, exportable signed logs, and retention policies aligned with partner/regulator needs.

badge lifecycle (beta)

states
- pending — nda cleared, public review queue
- provisional trust — nda cleared + endorsements but below quorum
- full trust — nda cleared + quorum reached
- flagged — active flags that require re-review
- revoked — badge removed after arbitration or nda staff/legal decision
assignment
when quorum is reached the system:
- records the endorsement snapshot (who endorsed, weights)
- signs the chain with an audit hash and issues the full trust badge
- updates partner/chain metadata and archive index
revocation
a badge can be revoked by:
- successful arbitration/appeal demonstrating harm or error
- nda staff legal block (e.g., contractual confidentiality breach)
- evidence of endorsement manipulation or moderator conflict
  revocation includes a public note explaining reason and links to the revised trace.

moderator tools & ux (what moderators see)

moderation dashboard: queue with filters (new chains, flagged chains, partner-priority, category)
diff viewer: compare revisions side-by-side, highlight redactions
talk pane: structured evidence/comments and voting controls
endorsement button with weight/justification field
flag modal with category selection (privacy, legal, safety, misinformation)
rollback + lock buttons with required reason field
audit link: view signed audit hash & export log
reputation panel: shows moderator’s reputation, past decisions, and conflict of interest disclosures

governance, elections & policy changes

policy council
steward group (mix of community-elected moderators, independent experts, and company liaisons) publishes and updates moderation policies.
policy change process
proposed changes go through: proposal → public comment period (30 days) → pilot → community referendum weighted by reputation → adoption.
moderator elections
trusted editors and moderators are elected periodically by community vote; election rules and eligibility are public.
transparency reports
quarterly governance reports with statistics (endorsement rates, revocation reasons, audit results) are published.

training, onboarding & community health

moderator training
required modules: privacy & redaction, bias & cultural sensitivity, legal basics, dispute resolution, safety triage.
mentorship
new moderators are paired with experienced stewards for early decisions.
community codes of conduct
enforce civility and evidence-based moderation; violations reduce reputation and can lead to suspension.
incentives
reputation points, recognition (badges, leaderboard), priority access to partner-sandbox features, and optional monetary grants for top contributors.

metrics & transparency (what we publish)

public dashboard with:
- number of chains published, provisional vs full-trust counts
- average time from publication → quorum (median)
- moderator count, active languages, top categories
- revocation events and reasons
- % decisions nda-cleared, % decisions fully trusted
per-chain metadata (visible on archive):
- decision id, nda-scan stamp, endorsement snapshot, signed audit hash, revision history, talk page link

sample policies & short templates

moderation guideline

focus on safety first: flag content that risks privacy, illegality, or serious harm.
prefer redact + suggest edits over deletion.
disclose conflicts of interest before endorsing.
provide justification with each endorsement.

conflict of interest disclosure

“i disclose that i am affiliated with [org] which is mentioned in this chain. my endorsement carries reduced weight until independent endorsements are recorded.”

appeal request (template)

“i request appeal for decision id [uuid]. reason: [new evidence]. suggested action: [re-evaluate/revoke badge/restore revision].”

technical endpoints (illustrative until launch)

GET /archive/:decision_id → public redacted trace + revision history + talk link
POST /endorse/:decision_id → moderator endorsement (signed, authenticated)
POST /flag/:decision_id → public flag submission
GET /audit/:decision_id → signed audit hash + export instructions (partner access under nda)
GET /moderation/queue → moderator dashboard feed (auth required)
POST /partner/request-badge → partner badge request (priority nda queue)

faq

q: how many public moderators are needed?
a: default quorum = cumulative endorsement weight equivalent to 5 independent verified moderators; this is public and configurable per category.

q: what if public moderators disagree?
a: a dispute triggers arbitration. the arbitration panel reviews evidence and issues a public decision which may assign/revoke badges.

q: can partners bypass public moderation?
a: no. public trust badge requires public quorum. partners may use private partner badges for internal workflows or on-prem archives, but these do not substitute full public trust.

q: how do you prevent manipulation?
a: reputation-weighted endorsements, rate limits, sockpuppet detection, conflict-of-interest rules, and visible endorsement snapshots make manipulation both costly and detectable.

Page updated

Google Sites

Report abuse