Skip to content

feat: wait for capacity and coordinate eligible workers - #43

Merged
spfuzzylink merged 31 commits into
mainfrom
spfuzzylink/paraphrase-quality
Nov 10, 2026
Merged

spfuzzylink merged 31 commits into
mainfrom
spfuzzylink/paraphrase-quality

Conversation

@spfuzzylink

@spfuzzylink spfuzzylink commented Oct 28, 2026 •

Copy link
Copy Markdown
Owner

A busy inference worker should wait before claiming work, or leave it for a ready worker. It should still return a completed exact result without waiting for model capacity.

This adds execution.NewDispatcher with bounded pending calls, admission waiting, explicit same-runtime nodes and optional read-only peer-result polling. It advances only on proven pre-claim deferral. Pending or uncertain work is never regenerated or handed off. The existing coordinator, exact identities, source expiry, budgets and quality checks keep their behavior.

The release also includes an opt-in bounded applicability assessment and adopted-update rejection guard. Free-form reuse remains experimental and unqualified; it is not enabled by default. The original #41 accuracy and whole-task savings goals continue in #44. #41 will be closed as superseded, not as a passing quality result. All 618 prior attempts, seven usage-unknown outcomes and 31 unfinished fences remain recorded.

Validation:

  • Full Go race/vet and Python checks, including exact reuse, recovery, scope and accounting regressions.
  • Fifteen dispatcher test groups, ten independent adversarial controls repeated three times, and an independent lost-CAS/peer-reuse regression.
  • Model-free HTTPS walkthrough: pressured node generates nothing; ready node handles the task once; another scoped client reuses the exact result under pressure with zero new tokens. Scripted providers on one physical host, not a model-quality or physical-cluster benchmark.
  • Explicit release allowlist now includes the new guide, complete examples and retained diagnostic evidence; archive tests execute that evidence verifier from packaged files.
  • Four native platform CI, source-container, isolated plugins and live self-hosted Langfuse checks must pass for the exact release commit.

Paid fallback is disabled; provider API spend for this work is $0. Questlock applied checked documentation revisions without model calls. This does not measure the surrounding Codex chat's allowance or establish invoice savings. No active client configuration was changed.

Refs #41. Follow-up: #44.

Add separately accounted source certificates and full-context question checks behind an explicit policy. Preserve default and exact reuse behavior. Include independent protocol tests, six-case live measurement and documentation disclosing failed qualification. No release or version change.
@spfuzzylink spfuzzylink added documentation Improvements or additions to documentation enhancement New feature or request labels Oct 28, 2026
@spfuzzylink spfuzzylink self-assigned this Oct 28, 2026
@spfuzzylink spfuzzylink changed the title feat: two-stage contextual reuse (qualification pending) fix: check saved answers before paraphrase reuse (qualification pending) Nov 1, 2026
@spfuzzylink spfuzzylink changed the title fix: check saved answers before paraphrase reuse (qualification pending) fix: verify paraphrase reuse with configurable resource admission Nov 1, 2026
spfuzzylink and others added 5 commits November 3, 2026 22:43
spfuzzylink added a commit that referenced this pull request Nov 4, 2026
Co-Authored-By: Claude Code <noreply@anthropic.com>
@spfuzzylink
spfuzzylink force-pushed the spfuzzylink/paraphrase-quality branch 2 times, most recently from 41cc5f9 to 3289d1c Compare November 4, 2026 09:02
@spfuzzylink
spfuzzylink marked this pull request as ready for review November 4, 2026 09:10
@spfuzzylink
spfuzzylink force-pushed the spfuzzylink/paraphrase-quality branch from 3289d1c to a2acc0a Compare November 4, 2026 09:18
@spfuzzylink
spfuzzylink marked this pull request as draft November 8, 2026 09:39
Restore the full evaluator, remove fabricated quality metrics, and port versioned rejection-only scope, role and reply-polarity checks. Preserve exact reuse and per-stage cost/uncertainty accounting. Keep the experimental release gated on local qualification.
Distinguish narrower questions from changed facts while keeping contextual verification mandatory. Reject explicit self-role changes and direct structured-output mismatches. Record the failed full regression and keep the release qualification gate closed.
Add identity-bound pair reasoning and sampling controls with bounded diagnostic stopping. Retain six failed development profiles and their measured usage. Exact reuse and legacy defaults remain covered by compatibility tests.

Draft checkpoint; quality qualification is still open. Refs #41.
Update the strict discovery contract for the two new optional capability lists. Existing-binary clean plugin checks pass for Codex and Claude, including revision and exact reuse, with zero model calls.

Refs #41.
Keep source bindings separate from the current target, reject explicit operation-object swaps, and allow bounded caller-selected thinking deadlines while preserving defaults. Unicode, privacy, legacy identity and pending-task fences retain regression coverage.

Retain eight failed diagnostics and the independently audited passing 20-request pilot, including unknown usage. Larger and fresh qualification remain open. Refs #41.
Replace fragile 50ms deadline assumptions with synchronized pair cancellation and explicit inherited-deadline checks. Preserve shorter resolver deadline coverage and pending-claim fences.

Validation: both focused tests passed 20 race-enabled runs; independent review approved the tests-only tree.
Record the 144-case population, including two unsafe approvals, a withheld contradictory verdict, one known output-limit failure and 107 unattempted rows. Keep all prior reports and unresolved outcomes.

Independent audit closes 586 attempts. Evidence replay and four semantic fault controls pass. This remains a draft; fresh quality and savings gates have not passed.
Add rejection-only CSV/TSV row and named-column checks with quoted, reported and ambiguous requests deferred to verification. Clarify the factorized target-scenario and verdict protocol, preserving the source frame and standalone/exact paths.

Full checks, focused race tests, independent guard and mock-interface controls, ten offline quests and packaged plugin checks pass. Native quality and savings qualification remain open; this is a draft checkpoint.
Add bounded diagnostic assessment before reuse decisions, reject explicitly adopted evidence changes, and preserve exact/default behavior. Extend matched measurement controls for explicit cases and budgets. Retain failed quality and critical-memory evidence; #41 remains open and no release is qualified.
…hecks

Add bounded dispatcher and read-only peer waiting, preserve exact/runtime ownership, package complete public examples and evidence. Scope the release to verified coordination; experimental reuse qualification continues in #44.
@spfuzzylink spfuzzylink changed the title fix: verify paraphrase reuse with configurable resource admission feat: wait for capacity and coordinate eligible workers Nov 9, 2026
@spfuzzylink
spfuzzylink marked this pull request as ready for review November 9, 2026 22:05
@spfuzzylink
spfuzzylink merged commit ed77fc6 into main Nov 10, 2026
16 checks passed
@spfuzzylink
spfuzzylink deleted the spfuzzylink/paraphrase-quality branch November 10, 2026 01:02
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

documentation Improvements or additions to documentation enhancement New feature or request

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant