Skip to content

PENDING.md

Latest commit

 

History

History
87 lines (61 loc) · 6.96 KB

PENDING.md

File metadata and controls

87 lines (61 loc) · 6.96 KB

PENDING — queued protocol changes

As of v0.5.1 (2026-04-09), the authoritative queue for protocol changes is the multicheck roadmap project board at https://github.com/users/Hulupeep/projects/25.

Individual tickets are at https://github.com/Hulupeep/multicheck/issues with labels:

  • v0.5.1 — small items, shipped or shipping soon
  • v0.6.0 — Mythos-era items, ship when Mythos lands or prototypes stabilize
  • v1.0 — major protocol additions, speculative
  • mythos-era — catalyzed by Mythos-class model capability
  • new-rule / recipe-refinement / tooling / documentation / meta-observation

This file is kept for historical reference and for meta-observations that don't fit the issue tracker well. New items go as GitHub issues, not here.

History of data-collection freezes

Freeze #1: 2026-04-07 through 2026-04-09

  • Freeze started: 2026-04-07 (at commit a83ac4a initial Phase 1 release)
  • Partial unfreeze #1: 2026-04-08 — item #5 pre-flight Q1-Q6 only, commit 7f73c1b
  • Full unfreeze / v0.5.0: 2026-04-09 — items #1-#4 from the reviewer handoff + M2 close, commits d544eee / 5064265 / 148d0cf, tagged v0.5.0
  • v0.5.1: 2026-04-09 — Mythos era prep: irreversible-action gate, evals in DoD, metrics model-tracking
  • Freeze rationale: 3-day window of rule stability for session-to-session data comparability
  • Outcome: 5 queued items, 2 partial unfreezes, 1 meta-observation closed (M2), ~30 metrics rows captured across two reference sessions

Items moved to GitHub issues

When the freeze ended and the issues got created (2026-04-09), the items in this file were transferred as specflow-compliant tickets. The PENDING.md history for each is preserved below as a pointer.

Originally-queued PENDING items

  • PENDING #1 Diff-content check for cascaded files → issue #1
  • PENDING #2 Branch-base sanity check → CLOSED (automation half shipped as hooks/pre-push.sh in v0.5.0; markdown half shipped as pre-flight Q2 in v0.5.0 partial unfreeze)
  • PENDING #3 Cross-layer value consistency (reviewer-time) → issue #2
  • PENDING #4 Specialist-persona sweep → issue #3
  • PENDING #5 Propagation manifest prototype → issue #4

Meta-observations

  • M1 Verify one level above → issue #5 (still open)
  • M2 Markdown rules have a ceiling of effectiveness → CLOSED in v0.5.0 by hooks/pre-push.sh
  • M3 Multi-reviewer asymmetry is multicheck's core thesis → issue #6 (documentation)
  • M4 External swarm as optional third reviewer layer → issue #7
  • M5 Multicheck's value scales with model capability → folded into the framing of the Mythos-era tickets (issues #8-#24); no standalone ticket

Mythos-era items (from analysis of the Mythos Preview System Card §4.1.1)

Original analysis proposed 8 items M-1 through M-8; the system card itself surfaced 9 more specific findings (SC-1 through SC-9).

  • M-1 [P-NNN] plan entry type → issue #8 (v0.6.0)

  • M-2 irreversible-request with [H-NNN] authorization → SHIPPED in v0.5.1, issue #9

  • M-3 Pre-flight Q8 spec conformance → issue #10 (v0.6.0)

  • M-4 Deployment gates → issue #11 (v0.6.0)

  • M-5 Evals in Definition of Done → SHIPPED in v0.5.1, issue #12

  • M-6 N-agent topology documentation → issue #13 (v0.6.0)

  • M-7 Metrics schema model-tracking → SHIPPED in v0.5.1, issue #14

  • M-8 Re-document Q7 purpose → issue #15 (v0.6.0)

  • SC-1 Cover-up detection patterns → issue #16 (v0.5.1 candidate)

  • SC-2 Blast radius check → issue #17 (v0.5.1 candidate)

  • SC-3 /proc/ and credential access restrictions → issue #18 (v0.5.1 candidate, partially covered by M-2 irreversible gate)

  • SC-4 Public-surface write authorization → issue #19 (v0.5.1 candidate, partially covered by M-2 irreversible gate)

  • SC-5 Scratchpad-reasoning unreliability note → issue #20 (v0.5.1 candidate)

  • SC-6 Evaluation awareness note → issue #21 (v0.5.1 candidate)

  • SC-7 No unauthorized side quests → issue #22 (v0.5.1 candidate)

  • SC-8 Git history integrity hard rule → issue #23 (v0.5.1 candidate, partially covered by M-2 irreversible gate)

  • SC-9 "Auto-mode is not sufficient" README repositioning → issue #24 (v0.5.1 candidate)

Partial v0.5.1 rationale

What shipped in v0.5.1: M-2 (irreversible gate), M-5 (evals in DoD), M-7 (metrics model-tracking).

Why these three:

  • M-2 is preventive against the worst failure mode Mythos documents (reckless irreversible actions). The Mythos System Card §4.1.1 gives Anthropic's own direct warning — "we do not expect [auto mode] to be sufficient to fully eliminate risk" — and the irreversible gate is multicheck's answer. Cannot wait for Mythos to ship; the pattern applies to Codex-class builders too (drop tables, force-push, rm -rf on wrong dir).
  • M-5 is trivial to add (one field in a template, one step in the reviewer's verification order). Useful even in the Codex era.
  • M-7 is schema prep for future capability-correlation analysis. Must ship before the data starts flowing from mixed-capability pairings, or we can't correlate.

Why not M-1, M-3, M-4, M-6, M-8, and SC-1 through SC-9: these are either Mythos-specific (only useful once Mythos is general-access) or depend on more session data before fold-in is informed. They're queued as issues and will be re-evaluated when the next session data arrives or when Mythos ships.

Several of the SC-1 through SC-9 items are strong candidates for an imminent v0.5.2 once the v0.5.1 irreversible-gate behavior is observed in practice. Specifically SC-1 (cover-up detection), SC-7 (no unauthorized side quests), and SC-9 (auto-mode warning) are low-risk additions that complement M-2 directly. The operator can decide when to fold them.