Skip to content

paper: v2 — extend the visibility-asymmetry claim cross-session (§3.6, +weave-ins) + governance-plan retarget - #3

Open
mavaali wants to merge 1 commit into
mainfrom
mihir/v2-cross-session-boundary
Open

paper: v2 — extend the visibility-asymmetry claim cross-session (§3.6, +weave-ins) + governance-plan retarget#3
mavaali wants to merge 1 commit into
mainfrom
mihir/v2-cross-session-boundary

Conversation

@mavaali

@mavaali mavaali commented Jun 14, 2026

Copy link
Copy Markdown
Owner

TLDR

Extends the paper's structural claim to the cross-session boundary as a structural argument + existence proof — without extending the experiment. The within-session variance result remains the headline. One new subsection (§3.6) + three surgical weave-ins + a governance-plan retarget. Diff is additive and small.

Discipline rule (cited verbatim from the brief): if a sentence states the boundary, it belongs in v2. If a sentence tries to operationalize crossing it, it belongs in ≤2 sentences of future work, or in paper two. When unsure, cut.

Every sentence in §3.6 and the §2.2/§7.3 weave-ins states the boundary or the existence proof. The only crossing-gesture is §7.9 (2 sentences, deferred future work).

Verified omnigent facts (Phase A2)

Re-verified from source on current main, fresh clone outside this repo's tree. Pinned commit: 0fe2d94cd2e8f77c6715a2d732b0882da2a97095 (omnigent-ai/omnigent). Prior Phase-0 gate reviewed abd110c; no drift on the three load-bearing facts:

  1. Spawn-tree-vs-independent-session boundaryruntime/policies/engine.py: root resolution self._root_conversation_id = root_conversation_id or conversation_id (L130); the per-session budget is routed to the root conversation so "approving once covers the whole spawn tree — a sub-agent runs as its own conversation" (L480–498), while a top-level session has root == self. Accumulates across a spawn tree, resets at the independent-session boundary.
  2. spec/types.pyStateUpdateAction = {SET, INCREMENT, DELETE, APPEND} (includes INCREMENT, L1097); PolicyAction = exactly {ALLOW, ASK, DENY} (L1081). ✓
  3. stores/conversation_store/sqlalchemy_store.pyset_session_state(conversation_id, state) "Serializes state as JSON and writes it to the session_state column" (L873–894). ✓

Convergence dates (Phase A3) — one honest caveat

  • Project public/software release: 2026-04-25 (CITATION.cff, git). arXiv v1 draft dated June 2026 (draft.tex:40). Omnigent main current through 2026-06-14; governance notes date omnigent June 2026.
  • I could not independently confirm the exact arXiv v1 submission timestamp from the repo (no arXiv ID present; STATUS.md still read "finalizing", which the author has confirmed is stale — v1 is submitted). Because v1 and omnigent are both mid-2026, the §2.2 sentence is framed as "independent, contemporaneous convergence" rather than over-precisely "post-dates v1" — the defensible framing given the date evidence. Not papered over.

Edit locations

Loc Change
draft.tex §3.6 (new, after §3.5) "The cross-session boundary" — three moves: categorical asymmetry, persistence objection answered, omnigent existence proof (commit pinned in footnote).
draft.tex §7.3 +1 sentence: load-bearing claim survives any session boundary, more starkly than within a session.
draft.tex §2.2 +1 sentence: contemporaneous industry convergence (omnigent Plan-then-Execute + per-session ratcheting budget).
draft.tex §7.9 ≤2 sentences: cross-session scope-and-reset (per-user vs per-user×agent, window, decay) named as the open problem; explicitly deferred.
notes/omnigent-governance-plan.md New §0 (2026-06-14): retarget cross-session/spawn-tree residual from Paper 2 → Paper 1 v2; compensation axis unchanged in Paper 2 (keeps Paper 2's axis clean).

Acceptance checklist

  • A2 facts re-verified from omnigent source on current main; commit SHA pinned (0fe2d94); no drift.
  • One new subsection (§3.6) with all three moves.
  • §7.3 extended by one sentence.
  • §2.x convergence sentence added (§2.2).
  • §7.9 future-work gesture ≤2 sentences.
  • No results section, table, figure, or eval artifact modified.
  • Diff is additive and small — a sharpened claim, not a second paper.
  • Optional IFC/scalar §4.4 paragraph omitted (actual §4.4 is "Trust budget"; slots awkwardly — excluded per author direction).
  • v1 headline (within-session variance) unchanged.

Notes for reviewer

…eave-ins)

Extend the paper's structural claim to the cross-session boundary as an
existence proof, without extending the experiment.

- New §3.6 "The cross-session boundary": categorical asymmetry, the
  persistence objection answered, and omnigent as the deployed-boundary
  existence proof (verified @ 0fe2d94).
- §7.3: one sentence carrying the load-bearing claim across the session
  boundary (more starkly — no context leakage possible).
- §2.2: one convergence sentence (omnigent ships Plan-then-Execute + a
  per-session budget that ratchets across a session/spawn tree).
- §7.9: ≤2-sentence future-work gesture naming the cross-session
  scope-and-reset policy as the open problem; explicitly deferred.
- notes/omnigent-governance-plan.md: retarget the cross-session/spawn-tree
  residual from Paper 2 to Paper 1 v2; compensation axis stays in Paper 2.

No experiments, scenarios, figures, tables, or eval artifacts changed.
@mavaali
mavaali changed the base branch from mihir/omnigent-policy-review to main June 14, 2026 22:02
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant