Missing-subject audit vs "Architecting the Agent Multiverse" (slide deck)¶

Source: Architecting_the_Agent_Multiverse.pdf (15 slides, NotebookLM deck of the Chronos Protocol / CausalFlow research). This maps every subject in the deck to the current library (v0.4.0) and specifies the missing features with enough context to implement each. Companions: HANDOFF.md, ANALYSIS.md.

Slide-by-slide coverage¶

#	Slide subject	Status in v0.4.0
1	Multiverse DAG of forked executions	⚠️ partial — see Gap 1
2	Entropy of autonomy (alignment decay over 70h)	✅ `drift()` entropy/health curve — Gap 6
3	Diagnostic trap (14% LLM-judge; hand vs brain)	✅ the library's premise
4	Execution-as-Code ("model may change, history must not")	✅ cassette + seeds
5	Workflow/Activity dichotomy, event ledger, replay mode	✅ record/replay; ⚠️ resume — Gap 5
6	OS-level substrates (DeltaBox/CRIU/Wasm table)	✅ documented out-of-scope
7	Execution Merkle DAG, CAS, O(1) branching, node = parent/action/output/fs hashes	⚠️ linear chain only — Gap 1; action/output hash split — Gap 7
8	Network determinism: VCR, virtual time, seeded entropy (uuid/rng)	⚠️ rng only — Gap 2
9	SCM (C_t → A_t → O_t → Y) + attribution formula	✅
10	do() calculus tracks (null vs forced intervention)	✅
11	Stochastic confound + Point-of-Commitment	✅
12	Shapley (AND/OR gates) + antithetic pairing	✅ ground-truth tested
13	Step-level faithfulness: causality vs coverage quadrant	❌ — Gap 3
14	Repair loop (isolate → inject → validate, minimality)	✅ repair v2
15	Multiverse Console: timeline graph, branch compare, REPL in frozen state, audit trail	❌ — Gaps 1, 4, 5

The gaps, with implementation context¶

Gap 1 — The Multiverse itself: first-class forking & persisted branches¶

Deck: slides 1, 7, 15 — "forking an agent at Step 45 creates a new node with a shared parent hash… instantly instantiates a parallel universe"; the console compares Failed vs Fixed timelines side by side. Today: forking exists implicitly (every ablation rollout is a transient fork via ReplayPlan), but nothing is user-facing or persisted: sessions are linear; a counterfactual run cannot be saved, named, listed, or diffed. Build (v0.5 candidate — highest value of these): - fork(trajectory, at_step, *, do=None, seed) -> Trajectory in a new multiverse.py: run ReplayPlan(held=<prefix, optionally forced>), record the live continuation with a branch-aware recorder, return a new Trajectory whose meta carries parent_session, fork_step, intervention. - Store: add parent_session / fork_step columns (nullable, migration like resamplable); CheckpointStore.branches(session_id) lists children. Storage is already O(1)-ish: forked steps dedupe through the CAS blobs. - diff(traj_a, traj_b) → first divergent step + per-step action diff — the "State Diff" the research UI shows. CLI: agent-replay fork / branches / diff.

Gap 2 — Virtual time & entropy (clock/uuid determinism)¶

Deck: slide 8 — "hook system calls (wasi:clocks/get_time) to inject logged timestamps; seed RNGs deterministically". Today: only ctx.rng is deterministic. An agent calling datetime.now(), time.time() or uuid.uuid4() silently breaks replay fidelity (held steps are fine, but resampled/live paths diverge from wall-clock and fresh uuids, and any recorded output embedding a timestamp changes its idempotency key across runs). Build (small, high leverage): - ctx.now() and ctx.uuid() on both contexts: during recording they capture real values as steps (kind=memory, name=__now__/__uuid__, resamplable=False); during replay they serve the recorded values — no API change to the engine, they ride the existing cassette. - instrument.RECIPES["time"] patching time.time/datetime.datetime.now and uuid.uuid4 onto those ctx ops for unmodified agents (opt-in, since global).

Gap 3 — Step-level faithfulness (causality vs coverage)¶

Deck: slide 13 — correctness ⊥ faithfulness quadrant (correct-but-unfaithful = "dangerous post-hoc rationalization"; wrong-but-faithful = "ideal debugging target"); FaithCoT-style masking. Today: the verifier scores outcome only; nothing measures whether the recorded reasoning steps actually drive the answer. Build (v0.5/v0.6, research-grade feature): faithfulness.py — - Causality score per reasoning step: mask/remove step i (existing remove intervention), hold everything else, N rollouts → shift in outcome distribution; AUC over progressive masking = trajectory faithfulness. - Classify runs into the 2×2 (correct×faithful) using outcome score × faithfulness score; flag correct-unfaithful runs in the report ("passed, but the reasoning did not cause the answer") — these are the silent risks. - Reuses AblationEngine wholesale; mostly a new scorer + report section.

Gap 4 — Multiverse Console (serve UI) with branch graph & frozen-state inspection¶

Deck: slide 15 — timeline graph with counterfactual branches, reverse-step navigation, "Interactive REPL inside frozen states". Today: static HTML report only. Build (already roadmapped as v0.5 agent-replay serve; the deck adds specifics): stdlib http.server browsing sessions → per-step state view (inputs/outputs = the app-level "frozen state"), branch graph once Gap 1 lands, and a "step back/forward" inspector. A true REPL into process memory is OS-substrate territory (out of scope) — the app-level equivalent is dumping the step's recorded inputs/outputs and context digest.

Gap 5 — Durable resume (event-sourcing recovery)¶

Deck: slide 5 — replay = "fast-forward through days of successful operation in milliseconds" to resume a crashed run, not only to analyse it. Today: replay serves attribution; there is no "continue this run live from step k" convenience. Build: thin wrapper over Gap 1's fork: resume(trajectory) = fork(trajectory, at_step=len(trajectory), seed=...) — hold the entire recorded prefix, run live from the end. Cheap once fork exists.

Gap 6 — Drift/entropy observability (nice-to-have)¶

Deck: slide 2 — alignment decay curve, context drift, silent failure. Today: nothing tracks per-step health over long runs. Build (optional): if the verifier can score intermediate state, record a per-step score series and chart it in the report — the "entropy curve" for a run. Low priority; useful mainly for long-horizon demos.

Gap 7 — Minor: split action vs output hash in the Merkle node¶

Deck: slide 7 — node = parent + action hash + output hash (+ fs root hash). Today: step_hash = link(parent, op_key, output_hash) — the parts exist but only the combined hash is stored. Persisting action_hash/output_hash separately enables "same decision, different observation" queries and cheaper diffing (Gap 1). Trivial store migration.

Priority recommendation¶

Gap 2 (virtual time/uuid) — small, closes a real determinism hole.
Gap 1 (fork/branches/diff) — the deck's headline; unlocks Gaps 4 & 5.
Gap 5 (resume) — nearly free after Gap 1.
Gap 4 (serve UI w/ branch graph) — already v0.5 roadmap.
Gap 3 (faithfulness) — differentiating research feature.
Gaps 6–7 — opportunistic.

Status (v0.6.0) — deck fully covered¶

✅ Gap 1 — multiverse.fork/afork/diff, CheckpointStore.branches, CLI.
✅ Gap 2 — ctx.now/ctx.uuid, instrument.enable_virtual_time.
✅ Gap 3 — faithfulness() + quadrant classification, CLI.
✅ Gap 4 — serve.py Multiverse Console, agent-replay serve.
✅ Gap 5 — multiverse.resume.
✅ Gap 6 — drift() per-step entropy-of-autonomy curve + optional intermediate-state state_scorer alignment-health overlay, silent-decay detection, DriftResult.to_html SVG chart, CLI drift. stats.binary_entropy.
✅ Gap 7 — Step.action_hash/output_hash.

Every slide subject in the deck now maps to a shipped capability. Gap 6 resolved its "needs an intermediate-state scorer" caveat by (a) providing the state_scorer hook for callers who have one and (b) always charting the verifier-only entropy-of-autonomy curve when they don't.