core/docs/adr/ADR-0220-engine-identity-vs-build-provenance.md

ADR-0220 — Engine identity vs. build provenance (code_revision in the identity hash)

Status: Accepted (ratified 2026-06-15 by repository owner). PR A merged; PR B merged (#772); PR C — identity/provenance split — implemented on feat/identity-substrate-hash-split. Date: 2026-06-15 Relates: ADR-0146 (engine-state persistence), ADR-0156 (atomic checkpoint), ADR-0157 (revision-mismatch warning), ADR-0219 (generation-dir atomic checkpoint), L11 identity continuity (core/engine_identity.py, commit f2dac1dc), L10-runtime-model-scope

This ADR documents a contradiction between two already-ratified decisions and proposes how to resolve it. It is the "PR A" of a 3-part sequence (A = this brief, B = safe operator ergonomics, C = the identity/provenance semantics change). It changes no runtime behavior. Per project doctrine (address critiques, don't waive them; fix upstream, not beside), the semantics change (C) must not ship before this decision is accepted.

Context — the incident

core always-on refused to resume a 26 396-turn lived life:

checkpoint was written under 75f3bb75649d…
but this build computes b0f89456fe26…

The checkpoint (engine_state/manifest.json, now backed up under engine_state/_life_backup_c6d0e2a_f5c6914d/) was stamped at revision c6d0e2a92004; the running build is f5c6914d0083. Under the always-on daemon's forced strict-continuity config the load guard raised IdentityContinuityError and the daemon exited (core/cli.py:393-401, exit code 2).

This was diagnosed and operationally resolved (back up the runtime files — not the directory; see "Operator ergonomics" below — then start a fresh life). The remaining question is architectural: is the refusal correct behavior, or a defect in how engine identity is composed?

The contradiction (grounded)

EngineIdentity is sha256(canonical(ratified_substrate)), where ratified_substrate is the five ratified personality packs plus the code revision (core/engine_identity.py:92-99):

substrate["code_revision"] = git_revision   # engine_identity.py:99

But the module's own docstring states the opposite contract (core/engine_identity.py:9-11):

"It is bumped only by a ratified change to the identity substrate … NOT by lived learning."

A git commit is not a ratified substrate change, yet it bumps the identity.

This collides head-on with ADR-0157, which already ruled that a revision mismatch on load is a non-fatal warning, not control flow. That decision is live in engine_state/__init__.py:419-433 (load_manifest emits an ADR-0157 RuntimeWarning on written_at_revision mismatch and loads anyway — "reboot is recovery, not control flow"). So the same git revision drives two opposite policies:

Field	Where	Same-rev change ⇒
written_at_revision (provenance)	engine_state/__init__.py:350	warn, load anyway (ADR-0157)
code_revision (inside identity hash)	engine_identity.py:99	hard raise under strict (runtime.py:790-804)

The provenance field and the identity input are the same value from the same source — get_git_revision() (engine_state/__init__.py:121-141) — wearing two contradictory hats.

Verified facts (black-box, against f5c6914d0083)

• The divergence is pure code-revision. Recomputing engine_identity_for_config(RuntimeConfig(), "c6d0e2a92004") reproduces 75f3bb75…637a29fee byte-exactly; with "f5c6914d0083" it is b0f89456…8bd6. All five ratified pack SHAs are identical across the two revisions (the default packs — default_general_v1, core_safety_axes_v1, default_general_ethics_v1, default_neutral_v1, default_unanchored_v1 — have identical blob hashes at both commits). Nothing about who the engine is changed; only the build did.
• The guard works as designed: identity computed at init (runtime.py:749-752), compared to the stamped value (runtime.py:787); under strict_identity_continuity it raises IdentityContinuityError, otherwise it warns and sets the queryable identity_continuity_break flag (runtime.py:790-804). The always-on daemon forces strict (chat/always_on_daemon.py:45-49).

Why this is a defect, not just naming

It is tempting to call this an optics/naming tension. It is sharper than that:

1. It defeats the stated telos during normal development. Because the daemon forces strict_identity_continuity=True, every commit between daemon restarts makes the engine refuse to resume the same life — directly contradicting both the "one continuous life" telos and the module's own "bumped only by a ratified change" contract. (The failure is fail-closed and safe — it refuses rather than silently forking — but the conservatism falls on exactly the wrong axis.)

2. code_revision is simultaneously over- and under-sensitive as an identity input:

   • Over: flips on every commit, even a docs-only or test-only commit that cannot change runtime behavior (proven above — packs byte-identical).
   • Under: get_git_revision() is HEAD-only — an uncommitted/dirty working tree does not change the revision. You can edit operator code and the daemon still believes it is the same life. It is also a 12-char short prefix (--short=12, collision-possible vs. the full SHA), and it returns the literal string "unknown" when git is unavailable — collapsing all builds in a git-less environment to one shared identity.

   A signal that is both too sensitive (every commit) and too coarse (ignores uncommitted edits, truncates the SHA, degrades to a constant) is a poor input to a "who am I" hash.

Decision drivers

• Preserve the L11 guarantee that distinct ratified substrate ⇒ distinct identity (identity is load-bearing and falsifiable).
• Keep the guard fail-closed: a genuinely unsafe resume must still be refused, not silently accepted.
• Stop a behavior-neutral code bump from forking the engine's identity.
• Do not weaken the L10 same-life proof surface or the lineage invariants.

Options

O1 — Status quo (block on any revision change). Honest about "code may have changed → don't resume," but contradicts ADR-0157 and the docstring, and makes continuous-life development impractical. Rejected as the resting state.

O2 — Demote code_revision to a warning everywhere (drop it from the identity hash). Identity = ratified packs only; build revision becomes provenance only, exactly like written_at_revision already is under ADR-0157. Simple and internally consistent. Risk: a code change that genuinely alters serialization/semantics of lived state would then resume silently under the same identity (mitigated by ADR-0219's schema versioning + ADR-0157's warning, but worth stating explicitly).

O3 — Split the hashes (recommended).

identity_substrate_hash = sha256(ratified identity/safety/ethics/register/anchor_lens packs)   # "who am I"
build_provenance_hash   = code_revision (+ optionally a content hash of the runtime)            # "which build"

Then the resume policy becomes an explicit, separately-governed question rather than an accident of putting code_revision inside identity:

same identity_substrate_hash + different build_provenance:
    → warn + stamp provenance, resume (default; matches ADR-0157)
    → OR require an explicit operator opt-in / migration / fork (strict)

O4 — Migration/fork command. Orthogonal to O1–O3: give the operator a first-class way to carry a life across an intended identity change (engine-state fork), instead of manual file surgery. Complements whichever of O2/O3 is chosen.

Decision (ratified 2026-06-15)

Adopt O3 (split), with the resume policy defaulting to warn-and-resume (O2 semantics) and strict-mode requiring explicit operator intent for a build-provenance-only divergence. A genuine substrate divergence (different ratified packs) still refuses under strict — that is the real "different identity". Sequence:

• PR A — this brief. No code. (merged)
• PR B — safe operator ergonomics (no identity-semantics change). (merged, #772)
• PR C — identity/provenance semantics. (this branch) Splits the hash so engine_identity = ratified packs only (NO code_revision); re-points the continuity guard at that substrate hash; records code_revision as provenance only (the manifest already carries written_at_revision); adds an identity-scheme marker so pre-split checkpoints (which folded code_revision in) migrate via warn-and-re-stamp rather than a flag-day strict break; and updates the L11 lineage / L10 soak proof surface — including a new test that a behavior-neutral commit (same packs, different rev) no longer breaks continuity, while keeping "distinct packs ⇒ distinct identity" green.

Honesty note on the split

The split is not a rename. It is the load-bearing semantics change and touches: the continuity guard (runtime.py:787-804), the manifest schema (engine_state/__init__.py:350-353), the lineage tests (tests/test_engine_identity_lineage.py, tests/test_identity_continuity_proof.py, which key off the single engine_identity field and assert engine_identity == parent_engine_identity for a stable life), and the L10 soak runner (evals/l10_always_on/runner.py:113). PR C must keep the L11 "distinct packs ⇒ distinct identity" test green and add a test that a behavior-neutral commit no longer breaks continuity. Note that build_provenance-as-a-distinct-on-disk-field already exists (written_at_revision); what does not yet exist is an identity_substrate_hash computed without code_revision.

PR B — safe operator ergonomics (no semantics change)

Two changes, shippable before the O3 decision, that would have prevented the incident's confusion:

1. core always-on --engine-state PATH. A trivial wire-through: run_daemon already accepts engine_state_path: Path | None (chat/always_on_daemon.py:132,152); cmd_always_on simply never passes it (core/cli.py:384-390), so it falls back to _DEFAULT_DIR. One add_argument + one kwarg surfaces the existing per-state-root concept so an operator can run a per-branch dev life without touching the package dir. (Note: a CORE_ENGINE_STATE_DIR env var already does this — engine_state/__init__.py:52-56 — so the flag is a second selection path to keep consistent, and --no-load-state already provides an ephemeral fresh start. The genuinely missing piece is a persisted fresh life under a new dir while preserving the old.)

2. A humane IdentityContinuityError recovery message. Today the handler (core/cli.py:393-401) states the mismatch but gives no recovery path.

   ⚠️ Footgun to avoid (this corrects the originally-suggested message). Do not tell the operator to mv engine_state … or rm -rf engine_state. When CORE_ENGINE_STATE_DIR is unset, the default engine-state directory is the git-tracked engine_state/ Python package (_DEFAULT_DIR = parents[1] / "engine_state"; engine_state/__init__.py is the only tracked file there, runtime data is gitignored alongside it). Moving or deleting it removes engine_state/__init__.py, breaking from engine_state import … across ~20 files — including the very core always-on command the message tells the operator to re-run (core/cli.py:341).

   The correct recovery guidance:

   This checkpoint belongs to a different engine life (the ratified substrate
   or build revision changed since it was written).
   
   Options:
     1. Resume the old life — run the build that wrote it:
          git checkout <checkpoint_revision>   # e.g. c6d0e2a92004
          core always-on
     2. Start a fresh persisted life under a separate state dir
        (does NOT touch the old one):
          core always-on --engine-state ./engine_state_<name>
        (or: CORE_ENGINE_STATE_DIR=./engine_state_<name> core always-on)
     3. Start fresh in place — clear ONLY the runtime state, never the package:
        back up/remove the runtime files inside the engine-state dir
        (manifest.json, the current pointer + gen-*/ dirs, recognizers.jsonl,
        discovery_candidates.jsonl, session_state.json, proposals.jsonl),
        leaving engine_state/__init__.py intact.
   

   The message should print the actual <checkpoint_revision> from the stamped manifest so option 1 is copy-pasteable.

"Start fresh in place" — clear the directory, not just the manifest

Removing only manifest.json is sufficient only in the legacy flat layout (where EngineStateStore.exists() keys on manifest.json, engine_state/__init__.py:456). In the production ADR-0219 generation-dir layout, exists() keys on the current pointer, so a manifest-only removal still leaves exists()==True and load_recognizers()/ load_discovery_candidates() ingest the prior life's derived state. Worse, proposals.jsonl is loaded independent of the exists() gate (runtime.py:761-763). A correct in-place fresh start clears all runtime files listed above. (The incident's dir was flat and was fully cleared, so its fresh start is clean.)

What this ADR does not change

• It does not modify core/engine_identity.py, the guard, or the manifest schema. PR B touches only the CLI surface (flag + message). PR C is gated on acceptance of O3.
• It does not weaken strict_identity_continuity; it proposes re-pointing what identity means so strictness lands on substrate, not build.

Corrected non-issues (so reviewers don't chase them)

• The stale always_on.lock is not a hazard. The single-instance lock is an advisory fcntl.flock the kernel releases on process death; the lock file is intentionally never unlinked (chat/always_on_daemon.py:55-64). A dead holder's marker does not block a new daemon.
• engine_state/ is not scanned by the architectural-invariant suite — there is no INV risk from runtime files living beside the package.

Evidence appendix

# pure-code-revision proof (RuntimeConfig defaults, packs byte-identical):
engine_identity_for_config(RuntimeConfig(), "c6d0e2a92004") == 75f3bb75…637a29fee   # stamped, old life
engine_identity_for_config(RuntimeConfig(), "f5c6914d0083") == b0f89456…8bd6        # current build
git diff --stat c6d0e2a92004 f5c6914d0083 -- packs/{identity,safety,ethics,register,anchor_lens}  == (empty)

# the two-hats-on-one-value relationship:
get_git_revision()  ──▶ code_revision  (hashed into engine_identity)   engine_identity.py:99   → strict raise
                    └─▶ written_at_revision (provenance, unhashed)      __init__.py:350         → ADR-0157 warn

Governance Cross-Reference (ADR-0225)

This late-corpus ADR is governed by ADR-0225:

• Safety boundaries: changes must preserve ADR-0027/0028/0029 identity and safety-pack boundaries; no identity, safety, or policy mutation is implied unless explicitly reviewed.
• Versor closure: runtime field paths must preserve versor_condition(F) < 1e-6; this ADR does not authorize hidden normalization or hot-path drift repair.
• Reconstruction-over-storage: evidence must remain reconstructive and content-addressed rather than duplicating opaque state.
• Replay-equivalence: serving, teaching, promotion, or checkpoint changes require a named deterministic replay / byte-equivalence gate.
• Mutation standing: any durable corpus, pack, policy, or epistemic-status mutation remains reviewed, proposal-only until accepted, or proof-carrying as applicable.