Capability: meta-identity reflection (answers about who you are to RAVANA)

Status: shipped (commits 3e9652f, f4682bd, branch auto/round-2026-08-15T1537Z). Verified: regression tests in tests/unit/test_round_2026_08_15T1537_meta_identity.py pass (2/2); live end-to-end probes reproduced below (real engine output, dim=64, seed=42, baby_mode=True, offline). Hardcoding self-audit clean.

What it does

When the user asks RAVANA to reflect on its accumulated model of the user — “do i seem like a real person to you”, “what am i to you”, “tell me something true about who i am”, “what have you learned about me” — RAVANA answers from its LIVE durable state, not from a biographical fact lookup (name/location) and not from an episodic echo of a stored turn.

The reply is composed entirely from runtime stores RAVANA grew autonomously:

• the user's real name (user_model.user_name), or an honest “i'm still learning who you are” when none is known (engine.py:2997, :3014-3017);
• the stance count + topics it has learned (UserStanceStore.stances, engine.py:3001-3002, :3029-3033);
• the fact count (PersonalFactStore.facts, engine.py:2999, :3020-3027);
• RAVANA's own identity strength + trend (IdentityEngine.state.strength / get_trend(), engine.py:3003-3004, :3006-3011, :3035-3037).

Real engine output (fresh persona corvin, taught “i love oysters” + “i think surveillance is wrong”):

Q: do i seem like a real person to you
A: i know you as Corvin. and from what you've told me i've picked up 2 stances
   you've shared and 1 facts about your life. you've let me see where you stand
   on things like oysters, surveillance. my own sense of self is still forming —
   my self-coherence sits around 0.25 and is holding steady.

The same answer renders for “what am i to you”, “tell me something true about who i am”, and “what have you learned about me” — because every word of content (except light connective scaffolding) is read from state, not authored.

Fail-closed. The meta detector is the first branch in _structured_recall (engine.py:2389-2398); it returns None when no meta signal is present, so:

• a plain biographical query — “what's my name” — is not intercepted; it still resolves on its own structured path (your name is corvin.);
• a query that matches no meta pattern stays on the normal honest pipeline.

No LLM, no per-topic reply table, no retraining. The user can correct any fact/stance and the stores merge on correction, so the reflected profile updates live.

How it grew from the conversation

The chat round that fed this cycle surfaced, among its residual limitations, a class of questions that were neither factual (name/location) nor episodic (repeat what I said) but metacognitive about the user themselves. The feature card (t_6fca4160, “Bug 5”) picked it as a concrete capability gap.

Root cause / prior behavior. Meta-identity queries were not recognised as a distinct intent. They fell through to one of two wrong paths:

1. the episodic cued-recall path, which echoed a stored turn verbatim (e.g. “you told me earlier …”), or
2. an authored “feeling-real” frame in response_gen.py keyed on the word “real” (\{subj\} is fuzzy for me...) — probe-tuned prose that was not derived from any runtime store.

Fix (commit 3e9652f). A single regex in _structured_recall (engine.py:2389-2398) detects the meta-identity intent and delegates to a new _meta_identity_reply (engine.py:2985-3039) that renders the reply from the LIVE stores listed above. The detection is fail-closed: it returns None for non-meta input, leaving every other path untouched.

Hardcoding audit (commit f4682bd). The authored probe-tuned “feeling-real” frame was deleted from response_gen.py. _meta_identity_reply contains no authored reply prose and no per-topic answer table — only connective scaffolding (and, you've let me see where you stand on things like) around slots that are read from state at call time (name, stance topics, counts, identity strength/trend). The 2 regression tests assert the deleted phrase (“fuzzy for me”) does not leak and that an episodic echo (“you told me earlier”) does not appear.

Where it lives (with line cites)

Test coverage

tests/unit/test_round_2026_08_15T1537_meta_identity.py (2 tests, both pass):

• test_meta_identity_reads_real_state — for each of the four meta phrasings, asserts the reply references the learned name (corvin), references a learned stance topic (oysters/surveillance), does not contain the deleted “fuzzy for me” phrase, does not contain an episodic echo (“you told me earlier”), and reflects identity state.
• test_non_meta_query_fails_closed_in_meta_branch — a plain “what's my name” is not intercepted by the meta branch and still resolves from the structured name path.

Run with:

RAVANA_OFFLINE=1 python -m pytest tests/unit/test_round_2026_08_15T1537_meta_identity.py -v