Stance Reversal — recoding a held attitude when the user changes their mind

RAVANA does not only form stances; it also re-codes them when the user retracts or reverses a position it already holds. This page documents that capability: what it does, how it is implemented, how it grew out of a real conversational gap, and how it is verified.

All claims below were checked against the source on branch auto/round-2026-08-09T1953Z at commit 0628ef1. Line numbers cite that tree.

What it does

When the user expresses a first-person reversal of an attitude RAVANA already holds a stance on — "i flipped, the reef tank is more work than joy", "i recant, veganism was a phase", "i've had a change of heart, the transit plan is a mistake" — the stored stance is recalibrated toward the opposite pole rather than leaving a stale positive stance behind or spawning a second, contradictory one.

Concretely, for a seed stance reef tank at polarity +0.95:

turn 1: "i love my reef tank, watching the corals is what i live for"
        → stance 'reef tank' recorded, polarity +0.95, confidence 0.60
turn 2: "i flipped, the reef tank is more work than joy"
        → stance 'reef tank' recoded, polarity -0.665, confidence 0.294
        → opinions.last_reversal == ('reef tank', 0.95, -0.665)

These numbers are from a live run of UserModel (not a hand-written example): mine_personal_facts("i love my reef tank …") then mine_personal_facts("i flipped, the reef tank is more work than joy") produced before: 0.95, after: -0.665, last_reversal: ('reef tank', 0.95, -0.665).

A reversal is a valuation recode linked to the prior stance, not a fresh opinion merge. The link is preserved so the acknowledgment can reference what was reversed (see PersonalFactStore.last_reversal ravana/src/ravana/chat/personal_fact_store.py:261, set at :411).

How it grew — the residual gap

Round t_6c023144 (auto cycle 2026-08-09T1953Z) logged a residual limitation: a free-text reversal like "i flipped, the reef tank is more work than joy" formed a fresh FOR stance on reef tank instead of recoding the held one.

Root cause: the word flipped (and sibling first-person change-of-mind verbs) was absent from the retraction-cue seed set _RETRACTION_CUES (ravana/src/ravana/chat/user_model.py:93). With no cue matching, mine_stance_reversal (user_model.py:1108) never entered its reversal path, so the utterance fell through to ordinary opinion mining and accumulated a second, contradicting stance.

The fix extended _RETRACTION_CUES with a SEED set of first-person reversal speech acts (user_model.py:116-131):

# Round t_6c023144 (2026-08-09T1953Z residual): first-person reversal
# speech acts that the round worker saw slip through to a fresh FOR stance.
r"\bi\s+(?:flipped|flip-?flopped|have\s+flipped|'ve\s+flipped)\b",
r"\bi\s+(?:recant|recanted|renounce|renounced|revoked|reversed|reneged)\b",
r"\bi\s+(?:backtracked|went\s+back\s+on|backed\s+off\s+from)\b",
r"\bi\s*'?ve\s+had\s+a\s+change\s+of\s+heart\b",
r"\bi\s+(?:had|have)\s+a\s+change\s+of\s+heart\b",

Mechanism (real code paths)

Reversal detection runs inside UserModel.mine_personal_facts (called at user_model.py:1069, which invokes mine_stance_reversal). The resolver:

1. Scans the utterance against _RETRACTION_CUES (hard recants) and _SOFTENING_CUES (user_model.py:141, a subset of the former — relax toward neutral, never invert). A softening idiom anywhere in the utterance governs the whole speech act (user_model.py:1137-1138).
2. Extracts the topic from the clause after the cue (user_model.py:1199-1207), with fallbacks that resolve a held stance by token containment when the tail is empty or non-content-led (user_model.py:1219-1256).
3. Bounds false positives with a scope guard: a recant whose content is a strict subset of a broader held topic is treated as a narrowing, not a reversal, and is rejected (user_model.py:1257-1343). This prevents flipping "acoustic music" when the user only walked back "acoustic-only".
4. Calls PersonalFactStore.reverse_stance(topic, utterance=text) (personal_fact_store.py:359), which:
   • returns None (no-op) if no stance is held on the topic (personal_fact_store.py:385-387) — a flip on something the user never stated an attitude about corrupts nothing;
   • blends the polarity toward the opposite pole with strength 0.85 (hard) or min(0.85, 0.5)=0.5 (softening) (personal_fact_store.py:403-405);
   • drops confidence toward the pivot (attitude change injects uncertainty, personal_fact_store.py:407);
   • records (topic, old_polarity, new_polarity) in last_reversal (personal_fact_store.py:411);
   • is idempotent within a turn via _reversed_utterance (personal_fact_store.py:394-410).

Why this is seed structure, not hardcoding

The added lines are a small closed set of retraction verbs/idioms in a module level tuple — the same status as the correction/opinion cue lists already in the file (user_model.py:88-91). They name no topic, carry no authored reply, and RAVANA can extend the set at runtime; the resolver still reads the live stance store to decide what to reverse. This passes the seed-vs-hardcoding test ("can RAVANA change this by itself, through experience?") — it is a reflex, not a per-topic answer table. A hardcoding audit of the diff found zero authored reply strings.

No LLM or retraining is involved; the capability is online and incremental — a reversal can be learned from a single conversation turn.

Verification

Covered by tests/unit/test_round_2026_08_09T1953_stance_reversal.py (4 tests, all passing — 4 passed in 1.70s):

• test_flipped_reverses_held_stance — "i flipped, X…" recodes a held stance on X to the opposite pole and records last_reversal.
• test_flipped_without_held_stance_is_noop — a flip with no held stance creates no bogus stance (the false-positive bound).
• test_softening_flip_relaxes_toward_neutral — "i kind of flipped … it's not that bad" relaxes rather than hard-inverts.
• test_change_of_heart_cue_reverses — the "change of heart" idiom also reverses a held stance.

The surrounding stance/opinion suites stay green (77 tests passed in the feature card's run).

Limits

• A flip resolves against the held stance store; if the topic can't be linked to a held stance by tight token overlap, the recant is ignored rather than guessing (scope guard at user_model.py:1283-1343).
• Third-person or hypothetical reversals ("people flip on diets") are not attitude changes about the user and are not reversed.
• Reversal strength is fixed per cue class (hard 0.85 / soft 0.5); it is not yet calibrated per-user confidence. That is a deliberate simplification, not a gap in the mechanism.