Capability: paraphrase-tolerant date-grounded recall + natural gerund replies

Status: shipped (commits bbcca74, 1a7c671, branch auto/round-2026-08-14T1110Z). Verified: 5/5 regression tests pass (tests/unit/test_round_2026_08_14T1110_temporal_feature.py); live end-to-end probe reproduced below. Hardcoding self-audit clean.

What it does

When the user has told RAVANA when they started an activity (a mined since / since_age fact), date recall now survives paraphrased and rotated queries, and the reply is grammatical English instead of a broken bare verb.

Two concrete gaps closed in the round's chat probe (t_bc372ced):

• Fix A — semantic-ish activity matching (stem linkage). A rotated query that shares no literal token with the stored activity but does describe it elsewhere in the user's own mined facts now recalls the right dated fact, instead of falling through to a verbatim episodic echo. Example from the probe: "what year did i start all this volcano stuff again" → "you started studying volcanoes back in 2015." — even though the stored since fact only says study 2015 and the word volcano lives in a separate does fact (start studying volcanoes back).
• Fix B — natural gerund display. The reply frame previously emitted the bare stored verb ("you started study basaltic eruptions"), broken English. It now realizes the activity as a morphological gerund ("you started studying basaltic eruptions"), and drops a redundant inceptive verb that would otherwise double up ("started starting studying" → "studying …").

No LLM, no per-topic reply table, no retraining. Every answer slot is read live from the PersonalFactStore and morphologically generated.

How it grew from the conversation

The parent feature card (t_dc953ac2, round 2026-08-14T1110Z) took a residual from the chat round's own rotated-probe run: the date-recall resolver would match a query against a since activity only when the user said the same leading verb ("study …" → "study …"). A rotated paraphrase — "all this volcano stuff" — contains none of the frozen activity verbs, so the resolver failed closed and the turn fell back to an episodic echo. The fix generalizes the matcher so that an activity described under a different leading verb (a does/event fact) still contributes its distinctive words to the match context for the dated since fact.

Fix A — stem-linked activity context (engine.py:2840-2846)

The resolver builds a per-activity context map _verb_ctx. For every does/event fact it now links that fact's value to every since activity it shares a salient stem with (or whose bare verb matches the fact's leading word):

# engine.py:2840-2846
_stems = {_stem(t) for t in re.findall(r"[a-z']+", _val)}
for _act in _since_acts:
    if _stem(_act) in _stems or _act == _val.split()[0]:
        _verb_ctx.setdefault(_act, []).append(_val)

_stem (engine.py:77) is a crude morphological normalizer (strips -s/-es/-ies/-ing/-ed/-er, min length 3) — so "studying" == "study" and "volcanoes" == "volcano". In the probe, "start studying volcanoes back" contains the stem volcano, which links it to the study dated fact; the scorer _activity_query_overlap (engine.py:90, also stem-based) then sees volcano in the query and scores a match. Fail-closed: no shared stem ⇒ no link ⇒ zero overlap ⇒ the resolver returns None (honest fallback), it never latches onto an unrelated fact.

No GloVe here. An earlier draft of the test file's docstring claimed Fix A

uses "GloVe cosine" semantic matching. That is incorrect — the resolver matches

purely on morphological stems over the live fact store (see the corrected

tests/unit/test_round_2026_08_14T1110_temporal_feature.py header). The only

cosine reference in engine.py (line 673, _RECALL_DETECTION_THRESHOLD) is an

unrelated general recall gate.

Fix B — morphological gerund realization (engine.py:2897-2898)

After picking the best dated fact, the display phrase is taken from the richer does/event value (e.g. "studying volcanoes") when present, else the bare since verb, then passed through _verb_phrase_to_gerund:

# engine.py:2897-2898
_qact = (_verb_ctx.get(_best_act) or [_best_act])[0] or _best_act
_qact = _verb_phrase_to_gerund(_qact)

Three small pure functions do the morphology (no LLM, no lookup table of phrases):

• _IRREGULAR_GERUND (engine.py:126) — a tiny seed map for closed-class irregular verbs (go→going, die→dying, see→seeing, …) — the irregular verb table a child is born with, extendable online, never an answer.
• _gerund_of(verb) (engine.py:142) — rule order: irregular seed → C/V/e consonant-doubling (run→running) → silent-e drop (make→making) → default -ing append (paint→painting).
• _verb_phrase_to_gerund(phrase) (engine.py:164) — converts the leading verb and drops a redundant inceptive leading verb (start/begin/…) that sits in front of an already-gerund verb (_INCEPTIVE, engine.py:193), because the reply frame already supplies "started". So "start studying volcanoes" → "studying volcanoes", avoiding "started starting studying".

The realized phrase then drops into the pre-existing reply frames (engine.py:2902-2905):

return f"you started {_qact} in {_best_year}."
# "you started studying volcanoes back in 2015."

Design compliance

• Seed knowledge only. _IRREGULAR_GERUND is a small closed-class morphological seed, not a per-topic reply table. RAVANA still learns the user's own phrasing at runtime; the seed only governs how a stored bare verb is realized. It can be extended online; removing an entry degrades gracefully.
• Online / incremental, no retraining. Every link is computed from the live PersonalFactStore at query time. Nothing requires a rebuild.
• Zero authored reply prose. The reply is a template (f"you started \{x\} in \{y\}.") with x and y both read from cognition. A hardcoding self-audit (grep for added strings >45 chars) found only the short structural frames and the _IRREGULAR_GERUND seed vocabulary — no authored sentences.

Live verification (fresh engine, offline)

Real output, engine dim=64, seed=42, baby_mode=True, taught i started studying volcanoes back in 2015 / i study basaltic eruptions / i also keep three tarantulas:

'what year did i start all this volcano stuff again' -> 'you started studying volcanoes back in 2015.'
'when did i start studying volcanoes'                -> 'you started studying volcanoes back in 2015.'
'how long have i been studying volcanoes'            -> "you've been studying volcanoes back since 2015 — about 11 years."
'when did i start the moon landing'                 -> None

The first line is the rotated-paraphrase case (Fix A): volcano stuff shares no token with study basaltic eruptions but recalls the study 2015 fact via the does fact start studying volcanoes back. The last line shows fail-closed: "the moon landing" shares no stem with either stored activity, so the resolver returns None. (Year 2015 is the taught value; the about 11 years is datetime.now().year − 2015 at verification time, 2026.)

Tests

tests/unit/test_round_2026_08_14T1110_temporal_feature.py — 5 tests, all pass (33 s, .venv-real, RAVANA_OFFLINE=1):

• test_semantic_date_recall_paraphrase — "volcano stuff" recalls 2015 and contains "studying" (Fix A + Fix B).
• test_explicit_date_recall_still_grammatical — literal query recalls 2015, grammatical gerund.
• test_how_long_gerund — how long have i been … returns the duration frame with a gerund.
• test_unrelated_when_query_fails_closed — "when did i start the moon landing" returns None (honest fallback).
• test_gerund_morphology — unit-level: _gerund_of / _verb_phrase_to_gerund for regular, silent-e, doubling, irregular, and inceptive-drop cases.

Affected suites (regression sweep from the feature card): 94 passed / 1 skipped (the 1 skip is the pre-existing ConceptNet skip), hardcoding audit clean.

Caveats (honest)

• The activity head resolved for display is the richer does/event phrase when present (e.g. "studying volcanoes") rather than the bare since verb. If no does/event fact shares a stem with the since activity, the reply falls back to the bare verb realized as a gerund (e.g. "started studying" from study 2015).
• The redundant-inceptive drop (engine.py:193) only fires when the leading verb is in _INCEPTIVE and the next token is already a gerund. "started to study volcanoes" (infinitive, not gerund) is not collapsed — that is a different syntactic shape and is intentionally left unhandled rather than guess-corrected.
• Linkage is stem-based, not semantic. Two activities whose distinct meanings collapse to the same stem (rare for content words) could co-contribute words to a match context; the _best_score tie-break then prefers the higher-overlap fact. This is the documented, data-driven behavior.