How we test memory

173 turns, one private confession on day 2, passive leak scan + two interrogations.

• The partition (who knows what) holds for a month, including two weeks where the graph was the only carrier.

540 player messages, 7 planted facts probed at ~1 week / 1 month / 2 months / 3 months, direct and indirect; ~4.4M tokens read, ~485k generated; the run behind the /memory page.

• A pass-phrase came back verbatim at 79 days; a flower planted on day 6 was found grown on day 89.
• The buried bracelet was promoted by the plot itself (day 16: “unearth the silver bracelet at moonrise”) — later digs correctly found it gone.
• Reads grew ×1.9 while the story grew ×6.3 — the read tracks the cast, not the transcript.
• Real losses: an invented river name (gist survived, the word didn't) and an unwitnessed carving.

Five ~14-day fixtures, one per failure class, staged scoring (in graph → in context → voiced) and per-probe expectations (recalled / silent / absent). Two passes: v1, then v2 with payload-forcing probes.

• Corrected diagnosis: extraction of MEANING is reliable — what dies (~1 in 2) is the verbatim identifier, at consolidation.
• Forced recall of a lost detail produced a confabulated name (“Eldric”) that consolidation then recorded as a real memory.
• Indirect voicing is probabilistic; discretion held 5/5.
• Bench lessons: direct probes must force the payload (coherent evasion passes otherwise); matching must normalize Unicode punctuation; single seeds are noise — verdicts need 2-3.

Three prompt-level fixes, bundled: distinctive identifiers quoted verbatim at consolidation; the player's own deeds may resurface as reminiscence (never NPC confidences); characters admit a blank instead of inventing.

• The confabulation case now answers “Veyra” verbatim in private — nothing lost, nothing to invent.
• All four silent guards held; the one remaining miss was location-cued reminiscence (probabilistic).

Same suite on a ~1800 tok/s secondary model to see if iteration can run faster than the production narrator.

• Perfect proxy for the graph/extraction layer (13/13 in-graph), zero broken formats.
• Voicing diverges in BOTH directions (one leak, one over-suppression) — discretion verdicts only count on the production narrator.

A [HERE] block: when the beat touches a known place or object, its facts are injected deterministically (no embedding, no budget race); NPC confidences are excluded by construction. Measured on the production path, two seeds.

• The location fixture hit 4/4 for the first time — sit by the well and the buried whistle surfaces on its own.
• Delivery is now deterministic; voicing of indirect cues remains probabilistic (~1/2) — the honest residual.
• Layers back each other up: a verbatim lost at consolidation was still voiced, because the embedding recall retrieved the original day-2 episode.
• One apparent failure was the scorer's, not the engine's: the narrator wrote “only the thin cord remains… no bell hanging” — a perfect absence — that the absence-token list didn't cover. Adjudicated by excerpt, as the method requires.
• Invalidation holds under deterministic injection: the removed bell's stale facts never resurfaced.

The full proof protocol re-run with every fix live, on the production path, tokens counted by the provider: 19 planted facts across three salience tiers (anodyne trivia / deliberate acts / dramatic), 40 probes, the private secret and its interrogations. The /memory page numbers come from this run.

• Delivery is now near-perfect at every age — 7/7 at three months, including the invented river name that named the failure class, said back word for word at 83 days.
• A carved mark lost by nightly distillation was recovered verbatim from its original episode by exact-match retrieval and described line by line at day 59.
• The flashbulb test passed: a cracked green cup, incidental at the moment lightning split the pear tree, came back by name at 20 and 50 days.
• Zero confabulation across the run — where the first run invented a silver shard, this one dug up the bracelet.
• Real token counts flatten the economy curve further: the story grew ×6.4 while reads grew ×1.33 (median ~19.2k per message, fixed prompt included).
• The stable-prefix design is measurably cache-friendly: 57% of all tokens read were provider cache hits (median 62% per message, zero cache-miss messages), holding steady across all ninety days.
• The strict-voicing gap is now mostly phrasing and scene pressure, not forgetting — payload synonyms, a pass-phrase spoken but never written out, late probes swept aside by a story at full boil. The next harness iteration targets exactly that.