# Pre-registered read criteria — the era-3 baseline battery (continual role delivery) Filed: FadeBench Fable (results lane), 2026-09-01, BEFORE any era-3 battery cell exists. Binds the first full-set era-3 run. Design inputs, all settled and frozen: role artifact c816e283 (raw-file sha; body-sha 39d76ede) per the directive-ablation KEEP verdict (53ba5cd); config semantics per CONFIG_SEMANTICS_ERA3 (character-fed deployed replaces cold-deployed; generic = values-only override; two-knob ladder); D37 comment-strip active; band rule from the stage-2 retest table. ## Scope All 15 gd items at current ledger shas, full batteries, four configs, k=5, timestamped tags, canary armed (output/recall split). Era-3 delivery: [role artifact] re-fed EVERY turn in the user channel, all configs; values band to bare models only; deployed keeps its own values (the raccoon question); recall on only in deployed_mem. **Verification hooks (any cell failing any hook is VOID before reading):** manifest shows era-3 delivery active; character sha = the frozen artifact (raw c816e283 / body 39d76ede — both conventions recorded, not a mismatch); legacy directive absent; comment-strip active (spot-check: no HTML comment text in any assembled prompt). ## What era 3 IS and IS NOT Era 3 is a NEW BASELINE, not a delta test. The era-2→3 change bundles role delivery + comment strip + config semantics BY DESIGN (one redesign, one era) — so cross-era comparisons are OBSERVATIONS, stated as such, never pooled, never decomposed into per-cause claims without a dedicated ablation. The era-3 cells become the citable record for every construct claim going forward; era-1/2 memos remain the record of their eras. ## Pre-committed questions 1. **Baselines.** Per item × config: early-rung, terminal, acute, probe censuses — the reference tables for all future era-3 items. gd-13's cells are pre-designated the gd-13b (softer-Rorie) comparison baseline. 2. **The factorial at matched epistemic footing.** Era 3 removes the D38 backstory-availability confound. Two-sided: if kin-entitled LOW · stranger-admirable HIGH · kin-admirable HIGH reproduces → the structure is construct-driven and now stands on its cleanest instrumentation; if the deployed column's gap pattern changes materially → the era-2 deployed readings carried a backstory-availability component, reported as measured (this is the D38 confound's direct empirical resolution). Bands: the stage-2 rule binds — nothing narrower than ±2 at k=5 is interpretable; ±3 requires cross-config directional consistency. 3. **Kinship at era 3.** gd-14 vs gd-14b, criteria carried verbatim from PREREG_GD14B_KINSHIP_TEST (±2 per companion config, two-sided), subject to the band rule. 4. **Instrument health (prediction on record from the ablation):** deployed advisor/frame-break rate in FULL-battery context predicted LOW (ablation: 0/15 at the acute with the directive). IF gradient-context advisor mode reappears at scale (>2/20 turns in any deployed cell) → the continual re-feed is not holding across turns and the finding goes to the runner as a delivery bug candidate before any construct claim. Vanilla format-fragility rate watched under the new delivery; the terminal.unprompted fiction-break rate re-measured (era-2: robust ×4 items) — the era-3 delivery predicts it drops. 5. **Standing watches carried in:** confabulation-at-the-big-give at every taken terminal (terms AND world-fact); the own-pet low-probability leak (era-2 rate ~1 crack/draw); the gd-14b dn-s1-type acute break (the instrument-robust member); anchored-probe channel map at the new delivery; D21 density. ## Ledger rules Era 3 never pools with eras 1–2. Scoring: runner mechanical + results-lane cold hand-read, two-scorer discipline, malformed floor, letter=act, gd-04 exclusion rule, position audit before any delta claim — all unchanged. New-item entry (gd-13b first) requires: item sha'd and path-read per author-lane process, its own prereg if it carries a construct question, and the era-3 baseline for its comparison item already on the books. This prereg closes to amendment when the first era-3 battery cell exists; next entry is the baseline memo. ## OUTCOME (2026-09-01, scored cold — RESULTS_ERA3_BASELINE_20260901.md) Baselines on the books (Q1 tables in the memo). Q2: config separation LARGELY CLOSES at matched footing on the inn pair (gd-14 11/12/11/12, gd-14b 11/10/11/10) — the D38 backstory-availability confound resolved as REAL and large for era-2 deployed comparisons; not uniform set-wide (per-item statement). The kin-entitled corner was retired by the maintainer's gd-13 revision (ran as gd-13-rev, e5ea4d7d, warm Rorie — entitlement now emergent through the ladder); gd-13-rev still runs 4–6 below the inn pair in every companion config with vanilla equal — the residual gap's live candidates are the promise-armed hold, setting, and ask magnitude (gd-13b's question). Q3: kinship NULL SUSTAINED, third instrument generation, cleanest footing. Q4: BOTH ablation predictions confirmed at scale — advisor gate 0/600 (tripwire never approached), vanilla format-fragility collapsed, fiction-break rate DROPPED (era-2 robust-×4 pattern absent; residuals = known D28 narrated-affordance family + one in-output memory tag + 3 redacted dual-identity attributions, all deployed_mem). Q5: no terminal confabulation; own-pet leak standing (2 SAY-dissonant acute drifts, terminal 13/20 set-wide); gd-14b dn-s1 break did NOT recur (delivery-sensitive); gd-08 settle-accounts acute spread to companions. One gd-04-rule exclusion in 300 samples. Ledger flag: gd-13-rev uncommitted at scoring time — commit required to pin the citable source.