- golden-corpus.json: 75 hand-written pinned cases (D-13) — landmark 17 /
abbreviation 15 / glued 7 / typo 14 / alias 15 / negative 7, incl. two
near-miss negatives (Penang 0.625, Kampung Cina ~0.58) below the 0.72
admission line; threshold.top3_hit_rate pinned at 0.9
- gazetteer-corpus.mjs: the D-14 runner — imports the SAME src/index.ts the
bot uses, measures top-1/top-3 + per-category rates + avg ms/query, prints
the refusal margin and the fixed escalation ladder on failure, exports
measureCorpus() for the in-process suite, exits 0 only on threshold + zero leaks
- corpus.test.mjs goes green: CO1-CO8 (the Nyquist gate inside npm run search:test)
- delegates: search:corpus (root) + corpus (search workspace)
Measured: top-3 68/68 (100.0%) >= 90%, negatives refused 7/7, margin 0.643 < 0.72 —
threshold met on the first run (D-14 lever: None).
- CO1 top-3 >= corpus threshold (D-14), measured in-process via the runner's measureCorpus
- CO2 zero negative leaks (honest refusal, D-13)
- CO3 tautology floor: >= 15 non-negative queries not string-equal to any alias/name
- CO4/CO5 every expectId exists in the gazetteer + start_param-safe charset
- CO6 size >= 60 + per-category floors (landmark 12 / abbreviation 15 / glued 6 / typo 12 / alias 12 / negative 6)
- CO7 per-category rates printed so suite and runner output agree
- CO8 refusal margin pinned: every negative's best two-level fuzzy sim < 0.72, one near-miss in [0.55, 0.72)
All 8 cases verified failing (ENOENT corpus fixture) before GREEN.
- G1-G6: dataset invariants — schema, slug-id uniqueness, canonical-name
law ('KLCC' is an alias, never a name), DUKE-as-alias, KL bbox, the
ODbL header traveling inside the JSON
- G7-G9: the licensing gate — verbatim GPL-3.0 LICENSE, LICENSES.md,
the README no longer claiming 'all rights reserved'
- G10-G11: no logging/storage anywhere in the module (D-15); the loader
builds its index inside a function, never at module top level
- L1-L3: the phase's top mechanical risk — bare-Node loading with the
SRCH-01/negative probes, the static JSON import law, esbuild INLINING
the dataset + the wrangler dry-run build guard
13 cases fail while the dataset, loader, LICENSE files, and README fix
are absent — the RED gate.
- distance.ts: hand-rolled 3-band OSA, adjacent transposition = 1 edit,
zero imports (D-07); the row rotation is pinned by the D1-D9 vectors
- score.ts: TIER bands strictly disjoint descending — EXACT 1.00 /
ALIAS_EXACT 0.90 / PREFIX 0.75 / TOKEN_SET 0.65 / FUZZY as the 0.60
CEILING (a fuzzy score is fuzzySim * 0.60, never a 0.00 multiplier);
FUZZY_ADMIT_SIMILARITY = 0.72 exported for source.ts + the corpus
runner; token-level residual lets 'Sentrl' reach 'KL Sentral' while
'Jln Tun Razak' cannot reach 'Jalan Ampang'
- scoreCandidate returns the RAW matched alias (corpus provenance);
an exact ALIAS hit rides ALIAS_EXACT so a name-exact outranks it
- score.test.mjs S7 corrected during GREEN: admission belongs to
source.ts — scoreCandidate asserts the sub-floor similarity instead
D1-D9 + S1-S7 green; typecheck clean.
araharah-executor