test_python_hygiene_budget forbids new root-level scripts/*.py; the brief named
scripts/muhurta_search_api.py. Same code at scripts/ondemand_calc/muhurta_search.py;
route, CLI import and tests follow.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
transit_date counted house_from_moon from the transit Moon when the caller gave
no natal Moon; the natal reference is now only what the caller supplied.
resolve_profile now refuses an unsupported ayanamsa name instead of passing it
to the engine.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The fixtures were captured on Windows; on Linux the Moon longitude differs in
the 13th digit (41.48538690748626 vs 41.485386907488675), so the exact-equality
freshness test failed. Rounding both sides to 9 decimals still catches any real
engine change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Base card gated by EVIDENCE_CARD_SPECS[domain].route: dasha systems keep Vimshottari/Narayana only, yoga detail only on yoga domains, annual brief only on the annual route; yoga row planets become names. Model-visible contract+card 46,664 -> 32,153 chars on average (-31.1%), max 54,532 -> 43,419 over 3 public charts x 12 domains. New size/route test; three existing assertions updated with old/new/reason notes.
Co-Authored-By: Grok (agent loop)
- career layer gets the D1 longitudes, so the paksha has one source (Obama
base vs career mismatch fixed)
- card carries movable_lagna_moon.reason, Narayana ad_status / ad_reason and
the mahadasha cycle_index; running line says "sub-period undecided"
- engine keeps the MD when the AD is undecided; timingFollowing skips
zero_duration rows explicitly
- timingKeys allowlist: adstatus / adreason / cycleindex / zeroduration
- fixtures + capture script, 16 engine tests, 7 card tests, PROGRESS, BUG-1812
Co-Authored-By: Grok (agent loop)
Owner 2026-11-09: the R2 widening is on hold until the astrologer rules on it.
STRENGTH_SCOPE_DEFAULT returns to seed_only_v0. All R2 code stays: local
propagation, the selected / equal / unresolved statuses, and
seed_ad_start_p43d_v1 as a named non-default profile. profile_config and the
cache keys record strength_scope = seed_only_v0. R1 and R3 are unchanged.
Tests that pinned the widened scope now pass it explicitly, with
原值 / 新值 / 原因 notes. Evidence-card golden recaptured (TTL=0; only the
profile labels changed).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Found in the round-7 cross-branch check: on public charts (Zidane, Obama)
the Neecha Bhanga row never reached yogas_detail. The compact detector entry
dropped formation_evidence, and the row sat past the 30-row cap. Keep the
witness fields when compacting and never cut a row that carries them (the
cap stays 30). The card's Narayana method_status also dropped the engine's
profile_config (timing key allowlist); carry it through.
Evidence-card golden recaptured with TTL=0 (Obama: row 30 is now Neecha
Bhanga); the load-dependent VedAstro closure reason is kept as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The 简明解读 list's grey line is the receipt row's result_line (fallback: what
the layer computed); firm → 已算, conditional → 有条件. One backstage filter
(isBackstageAuditRow: external / MEVG / real-case, plus provider, engine,
packet, calibration, Western) is shared by withTechniqueEvidence and the list.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Contract keeps BUG-1794 (no web / external / MEVG / real-case wording in either
mode) with BUG-1795's 深度解读 / 简明解读 names; the missing-audit-table repair
prompt keeps BUG-1794's "no verification rows" wording under the new name.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
r8 blind eval: 「这一层给出的结构与 D1 同向:土星在巨蟹」 in a D10 line was read
as a D1 claim (3 false mismatches). A layer token followed by 同向 / 一致 /
对照 … no longer opens its own span; the matched-claim note is only shown on
mismatches.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The high-rigor path overwrote can_apply with the technique contract's
confirmation_allowed, so a passing external validation could turn a
low-confidence or provisional engine result into an applicable one.
Both must now agree; otherwise the contract continues rectification
with local_engine_cannot_apply.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Raw totals are only compared when every candidate computed the same
(event, layer) evidence units. Otherwise winning_segment stays the raw
leader (provisional_leader) and comparable_leader re-ranks over the
units all candidates share; a missing unit is dropped for everyone,
never scored 0. The live stage Python engine now skips an event whose
Narayana is unavailable instead of scoring it 0, repeated facts count
once, and the stage labels an unequal-coverage leader as provisional.
占星师 2026-11-08 验收清单 §6 / §9 / §10.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
- R1: default AD direction rath_p50_md_parity_v1 (start parity kept as a research
profile); full profile config + algorithm version on every Narayana output, cache
keys (consult, api chart, probes, long report, synastry) and the rectification run.
- R2: one strength profile at the seed, AD start and p.43 (d); selected / equal /
unresolved kept apart; unresolved nodes block only what depends on them (seed
candidates kept, dated MDs before an unresolved length kept, AD-only loss in
rectification keeps the MD).
- R3: Neecha Bhanga witness fields (reference, relative_house, self_reference,
independent_confirmation_allowed, self_reference_only) from the main detector to
the evidence card yogas_detail rows; self-reference-only never counts as support.
- Sync her b54812c3 KP clock validation; port her round-5 / round-6 tests.
- BUG-1783..1787; golden recaptured with TTL 0.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
产品 2026-11-08 决定,偏离占星师「缺口不能隐藏」裁定(第五轮裁定 6 / 8):产品
选择不向用户展示;后台保留记录. The plain-mode gap sentence (BUG-1757) and the
professional audit table's four verification rows are no longer asked for;
the receipt's MEVG / Real Case Calibration rows are unchanged, and 「不写已验证 /
已校准」 stays.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
产品 2026-11-08(简明解读「解读依据(N 项)」): rows the evidence card speaks for
get techniqueAuditTable[i].result_line (one plain line built from card fields
only, never by the model) and display_status (firm / conditional from the
two-axis interpretation_status; Narayana with a method dispute -> conditional).
Provider, MEVG, calibration and Western rows get neither.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
占星师 2026-11-08 第六轮 §5: share of evidence-requiring judgments that trace
to the right layer, entity and reference, replacing raw counts. Same reader
as the runtime guard; reads the card the model saw from the payload.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
占星师 2026-11-08 第六轮 R4 / V6: a sentence whose explicit layer / entity
claim (varga sign, lagna, house lord, D1 house with 本命, Karakamsha sign /
house / lord / lord house, Arudha sign / house / lord) does not match the card
is rejected after the answer finishes, regenerated once, and dropped if the
rewrite still mismatches, adds an execution status or breaks the table. The
corrected answer replaces the streamed one; only validation steps are recorded.
Anonymous minimal counterexamples for the three r6 real errors fail on the
previous code and pass now. Harness records the guard and the pre-guard text.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
占星师 2026-11-08 第六轮 R3 / R4:
- independence gains derived (derived fact -> dependency); the Karakamsha
chain (AK D9 sign -> D1 projection) is derived, its lord's placement a
separate raw fact; evidenceCardFactRecords expands every item to the ten
fields, raw -> relation / derived -> conclusion.
- base.natal.jaimini.karakamsha carries karakamsha_lord_natal_house and a
reading line (113 of 210 blind answers copied the sign's house to the lord).
- yoga self-reference fields (self_reference / independent_confirmation_allowed
/ self_reference_only) read defensively: no support vote, her wording.
- runtime contract: plain layer names, mismatch rejection, self-reference rule.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The backend capability audit lists "Case Validation" as user-visible when
frontend source contains the gate's acronym; the row filter, comments and
the dev preview fixture now avoid the literal token.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Owner-approved chat answer-mode UI (2026-11-08):
- Modes named 深度解读 (professional, default) / 简明解读 (plain) in the UI,
the model-facing mode line and the runtime contract (names only).
- 解读方式 pill left of send (常用问题 pill look); menu reuses the preset
popup rows; 功能吉凶口径 (PVR 表 30 / BPHS 例外表) under 深度解读 only;
Drawer on phone. Remembered per user in localStorage; applies from the
next question; 重新生成 uses the current choice.
- Mode resolution: this question's words -> pill -> 深度解读; history no
longer sets the mode. Request carries answerMode (+ functionalProfile
only under 深度解读); engine gets analysis_mode / functional_profile.
- Receipt records answerMode / functionalProfile; 简明解读 answers show a
collapsed 解读依据 hairline list built from the receipt audit rows.
- External verification rows (联网 / MEVG / 真实案例) are dropped from the
rendered 深度解读 audit table and never listed.
- Composer stacks when the single-row field would be under 64px (320px).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8