A self-chosen side removes the line from the segment, so the saved record kept
nothing about what the other side changes. The decision now keeps each chosen
line with the charts that change across it, and the report's scenario statement
lists those domains. Grade caps, report text and the writing prompt are unchanged.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
A saved report draft lost all nine sections to claim_binding_missing although
each judgement carried claims, a grade within its cap and allowed refs.
- rebind: drop a non-fact `kind` label from a paragraph that carries claims,
and `refs` from a pure fact paragraph, before the strict parse; both logged
in auto_rebound. Every other check still runs.
- TIMING_TEXT: four decimal digits (15.0627) are not a year.
- dignity check: 「入旺星座」/「落陷星座」 names a sign.
- evidence grade: Ashtakavarga is a checked (never required) layer for
character, relationship, health pressure and timing claims, per the shared
baseline in strict-workflow-router.md §2. Caps unchanged.
Replay of the saved draft: 0 → 7 of 9 sections kept. Fresh run: 9 of 9.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
- scrub evidence-packet key names (upstream PRIVATE_EVIDENCE_TERMS) from reader text before the gate
- gate accepts "line label + later clause" as a faithful copy; practice failures carry item index and
the source line, and the program rebinds or drops the item after the rewrite round
- rule-built must_use_pair_facts (tight Moon/Saturn conjunctions, Nadi 0, Sun in the other's 8th)
in the thesis, emotional-safety and long-term sections, plus one prompt line
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Tolerances: jd_ut rounded to 1e-6 day (0.086 s) for the Kalachakra fixture; Muntha
ingress instants within 2 s (observed drift ~1 s between machines; a real change in the
search moves them by minutes). Every other field stays exact.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXp2Vp62Ur3C7yRWQdSq
After BUG-1989 a section read is up to ~16k characters. runSynastryReportTurn appended every
readSynastrySection result and threw synastry_chat_context_too_large once the total passed
SYNASTRY_CHAT_CHAR_BUDGET (20,000); the rewrite round re-reads the same section, so the
overview question on the public pair hit 24,693 and the route answered 503 without a trace.
Repeated identical reads now return the material already given; a new read is fitted to the
remaining budget. The route's catch records one error line and one failed observability
event with machine values only.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The main detector forms a cancellation for each debilitated planet with the same name and
rule_id. The detected-list merge, the route sieve and the yoga layer each treated the second
planet as a repeat, so Einstein's Moon kept its status label but lost its row, id and edge.
Per-object rules now dedupe by (rule_id, planet): instance key and evidence id name the planet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Evidence-pack step 2, one-topic vertical slice (owner 2026-11-18; the astrologer's 11-18
review §二.4–二.6, §三 questions 1 and 3, §六 step 2).
- The evidence base gains the one relation the data states with a named rule: the Neecha
Bhanga main detector cancels the debilitation effect of its planet (the placement stays),
and depends on the witnesses it cites. No weakens edge: there is no named rule for one.
- lib/evidence-pack.ts builds the wealth pack from the card's own indexed facts: direct
evidence; limits (cancellations with their witnesses, engine-stated yoga reductions,
coexisting pressure); time (only when the question asks about time, range from the
question); gaps and disagreements. The dependency closure follows every level with a
visited set, cycle record and budget; a stopped walk is a visible gap.
- CONSULT_EVIDENCE_PACK_DOMAINS (default off; wealth on staging only) swaps the wealth key
block for the pack; off, the model view is byte-identical. The time scope comes from the
split question, never from user feedback (test E).
- The backtest harness reads a whole facts packet; timing.horizon is registered as non-fact.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The card's model view gains evidence_index: the snapshot id and coverage, one id
template per fact path, status codes with a legend and the per-id exceptions.
Positions, houses, roles, afflictions, Arudha and derived houses, points, SAV,
KP, Argala, transits, double transit, Saturn cycles, the annual chart, every
dasha level, judgment and participation rows now resolve; rows that name their
own id keep it. Branch rows carry their #br: id from the facts packet. A
coverage check lists any card leaf that is neither a fact nor marked not
citable; it is empty on all public golden cards.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Per the astrologer's 2026-11-18 review of the evidence-pack draft (§二 1-5, §五.3, §五.5):
logical_key / snapshot-scoped evidence_id / content_hash; birth-branch ids bound to the parent
range and intervals; snapshot over opaque subject, birth input revision, calculation profile,
rule versions, retractions, observation instant and coverage (calculation time recorded only);
execution / formation / interpretation / source status; base-level edges with type, scope,
conditions, method version, basis and review status (engineering dependencies only);
dependency_closure with visited set, cycle report and budget; interpretation_key; a parseable
synthetic fixture and the review's test list.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Evidence-pack step 1 (owner 2026-11-18). One ASCII form, ev:<family>:<body>, for D1 and varga
placements, AL/UL derived houses, special points, Shadbala, dasha periods, transits, double
transit and yoga rule instances; an evidence base with links taken only from what the data
states; a snapshot id binding the birth input revision, calculation date and rule versions.
The internal-word check drops ids from the basis section and logs them in the body.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
(a) formed / disabled / observation (the yoga rule states on the card since
BUG-1967) and key: true|false fragments are now internal words; the table uses
only existing product wording (computed and conditions_not_met yoga texts,
未执行, 已停用, 燃烧 / 未燃烧, 逆行 / 顺行), a one-word key set to true keeps
the term, anything else is dropped. Offline rerun over 63 saved answers: 5 new
basis hits, all the two real leaks, no new false positives.
(b) The relations line states AL's 2nd / 11th (and UL's 2nd for marriage
continuity) against the natal house of the same number: same sign or not, and
where the natal lord sits. No voice rule added.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The chat now carries the birth window the personal report reads
(consultationBirthTimeWindow). The facts packet samples the chart every
60 s across it, bisects each change to 1 s and recomputes every dependent
field per branch (scripts/consult_facts/birth_range_branches.py). Without
a declared window, or with only the default 30-minute trial, the rows stay
not_evaluated; a sample that does not reproduce the card is not branched.
A self-chosen side only orders branches. The key block lists the branches
and the relations line states what they share first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The marriage card reads the question by fixed Chinese and English patterns
(consultation-marriage-intent.ts, no model call). For marriage_continuity
UL's 2nd joins the key block and its relations line (role main); for any
other marriage or love question it stays in the participation rows as an
auxiliary layer with an analysis-only note, per the astrologer's
2026-11-17 batch-2 item 6.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The astrologer's 2026-11-17 batch-2 reply (section 3, items 2, 3, 4, 6) asked that
the chat card stop treating the first 30 yoga rows as the whole list. The owner
approved editing scripts/consultation_native_layers.py for this item.
- Native layer: no row cap; each row has rule_id (engine id, or a deterministic
derived:<category>:<name>), method_version and rule_state (formed / candidate /
not_formed / not_executed / disabled; observation for structural rows); the
layer adds rules_version and not_formed / not_executed / disabled / superseded
lists. Same-named rules with different ids are no longer merged; an id-less
chart.yogas row adopts the engine's id. Main detectors list their reference
rules (parent/evidence, not one id). Parvata is disabled_source_mismatch;
retired legacy spellings (Voshi, Sunapha, ...) point to their precise rules.
- Card: selectEvidenceCardYogaDetail keeps the rows the card's domains read and
records selected_by; later-house lord rows are listed short; omitted is a
structured count; the full layer is the new yoga_conditions lookup section.
- Personal report keeps its previous rows exactly (report-side dedupe + cap);
its yoga field list now matches the card.
- Size budgets raised with recorded old/new values and reasons.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Her second-batch items 4-6: main evidence of their own view, parallel to the natal
houses. Derived product-side from the packet's AL / UL sign (D1, inclusive count;
AL in Pisces -> Aries): target sign and natal house, occupants, lord and its natal
place, aspects, rule source. A missing start stays 派生未计算. They join the wealth /
marriage key block and its relations line; the report reading material treats them
as executed. The yoga card (30-row cap, rule ids) is generated in a frozen file and
is not changed here.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Shadbala reads 达到 / 未达到最低线 (her item 9; owner agreed) and the card carries
min_required for the 判断依据 line; required_strength_ratio beside ishta_bala_pct;
remedies no longer trigger from Shadbala (unaccepted_threshold). Education uses D24,
not D5 (item 7). Vighati keeps its position only (item 1); the 64th Navamsa from the
Moon and Pranapada take part as auxiliary layers with her fields (items 2, 3). The
annual natal two-Hora D2 gives Hora membership instead of the lagna lord's house,
and its birth-time stability check compares that membership (item 8).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Before/after on two public charts (4 answers each, DeepSeek flash): the
after answers cover study, moves, contacts and pace more often; no new
echoes, hedging lines or English leaks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXp2Vp62Ur3C7yRWQdSq