• The planets sit within 0.40 arcseconds of JPL Horizons DE441 across 1900 to 2050, checked at fourteen instants. Before 11 September they sat up to 580 arcseconds off, from truncated tables, no light-time correction and the Sun's aberration applied to every planet. No earlier test could see it, because each looked at one date.
• The true lunar node was a five-term approximation, up to 12 arcminutes from the real osculating node. It is now solved from the engine's own Moon, within 31 arcseconds of Swiss Ephemeris. The mean node, which is the default and KSK's practice, was right all along.
• All twelve Placidus cusps and the ARMC agree with Swiss Ephemeris to 0.05 arcseconds. Eclipses are solved, not screened: every one of the 50 in NASA's 2000 to 2010 canon classifies with its type, and the eclipse tool now returns local circumstances.
• A planet or cusp sitting within a few arcseconds of a sub or sub-sub boundary can change lord after this. Readings made before 11 September may differ at that level.

True Chitra is computed from Spica now. Until 13 September it returned Lahiri's value, so the engine offered six ayanamsas and computed five. Spica's apparent place is checked against Swiss Ephemeris at five epochs to 0.32 arcseconds, and True Chitra runs 18 to 64 arcseconds below Lahiri across the era. If you chose true_chitra, your positions move by that much.

• get_frame_spread builds the same chart under every ayanamsa and names exactly which positions change lord and which hold. The frames disagree by a fixed 5.6 arcminutes, so the set that moves is small: about 6% of cusps and planets at the sub level, 56% at the sub-sub level, and roughly half of charts are frame-independent at the sub level and can say so.
• get_paradigm_panel reads one matter through all seven systems as parallel arguments, never as a vote. Pass frames and it says which verdicts change with the zero point alone. Measured over 120 panels, one in five has a verdict that does.
• A conclusion that changes with the ayanamsa is a property of the frame, not of the chart, so it ships as both readings with the frames named. Nothing here is scored, averaged or ranked.

• A denial house never denies on its own. DENIED means the cuspal sub lord reaches none of the required houses. 22.8% of verdicts left DENIED when this was corrected, and none entered it.
• Every event rule now carries the page it rests on. 16 of the 78 rows shipped a house group no page states and were corrected from the books; two had denied on a house the corpus requires. 12% of verdicts moved.
• Dignity, retrogression and combustion are reported and gate nothing, because KSK rules all three out by name for natal work. Retrogression carries forward as a delay signal to the timing phase, where KP does use it, and the delay's lift date is returned.
• A horary number fixes the Ascendant at the commencement of its sub, and the other eleven cusps are solved back from it. A rounding error had put 77 of the 249 numbers one sub lord off.

Every claim names its source. A response carries the rule it applied, with the text and the locus, in one of eight source classes from corpus (the page is quoted) down to unattributed (nobody has looked). 51 of the 78 event rules quote their page; the 19 that cannot say so on the wire. Verdict scores that are conventions, not book values, are labelled as such.

18 new tools, grouped by lineage. Every non-KP tool says which system is answering in its first line.

• Tajik has its own judgement mechanism now: the sixteen yogas with their orb table, panchavargeeya bala out of 80, harsha bala out of 20, the hadda table, and the mudda and patyayini dashas.
• The year lord's strength tie-break runs, so 35.5% of years stop being undecided, and where the text leaves a rung open both readings are returned.
• get_tajik_variant_spread does for Tajik's strength systems what get_frame_spread does for the ayanamsa: it says which conclusions move between the variants the sources leave open.

• Shadbala is the BPHS procedure in virupas, and it says so where it disagrees with the index it sits beside.
• Vimsopaka scores all sixteen vargas on the classical ladder.
• Ashtottari and Yogini started the wrong lord and carried no balance at birth. Both are rebuilt on the verses.
• Rashi drishti was inverted for eight of twelve signs. The negative yogas panel was missing a row. get_multi_system_verdict's custom house_group path never worked. All fixed.

Nothing is deleted on the wire. Every response compactor is now under a recorded-payload test and may drop only what it declares. Five were found deleting claims on the way to you, among them the ruling planet roles, a tiebreak rank and the promise reasoning. All five are fixed.

The server speaks MCP protocol revision 2026-07-28 and still answers 2025-11-25 clients from the same handler, so nothing has to be cut over on a date.

• A developer portal at developer.lumin.guru: up to five named API keys, a monthly meter, usage analytics by day and by tool, and the call pack ledger.
• The free allowance is metered per account across every key you have held, revoked ones included, so a new key never means more free calls. Failed calls are free.
• Call packs come in any amount, starting at $1 for 400 calls. A key you delete stays deleted.

• The Claude Code plugin is at 0.7.0. Four writing registers choose how a finished reading is written: consulting (the default, the answer the way an astrologer says it across a table), practitioner (the full written verdict with every provenance field shown), teaching, and report. Each changes only how the reading is written, never what is computed, and the plugin fails its own validation if that sentence goes.
• A kp-build skill for building a product on the server: input classes, the call shape, metering, paging and the copy rules.

The benchmarks page at lumin.guru/benchmarks shows every figure on the site against the oracle it was checked with, and what is still open. The public lumin-evals repository carries the same sky fixtures, the scorecard and the findings index.