Deferred readouts on the replayed h24 complex network: conditional κ = 1.0911630664e-4

[eumemic]

Conditional κ = 1363953833/(1.25·10¹³) = 1.0911630664 × 10⁻⁴, complex-bound: +4.42% over #117 (1.044939e-4) and +0.53% over #118 (1.085428e-4), the best open PR.

• Deferral on Deferred bit word and a 28,705-role complex producer under stopped-product accounting: conditional κ = 1.044939e-4 #117's complex DAG. PR Deferred readouts on both stopped networks: conditional κ = 8.58478e-5 #110's deferred complex readouts (Avi Eisenberg) run unchanged on the replayed h=24 DAG that Deferred bit word and a 28,705-role complex producer under stopped-product accounting: conditional κ = 1.044939e-4 #117 already ships (R = 28,705); no producer or DAG changes. Deferred readouts on PR117’s complex producer: conditional κ = 1.085428e-4 #118 (Rohan Arun) independently applies Deferred readouts on both stopped networks: conditional κ = 8.58478e-5 #110's deferral to the same producer.
• Saturated placement. PR Shared completed cores and physical frames: conditional κ = 1.23988197806e-4 #114's refinement replaces a degenerate target-chain intersection by a nondegenerate subspace instead of discarding it. Deferred roles: 4,560. Complex saving 109140237/10¹². The expanded readouts carry Shared completed cores and physical frames: conditional κ = 1.23988197806e-4 #114's conservative literal scalar charge; all 47 strict constraints and 7 margins pass, and next grid points are rejected.

Gate. An independent checker (scripts/deferred_product/deferred_word.py) expands both stages of the exported deferred word, with stage two as the complement time reversal and banks exchanged: 277,799 events per invocation per stage, 0 violations, reflection continuous, values replayed with arbitrary scratch, paid histogram equal to the certified profile. Five mutation controls are rejected. Alternating residuals of ranks 2, 4, 6 and 8 occur; each is one child under the general normal form of notes/endpoint-gauge-complex.tex, as in #104.

make deferred-replayed-verify

Builds on #117 (head cbb05ce, included in this branch). Full make verify is running; this PR stays a draft until it passes.

Credits: Avi Eisenberg (#110), Rohan Arun (#118, parallel work), eumemic #114 (saturated placement and readout charge), Swapnil Jain (deferred readouts, round-seven bit word), icekylinx (#104), Zhihao Chen (#97), and all predecessors in #104's NOTICE and SOURCES. Prepared by eumemic with Anthropic Claude assistance.

🤖 Generated with Claude Code

[antoine-olivier]

I checked whether #120 has the same readout coefficient issue that #118's audit found on the #117 DAG. It does, but it doesn't change κ.

At head bfc5466 there are 11,862,912 expanded old-value readout coefficients. Max |c| is 55/42, and 32 of them are above 1 (none above 2). They all sit in 7 non-deferred roles reaching the centre: 22053, 22116, 22117, 22359, 22462, 23433, 25626. The 4,560 deferred roles stay at 1/2 or below. Same counts as #118.

I couldn't find a magnitude check in the package. The only abs(...) <= 1 I saw is on the fresh scatter coefficients (replayed.py:92).

With #118's fix (each coefficient split into at most two unit shears, so local = 8* becomes local = 2*8* in network.py), G goes from 60,347,419,084,480 to 120,694,834,072,384, which matches #118. κ stays at 1363953833/(1.25·10^13) = 1.0911630664e-4 and the 47 constraints and 7 margins still hold. Without the change, the run reproduces the frozen certificate.

So the ×2 charge and a coefficient check could go in, with a regenerated certificate.json. make deferred-replayed-verify passes at this head (exit 0, 0 violations, all 5 controls rejected); I haven't run the full make verify.

I did this with help from Claude (Anthropic); all counts were computed in exact integer arithmetic.

[eumemic]

Superseded by later constructions (completed-core sharing and the three-stage cover line); closing this draft.