docs/HANDOFF.md

Handoff — Quantum Chemistry & Clean Energy (Hamiltonian Block Encoding)

Current state — all three benchmarks optimized (2026-11-04)

Latest goal achieved: all three kickmix circuits optimized and assembled in submission/ (invsqrt.kmb, dqi.kmb, dfthcsos/), mirroring sample_submission/. Every result below is from the unchanged verifiers with salts not used during development. Details: submission README.

Benchmark	Baseline score	This submission	Improvement	Verification
I. invsqrt beta/sqrt(x) phase	2,076,774.23	684,969.09	−67.0%	Official run, 25,067,616 shots (8/x), salt final-v2-8shot: PASS, worst error 0.661× tolerance (also PASS at 1 shot/x, salt test-v2-3, identical hash)
II. DQI RS syndrome decoder	24,439,463,583,672.93	17,673,729,048,937.30	−27.7%	Full official run, 102,400 shots, salt dqi-v6-full: PASS, 0 value / 0 phase errors; second salt (10,240 shots) and adversarial max-weight full-size shots also PASS, 0 errors. v1 was 20,679,670,765,943.19 (−15.4%)
III. DFTHC-SOS FeMoCo	3,355,599,588.32	2,309,230,123.93	−31.2%	Fresh salt final-fresh-check: PASS, 254,464 address-shots; error_2 = 19.31 mHa is the thinnest margin (gate 20 mHa)

Usable Python 3.13 in this container is /usr/local/python/3.13.8/bin/python3 with PYTHONPATH=kickmix:.venv313/lib/python3.13/site-packages:verifiers; the wrapper work/py.sh encodes exactly that (./work/py.sh …). Note: .venv313/bin/python in this checkout lost its execute bits/stdlib home; chmod +x recovers the binary but not stdlib resolution, so prefer the 3.13.8 interpreter above.

Benchmark I — invsqrt

Direct piecewise-cubic phasing circuit (no Newton iteration), inspired by the semi-log table and product-decode lookup ideas in the public portfolio notebooks (vqcc-own-circuit-portfolio-notebook.ipynb, quantum-circuits-four-loaders-eight-toffolis.ipynb): 128 semi-log intervals over M ∈ [1, 4) (64 per octave), per-interval cubic c0 − c1·d + c2·d² − c3·d³ fitted by reweighted NNLS, coefficients loaded by a Clifford Möbius/XOR network over the 7 case-bit monomials with a shared AND ladder, and mixed-precision Horner (widths 29/23/18/12; early-step errors are damped by powers of d ≤ 1/16). The 10 thermometer normalization flags are compressed to a 4-bit binary s before the Horner peak, and de-normalization is a binary barrel shift (stages 1, 2, 4, 8). Builder: work/build_invsqrt2.py; exhaustive model: work/model_direct.py (max relative error 2^-24.60 over all 3,133,452 inputs vs the 2^-24 tolerance). Final: 171 qubits, 3,757 Toffolis (reference: 260, 7,735); score 684,969.09 (−67.0% vs baseline, −62.8% vs the previous design below). The previous Newton–Raphson design (8-interval minimax cubic seed + one Newton step, stage-ordered allocation freeing c3/selectors before the Newton stages; builder work/build_invsqrt.py, fit/model work/fit8.py, work/model8.py; 240 qubits, 7,422 Toffolis, score 1,840,306.98) is retained. Its measured rejections stand for that structure: reference precision is near-tight and Horner per-term flooring caps the seed accuracy at ~2^-13, which is why the new design evaluates the cubic directly at final precision instead of polishing a seed with Newton.

Benchmark II — DQI

Reference Proos–Zalka EEA decoder with the GF(2¹²) inverse replaced by a composite-field inverse (work/gf_inv12.py): GF(2¹²) as GF(2⁶)[y]/(y²+y+ν), ν=32; inverse of b+ay is d+cy with Δ = νa²+ab+b² (the norm), d=(a+b)/Δ, c=a/Δ, and 0↦0 as kickmix expects. Basis changes are Clifford linear networks; the scaffolding del path uses only del_* primitives, so it is Toffoli-free. Forward cost 126 Toffolis (vs 315 for the built-in Itoh–Tsujii inverse). The primitive was verified exhaustively (all 4,096 field elements: value, cleanliness, phase) before integration into the Chien/Forney loop (4,095×) and the EEA quotient division (888×, reimplemented as inverse+multiply), and the decoder was verified at (m=63,n=10), (m=255,n=20), then full size. Builder: work/build_dqi.py (copied reference with the two inverse sites patched). Final: 1,905 qubits, 4,783,674 Toffolis (reference: 1,893, 5,725,461). Raw verifier JSON: work/dqi_verify_full.json, work/inv_verify_full.json, work/dfthc_verify.json.

DQI v2+ (2026-11-04, current submission): score 17,673,729,048,937.30 (−14.5% vs v1, −27.7% vs baseline). 1,884 qubits, 4,179,942 Toffolis. Refinements adapting the measurement-cleanup / shared-product ideas of the public portfolio notebooks (vqcc-own-circuit-portfolio-notebook.ipynb, quantum-circuits-four-loaders-eight-toffolis.ipynb) to this decoder: (1) the GF(2⁶) sub-inverse inside the composite inverse is itself a GF((2³)²) tower (work/gf_inv12.py init_inv6); generated bilinear multiplies throughout (work/gf_mul12.py: GF(2¹²) in 54 products, GF(2⁶) in 18, GF(2³) in 6 via a Chudnovsky evaluation; controlled form 66 vs 75, product registers erased with the Toffoli-free del_gf2_mul), bringing the GF(2¹²) inverse to 79 Toffolis vs 126; (2) tower basis changes in the inverses are applied in place (Clifford CNOT/SWAP networks, _apply_linear_inplace), and the quotient product accumulates directly into q_i, cutting peak qubits from 1,914 to 1,884 (with degree counters also narrowed to m.bit_length() bits); (3) the Chien σ==0 test is an 11-Toffoli AND tree erased measurement-style (_init_zero_flag/_del_zero_flag in work/build_dqi.py) instead of two 12-Toffoli flip_if_equal passes. Primitives verified exhaustively (values, ancilla cleanliness, phase; work/test_gf_primitives.py), decoder verified at (m=63,n=10), (m=255,n=20), then full size: 102,400 shots, salt dqi-v6-full, PASS with 0/0 errors (work/dqi_verify_v6.json); second salt at 10,240 shots and adversarial max-weight shots at full size also PASS, 0 errors (work/dqi_verify_v6_salt2.json). Explored and rejected this round: EEA support-based swap/shift pruning (packed simulation showed nearly every slot can be nonzero), reducing n_iters (near-tight: n=20 fails at 69 of 72 iterations, n=10 at 36–37 of 42), shrinking the EEA window below m=n+3 (fails outright), and a reversible Berlekamp–Massey core (its per-iteration discrepancy/ratio temps do not clear without either a ~840-qubit journal — whose Q² penalty eats the gate savings — or singular-prone recovery formulas; left as the documented next lever in the section below).

Benchmark III — chemistry (history below)

The DF-THC-SOS work under experiments/ and research/ stands as recorded in the rest of this document; the frozen 31.18% submission was copied unchanged into submission/dfthcsos/ and re-verified with a fresh salt.

Objective

Latest30% user objective achieved:31.182788% below the original baseline. Frozen ready submission: experiments/coordinator_verified/merged_C28_b10_30/submission/. Default score2,309,230,097.5865555; independent-salt score 2,309,230,043.6354394. Both pass the unchanged full verifier, exceeding the30% target2,348,919,711.82. Read first: experiments/coordinator_verified/merged_C28_b10_30/BEST.md.

Final parameters R65/B60/C28, brot10/J11/K7; T65095.56294, Q741. Seven128-qubit circuits use12 selection+110 word+6 workspace qubits. Lambda294.02612134123535, lambda_eff47.873768372534656, gap3.923618193151924; error_1=12.935649mHa, error_2=19.311558mHa. Canonical digest: 7b79a311bfb66d9e9573cb8b12023ad1b852ed07ae7dae75f79037fde98171ea.

The numerical method alternates whitened coefficient fits with coupled integer angle moves, compresses copies toC28, restores the admissible gap, and applies exact coefficient-copy rotations. One-body words are regenerated. The circuit method uses a high-bit Möbius/ANF lookup and merges deferred measurement phases. All branches are corrected; arbitrary word inputs are supported. Exact proofs, independent review, primitive exhaustive tests and 222,656 arbitrary-word/inverse cases on the final files supplement both full verifier runs. Original scoring sources and reference data are unchanged; the certificate records their hashes and all frozen submission hashes.

Actual tool-enabled Claude supplied the ANF prototype; native review corrected its branch-test coverage overclaim. Actual tool-enabled Qwen's initial cost analysis was rejected, then a narrow correction was independently checked. Native Codex workers completed numerical fitting, synthesis, proofs and review. All workers have stopped; optional shape checkpoints are preserved. No Modal compute was used. Compact reports are under research/{precision_30,qrom_30,compression_30,shape_30}/RESULT.md.

Earlier milestones remain intact:21.7446% under experiments/coordinator_verified/anf_C28_21/;25.87% under experiments/qrom_30/merged_C60_b10_gap3922/.

Previous user objective achieved: approximately20% below the original score. Frozen best: 2,681,652,028.8712068, 20.084266% improvement. Independent-salt verification also passes at 20.084237% improvement. Historical evidence: experiments/coordinator_verified/gray_C28_refit20/BEST.md.

Co-optimize three coupled components so that the electronic structure of FeMoCo — the FeMo cofactor of nitrogenase, the biological catalyst behind nitrogen fixation — can be block-encoded and simulated on a quantum device. Here “Co” means cofactor, not cobalt. The computational task here is Hamiltonian block encoding; passing its verifier does not establish an energy-saving ammonia synthesis process.

The three co-optimized components:

1. Classical tensor factorizations (DF-THC-SOS) — represent the Hamiltonian using one-body contributions and R·C squared generators built from coefficients and rank-one projectors u_b^(ρ) u_b^(ρ)ᵀ, with u_b^(ρ) over N = 76 orbitals, B basis vectors per rank component, and C coefficient copies (X^(ρc) = sum_b w_b^(ρc) u_b^(ρ) u_b^(ρ)ᵀ): score = λ_eff × T_total × Q_total, where λ_eff depends on the reconstruction quality (Λ, E_SOS, E_gap).
2. Givens rotation trees — encode every vector u^(ρ) as a sequence of 75 angle words decoded by a spanning tree of Givens rotations over the 76 orbitals. tree_edges (75, 2) is part of the submission.
3. Quantum lookup tables (QROMs) — the ANGLE_k kickmix circuits (K = ceil(75/J) files angle_k_0 … angle_k_{K-1}) reversibly load, for every vector address a < 76 + R·B, the J-word batch of row a (brot bits per word).

The three interact: brot (word precision) feeds both the QROM width (J·brot persistent qubits) and the Givens rotation cost; the factorization shape (R, B, C) sets the state-preparation costs; a cheaper QROM with more qubits trades off against Toffoli count.

The problem (competition task "dfthcsos")

• System: FeMoCo76o113e — N = 76 spatial orbitals, η = 113 electrons.
• Fixed constants: E_upper = -22140.27335 Ha; fixed symmetry shifts S, β₁ of the provided baseline (submitted H1_I, Hsym_lower ignored).
• Score (lower is better):

  score      = lambda_eff * T_total * Q_total
  lambda_eff = sqrt(E_gap * (2*Lambda - E_gap))
  T_total    = ANGLE_k Toffolis (forward + reverse pass, simulated)
             + outer_state_prep() + inner_state_prep()   [SpinFreeData_V1]
             + 2*(N-1)*(2*brot - 1)                      [Givens rotations]
  Q_total    = outer_persistent + inner_persistent + J*brot
             + max(outer_temp, inner_temp, angles_temp)
  

• Acceptance gates (else FAIL):
  • Lambda > 49.0, E_gap > 3.914898 Ha
  • 0 < E_gap < 2*Lambda
  • error_1, error_2 ≤ 20 mHa (reconstruction residuals vs FCIDUMP; equality is accepted by the current code)
  • every ANGLE_k circuit ≤ 128 qubits (QUBIT_CAP)
  • circuit contract: selection = ceil(log2(76+R·B)) qubits; word_0 … word_{J-1} = brot qubits each; workspace returns to |0> after each circuit; selection is unchanged and no residual phase is allowed; no Z_POW (phases must be multiples of π); reverse traversal must clear the word registers. Zero padding beyond angle 74 is checked.
  • tree_edges must be an integer spanning tree rooted at orbital 0, with parent-before-child ordering; vertices must lie in [0,76).
• Submission format: one directory with angle_k_0.kmx … angle_k_{K-1} (or .kmb) + submission.npz containing U (R,B,76), W (R,B,C), H2_I (R,C), H1_I, Hsym_lower (2926,), tree_edges (75,2). NPZ only — pickles are rejected. U supplies the shape; the actual vectors are decoded from the circuits. Stored W is native lowercase w, and the native identity offset is H2_I - sum_b W. Submitted H1_I and Hsym_lower are ignored for metrics, but remain part of the submission hash.
• Grading is stochastic: the verifier simulates each address 64× with a seed derived from SHA-256(submission) + a private grader salt. The public CLI defaults to --salt ""; rerunning the same submission with the same salt repeats the samples. Passing sampled HMR branches is not a proof of correctness on every branch or arbitrary word inputs.

Each circuit is an XOR delta oracle, not a fresh overwrite of its batch: delta[0,a] = batch[0,a], delta[k,a] = batch[k,a] XOR batch[k-1,a] for k > 0. This lets successive calls stream batches through the same word registers.

Local files

Path	Content
verifiers/dfthcsos/SCORING_README.md	The full rules (contract, scoring, thresholds) — read this first
verifiers/dfthcsos/verify.py	Entry point: verify(circuits, factorization) -> {status, score, …}
verifiers/dfthcsos/scoreinfra/	FCIDUMP reader, DF-THC math, circuit costs, register contract
verifiers/dfthcsos/FeMoCo76o113e.fcidump	Reference integrals (128 MB, SHA-256 checked)
verifiers/dfthcsos/FeMoCo76o113e.fixed_shifts.npz	Baseline symmetry shifts (used by verifier)
verifiers/dfthcsos/example_submission/angle_k_reference_circuit.py	Builds plain-QROM ANGLE_k circuits from a submission.npz
sample_submission/dfthcsos/	The baseline example submission (10 .kmb circuits + submission.npz)
kickmix/kickmix.so	Compiled kickmix simulator (verified with Python 3.13; incompatible with the local 3.12 environment)
paper/verifiable_circuit_compilation.pdf	The underlying paper (arXiv-2610.23832v2 source also included)

verifiers/dqi/ (Benchmark II) and verifiers/invsqrt/ (Benchmark I) are the other two competition tasks; all three are now in scope and optimized — see the current-state section at the top.

Working setup

• The repo's .venv is Python 3.12 — kickmix.so won't load on it (undefined symbol PyObject_ClearManagedDict). Use Python 3.13:

  cd /home/gwen/Quantum
  /usr/bin/python3.13 -m venv .venv313
  .venv313/bin/pip install numpy scipy
  

• kickmix.so is an extension module inside kickmix/, so the path to add is the folder itself, not its parent:

  cd /home/gwen/Quantum/verifiers
  PYTHONPATH=/home/gwen/Quantum/kickmix ../.venv313/bin/python \
      dfthcsos/verify.py ../sample_submission/dfthcsos/
  

(Runtime depends on the host; the supplied README estimates ~25 s. Every address is simulated 64×, so larger R·B increases the workload.)

Verified baseline (starting point)

Sample submission = starter kit run on the provided baseline factorization (log-depth Givens tree, brot = 12, J = 8, R=65, B=C=60, 3976 addresses, 10 QROMs, 119 qubits):

status:  PASS
score:   3355599588.32   (49.301369 × 93237 × 730)
Lambda:  308.4559        E_gap: 3.965477 Ha
error_1: 3.19 mHa        error_2: 18.64 mHa
ANGLES:  3974 Toffolis ×2 passes per circuit; 13 757 heuristic Toffolis

Reverified on 2026-10-23 with the unmodified CLI, empty salt and 64 shots per address (254464 address-shots). Verifier submission digest: 276d464ff14da828b6b33e8487d99975494cbdf0c3cb2bd37dd3c3ddc4002695.

angle_k_reference_circuit.py regenerates these exact circuits from the angle_words stored in submission.npz — a good template for building new ones.

Current improvement and round-zero findings (2026-10-23)

Completed continuation: the user raised the objective to approximately20% below the original baseline (target2,684,479,670.65). Native workers combined C28 coefficient compression and refitting with exact Gray-code QROM traversal. The frozen submission passes the unchanged full verifier with two salts and exceeds20% in both runs. All workers finished; no Modal work was launched.

• research/ROUND_ZERO.md records a completed local sensitivity analysis; research/round_zero.py reproduces it and writes the numerical evidence and a 106156-row heuristic resource grid.
• The full-precision residual of the shipped U is 16.1332 mHa. The tensor perturbation from baseline quantization has norm 9.4019 mHa; the combined residual is 18.6396 mHa, including a cross term. These norms do not add.
• Direct 11-bit re-encoding fails at error_2=25.0354 mHa. At 13 bits it reaches 16.7973 mHa, buying pruning margin at a 1.3074% score cost before any pruning. All baseline delta tables have 3976 distinct rows, so identical-row subtree compression offers no Toffoli saving here.
• A bounded coefficient-copy rotation pilot lowers Lambda while preserving the reconstructed Hamiltonian, h1_prime and gap. It right-rotates both stored W and H2_I by the same per-rho orthogonal transformation, keeping the angle circuits unchanged. See research/copy_rotation_pilot.py and experiments/copy_rotation_pilot/.
• Historical verified20% milestone: experiments/coordinator_verified/gray_C28_refit20/submission/. Score 2681652028.8712068, 20.084266% below the original baseline. A fresh independent-salt run also passes, score 2681653027.584237 (20.084237% improvement). Tiny score variation reflects sampled counts of correct conditional phase-cleanup operations.
• Parameters R65/B60/C28; brot=11, J=10, K=8. T≈75468.00, Q=741. Every circuit uses exactly128 qubits. Lambda=294.519395261239, lambda_eff=47.95361499286076, gap=3.93012327384713 Ha; error_1=6.392636 mHa, error_2=19.569349 mHa. All acceptance gates pass.
• The method compresses coefficient copies toC28, refits against the exact shifted ERIs, polishes integer11-bit words, refits coefficients again, and applies exact copy rotations. One-body angles are regenerated from the decoded two-body representation. A Gray-code XOR lookup retains one page indicator and costs66,688 expected ANGLE Toffolis. It preserves arbitrary word inputs and cancels phase in every measurement branch. Proof, independent review,534,528 exhaustive cases and655,360 independent kickmix shots supplement full verification. Original scoring sources, reference integrals, fixed shifts and baseline remain unchanged.
• Historical20% reproduction and evidence: experiments/coordinator_verified/gray_C28_refit20/BEST.md. It links the ready submission, both full results, certificate/file hashes, reproduction command, and compact research reports. Canonical digest: 0ed3e62e8c607e5f031c8e84510bfd91a39de859c1b871bc34f4acb0b9d57ffc.
• The previous11.708% frozen best remains at experiments/coordinator_verified/gapconstrained11_j10/. C30+Gray also passes at2,691,626,527.68803 (~19.79% improvement), under experiments/qrom_next/gray_C30_offset/.
• A simpler ordinary-reference QROM fallback also passes at experiments/precision_refit/gapconstrained11/submission/, score 2984514478.377341 (11.05868% improvement), J9/K9, Q733. The prior exact-copy-rotation candidate remains preserved in experiments/coordinator_verified/exact_rotation_cyclic150/.
• The pi harness is installed at /home/gwen/.local/share/pi-aristote/node/bin/pi (add that directory to PATH). Two fresh ilaas/qwen-3.8-27b sessions supplied a derivation and review; the coordinator corrected factual/API errors before implementing the pilot. Evidence and model assessment: research/pi_trial/.
• Modal profile ancktsang is configured and active, with a user-stated $30 this-month allowance. No Modal compute has been launched. The earlier $120 is a prospective aggregate ceiling, not an independently verified balance or authorization to charge this account beyond $30.

Remaining improvement levers

• λ_eff: better DF-THC-SOS factorization. At fixed gap, smaller Λ lowers λ_eff. At fixed Λ, dλ_eff/dE_gap = (Λ-E_gap)/λ_eff, so near the baseline smaller E_gap lowers the score, subject to the strict gap floor. This direction reverses for E_gap > Λ; score the full formula. The original dfthc training library is not shipped, but is not required by the verifier; a compatible optimizer can be implemented locally.
• ANGLES cost: the verified best uses an exact Gray-code page traversal and an output pivot to reduce workspace and fit J10 at11 bits. Further lookup changes require renewed phase and measurement-branch review. The baseline has 9 spare qubits under the 128-qubit cap. Reducing its 11 workspace qubits alone does not lower Q_total: state preparation dominates the shared workspace at 544 qubits. Extra workspace can be useful if it lowers Toffolis while respecting the circuit cap.
• brot / J trade-off: brot > 12 raises the Givens term 2(N-1)(2brot-1) and inner-prep temp qubits (544 → 1087 at brot=16) but may allow a better factorization; J changes K (number of circuits) and the J·brot persistent qubits.
• State-prep costs (heuristic, from R,B,C,N,brot and fixed chi=10): reducing ranks can lower prep costs (846 + 9461 Toffolis at the baseline). Recompute the exact model: optimized block sizes introduce discrete changes in costs — co-design with the factorization.
• tree_edges: choose a valid root-0 spanning tree (path, log-depth, or custom). Shape, integer vertices, connectivity and order are checked by evaluate_dfthc_angle_words. The tree affects decoding error and word-table structure at fixed brot; the heuristic Givens Toffoli count depends on N and brot, not tree depth.

Remaining levers (Benchmarks I / II)

• invsqrt: score is Q × C at 171 qubits / ≈4,006 Toffoli-equivalents. Horner multipliers now dominate (~2,960 of 3,757 static Toffolis, both directions, at the ~1-bit-per-cell AND+adder floor; a controlled iadd costs the same as the AND+adder rows it would replace). Rotation synthesis consumes ~249 of the 4,006 (36 rotations at eps ≈ 1.03e-12). Further wins would need a cheaper multiplier structure (rounding or guard-bit Horner steps to shave register widths, a binary-barrel normalize to replace the unary one), a measured-cleanup (HMR) scheme that erases Horner intermediates without a second arithmetic pass, or the notebooks' two-pass phase structure: the public portfolios report 62 qubits / ≈1,585 Toffoli-equivalents, so a different evaluation structure may still have substantial headroom.
• DQI: v2 (above) took the GF((2³)²) sub-inverse, the Karatsuba sweep multiply, and the measurement-erased Chien zero-test; the EEA core remains the floor (≈3.6M of 4.26M Toffolis). Next levers, in expected order: a Berlekamp–Massey decoder replacing the synchronized EEA (needs a reversible in-place update design; score scales as Q², and a syndrome-register-resident design could also cut the 1,914 qubits — the public portfolios' ~891-qubit / ~397k-Toffoli decoders show the headroom, but their circuits are not in this checkout); window-pruned EEA gate emission (bounds must be re-proved per emitted gate — rejected as unsafe without that proof; packed-simulation support analysis found nearly all EEA slots live, and n_iters is near-tight, so neither pruning direction showed safe headroom).

Open questions

• The local contract accepts new factorizations subject to the same gates, and the included mathematical specification explicitly describes their optimization. External competition participation or compute restrictions are not established by this checkout.
• The public scoring harness is verify.py (CLI and Python API), with reproducible empty-salt sampling. No additional harness was found in the listed task files.
• Native Codex agents successfully performed numerical fitting, circuit construction, and independent checks with tools. External Claude through its CLI supplied the exact pair-rotation derivation and implementation. Qwen tool access was repaired and smoke-tested, but its subsequent compression task did not deliver a valid numerical result; native agents completed the screens. Read compact RESULT.md reports rather than logs. ORCHESTRATION_PLAN.md is historical planning; BEST.md and this handoff describe the current verified state. No Modal compute has been launched.

Compact research reports

• research/precision_30/RESULT.md: successful ten-bit fitting, coupled integer-word solver, C28 compression, gap restoration and exact rotations.

• research/qrom_30/RESULT.md and PROOF.md: actual Claude ANF contribution, native generalization, merged phase cleanup and independent branch tests.

• research/shape_30/ANF_REVIEW.md: independent merged-phase review. RESULT.md records basis reduction, relaxed spectral bounds and checks.

• research/compression_30/RESULT.md: nonwinning low-C experiments and actual Qwen assessment; corrected cost row independently reproduced.

• research/compression_next/RESULT.md: successfulC28 compression, fitting, polishing and copy rotations; both full verifier results. Stronger numerical-onlyC28/C30 checkpoints are retained for future work.

• research/qrom_next/RESULT.md and PROOF.md: successfulGray lookup, exact branch proof, exhaustive tests, verifiedC60/C30/C28 milestones.

• research/precision_next/REVIEW.md: independent Gray-QROM audit and additional266,240 all-branch checks. RESULT.md records a rejected10-bit experiment (28.5647 mHa); it establishes no impossibility bound.

• research/joint_fit/RESULT.md: continuous jointU/W fit lowered the full-precision residual to14.4147 mHa but did not give a better quantized submission; native-offset optimization produced numerical-only variants.

• research/precision_refit/RESULT.md: successful coupled weight/word fit and gap adjustment; ordinary reference-circuit fallback already exceeds10%.

• research/qrom_construction/RESULT.md: successful lookup construction, exact phase proof and exhaustive primitive validation.

• research/qwen_compression/RESULT.md: repaired Qwen tools, model delivery failures, and native rank-compression screens with preserved checkpoints.

• research/tree_precision/RESULT.md: external Claude's fixed-weight quantization experiments did not reach the accuracy gate. Its tree surrogate is a heuristic, not a proof that other trees cannot succeed.

All workers finished. Final optimization used local compute; Modal was not used.