# Research: primitive odd modular periods Checked definitions: PARI mfperiodpol(mf,f,-1) gives the odd part of the upward integral from zero to imaginary infinity of f(tau)(X-tau)^(k-2). Even k makes this identical to the (tau-X) convention. Paşol–Popa Section 1, Proposition 5.11(a,c), and equation (5.7) fix rationality, real phase and the arithmetic Mellin factors. The primitive integral polynomial has positive leading coefficient and coefficient gcd one; r_f^-=Omega_f^-*P_f^- defines the signed scale. For odd m, [X^(w-m)]r_f^-=binom(w,m)*(-1)^((m-1)/2)* m!/(2*pi)^(m+1)*L(f,m+1), w=k-2. The parity average includes 1/2. Checked Delta control: P=4X^9-25X^7+42X^5-25X^3+4X; hence Omega=10*L(Delta,2)/ (4*(2*pi)^2). The first direct sample is 0.009269276162370163236258034682... No absolute value, conductor factor or rational reconstruction is used. Live discovery: authenticated catalogue cached once in catalogue.json and catalogue-pages.json; 762 tables at this planning pass. already_here and already_asked saved in cheap-screen.json. Related period, Manin/Shimura and primitive-period candidates inspected. T696 contains the exact shape, T765 the separate even scale (offered_for_review); T323/T683 arithmetic critical values, T326/T684 Petersson products. No odd scalar table was found. The source Delta L-value lookup gives a positive T323 search control; source Manin-period probes give no match and are not a uniqueness argument. Origin #262 was read live and supports explicitly normalized periods, broad finite weight/level slices and independent Mellin checks. #303 and #309 are context for existing companions, not direct origins of this table. The exact initial message supplied for proposal #349 cites only #349 and #262. All four suggested live tags apply: modular form, number theory, period, special values. No tag departure is needed. No issue is to be closed. Fresh T696 metadata and document are cached. Its generator source, hash de34dd03ca07e1eebe7af6ec731422362a8291b84204844f8a0ed76bd75e435a, was downloaded from its live attachment and isolated as forms.py. It is a reusable exact algorithm to recheck, not an oracle. Roster comparisons use its exact T694-roster attachment. All selected shapes and indices must agree. A conflict is a mathematical dependency blocker; published tables stay intact. Shared even-builder computation components were reused for independent-route code (explicit Fourier sums and tails), not its scalar results or verdict. The split point is 1/sqrt(N), with epsilon=(-1)^(k/2)*Fricke. Sage's raw Atkin–Lehner operator is divided exactly by N^(k/2-1), as documented. A Fricke/Fourier comparison with PARI will validate its sign. The Fourier bound |a_n|<=d(n)*n^((k-1)/2), Bringmann–Han–Heim–Kane Theorem 2.6, p.4, implies |a_n|<=2*n^(k/2) and an explicit positive geometric tail after expanding the integer incomplete gamma. Checked source locations and full cached-file hashes are in source-hashes.json. Full coset Haberland pairing is not asserted. Coordinated instructions: manifest.json, unchanged from agents/runs/20261119T222759Z-build-skills/manifest.json, SHA256 7629363ce07d2e1340528294fe639e6fc7a5b18da808ec52558687fc1db91534. All archived resource hashes verified; valid local/live bundles match b5c3c7985b7c9e178b4899d083a9e3bc4284989dff76f05e5b0e52835d47eed2. Core and all six task-relevant references read. Independent critique and argument-selection verdict remain controller stages, not this build's output. Checked build results: every one of the 18 odd scales passes all 111 odd coefficient tests at both independent Mellin settings (150 digits/450 terms, 190/600). All exact shapes match live T696; all 33 embedded-form indices and 21 space dimensions agree. Each rational form's 600 exact Fourier coefficients match between Sage and PARI. Stored intervals contain every nonzero coefficient-derived scale; worst midpoint residual <0.48 stored ULP. Narrowest independent scale accuracy is 520 bits. Both wrong signs and factor-two errors are rejected for every form. The table is conservatively heuristic (agreement-checked); no critic verdict has been written by this run. The sole full computation is preserved in checks.log, with per-form CHECKPOINT lines, complete JSON artifacts and dependency versions. No prior sibling numerical result or failed-run checkpoint was reused.