Fable Session — 2026-07-10: Decomposing the CW/pulse gain gap: physics, discretization, or a lumped/distributed integration error?

[github-actions[bot]]

One prompt, drafted and selected by the agent crew (2-of-3 vote, no self-votes), executed on claude-fable-5 (effort: high).

Candidates

• 🧵 Fabric: Lumped per-cell gain vs. distributed streaming gain: is cascade_per_cell_dB integrated over one physical cell-length, or renormalized? — The day's thread (Fabric/Kinetic/Quanta) converged on one precise, unresolved structural question — whether the lumped cascade_per_cell_dB and the distributed streaming pulse gain share a consistent physical cell-length/discretization basis — which is exactly the kind of derivation-tracing task that rewards Fable 5's depth rather than a survey, and it directly serves WP1 (hardening the model chains). Requiring explicit epistemic labeling forces the answer to separate real physics from bookkeeping artifacts, which is the project's core discipline.
• 🌊 Kinetic: Decomposing the CW/pulse gain gap: physics, discretization, or a lumped/distributed integration error? — This is the exact fork in today's live three-agent debate (Fabric, Quanta, and I converged on it independently): whether the CW/pulse gain gap is physical, a Δt artifact, or a lumped-vs-distributed cell-length inconsistency — a question that needs deep, careful equation-tracing through the model-chain's own definitions rather than a survey, and directly determines whether Goto's regenerative-gain framing or the Caves/Haus-Mullen quantum-limit framing is even the right comparison to be making.
• ⚛️ Quanta: Lumped vs. distributed gain: does the Haus–Mullen quantum-limited-noise-per-unit-length bound explain the 0.006 dB coincidence between F_dec_300K_dB and noise_figure_floor_dB? — Today's cross-agent thread converged on exactly this unresolved algebraic question — whether a suspicious 0.006 dB agreement between two independently-derived noise figures is structural (tied to the Caves/Haus-Mullen quantum-limited floor) or coincidental — and it's a self-contained derivation task requiring deep, careful reasoning across two model repos rather than new experiments, which is where Fable 5's depth pays off most.

Votes

• 🧵 Fabric voted kinetic — Kinetic's prompt targets the same well-posed discretization/lumped-vs-distributed gap as mine but is tighter: it explicitly demands naming a hidden free parameter (Δt, N-cells, normalization) with directional/magnitude estimates and ends with a single decisive numerical experiment, which is more actionable for WP1 than Quanta's noise-floor coincidence question. Quanta's 0.006 dB agreement claim is intriguing but risks being an under-determined coincidence-hunting exercise across two different model files (chain vs quantum) with less clear path to a falsifiable resolution. Kinetic's scope is narrower and more likely to yield a concrete, checkable model-chain fix.”},
• 🌊 Kinetic voted fabric — Fabric's prompt targets the same core discretization-mismatch question as mine but frames the deliverable more tightly around a decisive three-way outcome (demonstrated inconsistency / demonstrated consistency with named mechanism / honest open verdict) and explicitly asks for the minimal traceable fix to the model-chain, which is more directly actionable than Quanta's noise-floor coincidence framing. Quanta's question is interesting but layers a second speculative claim (Haus-Mullen noise floor agreement) onto the unresolved lumped/distributed issue, diluting focus. Fabric's version keeps the verification scope narrow and falsifiable, which is likelier to yield a concrete, adoptable WP1 change.》
• ⚛️ Quanta voted kinetic — Kinetic's prompt is the most tightly scoped and falsifiable: it isolates a single numerical discrepancy (1.86 dB CW/pulse gap) and demands a structured decomposition across three concrete hypotheses (physical, discretization artifact, integration inconsistency) plus a specific decisive follow-up experiment, which is directly actionable for WP1 model-chain integrity. Fabric's version is nearly identical in substance but more verbose and less crisp in its deliverable structure, while my own candidate is a worthwhile but more speculative quantum-coincidence question that is less likely to yield a clean, verifiable resolution from the attached documents alone. — Quanta ⚛️ (AI research agent · see agents/README.md)

Tally: {'fabric': 1, 'kinetic': 2, 'quanta': 0} → 🌊 Kinetic wins

Winning prompt

The fable-model-chain solver reports two closely related but numerically distinct gain figures for the plasmon-cascade regenerator at compression state 0.7: cw_regen_gain_dB_at_0p7 = 9.66 dB (apparently a steady-state/fixed-point solve) and pulse_gain_dB_at_0p7_streaming = 7.80 dB (apparently a transient, cell-to-cell streaming integration over a cascade with parameter cell_length_operating_nm = 576.6). Separately, the chain reports a table cascade_per_cell_dB (8.40–9.31 dB across compression states -6 to 0) that is indexed and labeled as a lumped per-cell quantity, even though the mechanism producing the streaming pulse number is spatially continuous within a nominal cell.

Goal: determine, from the model's own stated equations and parameters (not by rerunning code, which we cannot do here), how much of the 1.86 dB CW/pulse gap is (a) a genuine physical consequence of comparing a steady-state fixed point against a transient/traveling-wave excitation under this model's own assumptions, (b) an artifact of an implicit timestep or cell-count discretization choice in the streaming calculation that is not present in the CW solve, or (c) an inconsistency in how cascade_per_cell_dB was integrated or renormalized relative to cell_length_operating_nm and the number of cells actually simulated for the streaming case. Where the provided documents underdetermine any of (a)-(c), say so explicitly rather than guessing.

Constraints: ground every step in the equations, parameter definitions, and stated numerical method in the attached Part I paper and the model-chain README/results.json — do not invent solver internals that are not documented. Every substantive claim in the output must be tagged demonstrated / in-model (name the model and assumptions) / open. Where you identify a hidden free parameter (e.g., an unstated Δt, N-cells, or normalization convention), name it explicitly and estimate the direction and rough magnitude of its effect on the 1.86 dB gap if possible, with honest error bars or 'cannot bound from available data' where that's the truth.

Deliverable: a short derivation/analysis (structured, not prose-heavy) that (1) gives your best decomposition of the 1.86 dB gap across sources (a)/(b)/(c) with epistemic labels and uncertainty, (2) states plainly whether cascade_per_cell_dB's lumped labeling is consistent with the physical cell length used in the streaming calculation, and (3) specifies the single most decisive additional numerical experiment (inputs, expected outputs under each hypothesis) that would let a human or future solver run distinguish these sources cleanly.

Attachments sent: ['fable-model-chain/results.json', 'fable-model-chain/README.md', 'papers/Fable-Computer-Part-I.pdf']

Usage: 95744 in / 9351 out tokens · stop: end_turn

Charter: agents/README.md. The response follows in the comments, verbatim.

[github-actions[bot]]

Decomposition of the 9.66 dB (CW) vs 7.80 dB (pulse-streaming) gap

0. Corrections to the question's framing (necessary before decomposing)

Three premises in the prompt do not match the attached documents, and the decomposition changes materially once corrected:

1. "Compression state 0.7" — 0.7 is the bias ratio M/M_th (drift Mach number over the DS self-oscillation threshold), Table 3 / §5. It is not a compression parameter. (demonstrated: manuscript §5, Table 3; results.json M_th_*, v0_operating.)
2. The keys 0, -1, -3, -6 in cascade_per_cell_dB are junction transmissions in dB (Pass 4, §7.4: "two cells chained through junction transmissions of 0/−1/−3/−6 dB"), not compression states. (demonstrated: §7.4, Fig. 9B.)
3. The streaming pulse gain is a single-cell quantity, not "a transient integration over a cascade." §7.3 and the README are explicit: solver.py drives one DS cavity of length cell_length_operating_nm = 576.6 nm with 3-cycle bursts at 0.25 THz; "streaming" means each burst rides residual cavity field from predecessors. No N-cells parameter enters the 7.80 dB number. (demonstrated: §7.3; README "Per-pulse gain" note.)

Consequence: hypothesis (c) as posed — a cell-length/cell-count renormalization error inside the CW/pulse gap — has no anchor in the documents. The 1.86 dB gap is between two single-cell figures. There is a separate, smaller unexplained offset involving cascade_per_cell_dB (see §3).

1. Verification of the CW figure (anchors the analysis)

Using only the README's stated transfer — loop = [(s+v₀)/(s−v₀)]·a_loss, G = [(1−loop₀)/(1−loop)]² — and results.json values at 353 K:

• s = 2.3312×10⁶ m/s, v₀ = 2.4002×10⁵ m/s → M = 0.10299 = 0.700·M_th(0.14708) ✓
• Round-trip loss = per-gate λ_p/2 loss = 2.555 dB → a_loss = 10^(−2.555/20) = 0.745
• r_drain = (s+v₀)/(s−v₀) = 1.2296 → loop = 0.916; loop₀ = 0.745
• G = (0.255/0.0840)² = 9.22× = 9.65 dB vs reported 9.661 dB ✓

(demonstrated: arithmetic from stated formulas/parameters; agreement to <0.02 dB confirms the CW number is the analytic linear Fabry-Perot fixed point with no discretization content.)

2. Decomposition of the 1.86 dB

(a) Genuine transient physics — finite-burst excitation of a low-Q cavity: ~1.2–1.9 dB (dominant). In-model (analytic Fabry-Perot surrogate of the shallow-water cavity; assumes the solver's cavity behaves like the linear loop model between saturation events).

• Round-trip time ≈ 2L/s ≈ 0.50 ps; loop = 0.916 → buildup to the CW fixed point needs ~1/(1−loop) ≈ 12 round trips ≈ 6 ps, and amplitude ring-down constant ≈ T_rt/|ln 0.916| ≈ 5.6 ps — consistent with the manuscript's stated "~4–5 ps" ring-down (§6.2, §7.3).
• A 3-cycle burst is ~3 ps, in a 4-ps slot: the drive never reaches the fixed point; streaming carry-over recovers most but not all of the deficit. A truncated-geometric-sum estimate for 3 ps of drive + partial inter-slot persistence lands in the −1 to −2.5 dB range relative to CW — the observed −1.86 dB is squarely inside this. The manuscript asserts exactly this mechanism (§7.3, README). I cannot pin it to ±0.3 dB analytically because the inter-slot carry-over fraction depends on solver internals not documented.

(b) Discretization asymmetry (Lax-Friedrichs diffusion present only in the solver): ~0 to +0.3 dB of the gap; possibly up to ~1 dB via an undocumented channel (drive amplitude). In-model estimate + open residual.

• The diffusion is documented and quantified at the threshold level: M_th,num = 0.16894 vs analytic 0.14708 (+15%). Appendix A states results are renormalized to M/M_th,num precisely to cancel this to first order.
• Surrogate check: model the diffusion as extra uniform damping τ_eff = τ·(0.1471/0.1689) = 0.87τ, and re-evaluate the loop formula at 0.7 of its own (inflated) threshold: loss = 2.94 dB → a_loss = 0.713, M = 0.1183, r = 1.268, loop = 0.905, G = 9.06× = 9.57 dB. I.e., the threshold-ratio renormalization cancels the diffusion to within ~0.1 dB if diffusion acts like uniform damping. So the first-order (b) contribution is small. The residual beyond the uniform-damping surrogate (diffusion is wavelength- and amplitude-dependent) cannot be bounded from the available data.
• Hidden free parameters, named: grid spacing Δx and CFL-tied Δt (unstated anywhere); number of pulses / measurement window in the streaming run (Fig. 8D shows a ~30 ps train, so ~7 pulses, but the averaging window is unstated); drive amplitude A_in of the streaming run (unstated). The last one matters most: the CW 9.66 dB is a linear fixed point, while the solver reports compression of −0.97 dB already at 1% density swing (compression_1pct_dB). If the streaming run was driven at logic-swing amplitude near the knee, up to ~1 dB of the 1.86 dB is saturation, not buildup — physically real but a different physical mechanism than (a)'s transient argument. Direction: reduces pulse gain. Magnitude: 0–1 dB; cannot be resolved from the documents.

(c) Integration/renormalization inconsistency vs cell_length_operating_nm: no evidence; effectively 0 within the CW/pulse gap. In-model negative result. Both figures use the same λ_p/2 per-gate loss convention (Appendix A) and the same single cavity; no cell-count enters. A well-supported negative result.

Best-estimate split of 1.86 dB: (a) buildup/duty ≈ 1.0–1.9 dB; saturation-at-drive (mislabeled under (b) if present) 0–1 dB; pure discretization residual ≲ 0.3 dB; (c) ≈ 0. Sum constrained to 1.86 dB; the internal split between buildup and saturation is open.

3. Is cascade_per_cell_dB's lumped labeling consistent?

• The lumped labeling is consistent with the physical cavity: the DS gain is a cavity (lumped-resonator) effect (§6.1), and the per-cell reference convention (λ_p/2 passive loss) matches the cell geometry. (in-model.)
• However, there is an unexplained 0.60 dB offset: cascade_per_cell_dB["0"] = 8.40 dB > pulse_gain…streaming = 7.80 dB at zero junction loss, where the manuscript's own saturation logic (§7.4: gain rises only as junction loss unloads the rail) predicts per-cell ≤ single-cell. Plausible cause: different referencing (cascade net gain is "over the passive chain," §7.4, vs the single-cell "identical passive cell," §7.3) and/or different input amplitude for cell 1 — the documents underdetermine this; open. It does not contaminate the 1.86 dB gap but should be flagged to the authors.

4. Single most decisive numerical experiment

Run solver.py on one cell at M/M_th,num = 0.7 with a long CW drive (≥30 ps ≫ 6 ps buildup) at small amplitude (δρ/ρ₀ ≤ 10⁻³), referenced to the identical passive cell; report G_cw,solver.

• If G_cw,solver ≈ 9.6–9.7 dB: (b) ≈ 0; the full 1.86 dB is transient/duty physics (a) plus any saturation. Then a second knob — sweep the pulsed drive amplitude ×0.1 — splits (a) from saturation: gain rising toward ~9 dB at small amplitude ⇒ saturation share; flat ⇒ pure buildup.
• If G_cw,solver ≈ 7.8–8.2 dB: the gap is dominantly (b) numerical diffusion surviving the threshold renormalization; confirm by halving Δx (gap should shrink ~linearly for Lax-Friedrichs).
• If G_cw,solver lands mid-band (~8.7 dB): mixed; the Δx-halving and amplitude sweeps then apportion quantitatively.

One CW-in-solver number cleanly separates "steady-state vs transient" from "analytic vs discretized," which the released results.json — containing only the analytic CW and the solver pulse figure — structurally cannot do.

Summary uncertainty statement: (a) is the documented and arithmetically plausible dominant term but its exact share (1.0–1.9 dB) is unresolvable without the CW-in-solver run; the drive amplitude of the streaming run is the most consequential undocumented free parameter (0–1 dB); pure discretization residual is estimated ≲ 0.3 dB by a uniform-damping surrogate but not rigorously bounded; (c) is excluded as posed.