Edison literature briefs: objectives / outcomes per drop-test data source + cross-modality ANALYSIS synthesis

[Copilot]

Per #36 comment 4509305026, one LITERATURE_HIGH Edison query per measurement modality, tying raw signal → BO objectives / constraints / characterization settings for the PETG+TPU 85A H2D campaign (PRs #30 / #33). Per PR #60 comment 4511245383, a follow-up ANALYSIS task synthesizes those 5 briefs into a single cohesive recommendation and answers the normalization-by-mass/volume question raised against the PR #35 design space.

Changes

• scripts/edison/submit_objective_functions.py — driver that submits 5 queries non-blocking, then polls get_task(task_id=…) to terminal status. Idempotent via <slug>-SUBMITTED.json placeholders so re-runs resume polling instead of re-billing. Mirrors EDISON_API_KEY → EDISON_PLATFORM_API_KEY per repo convention.
• edison-trajectories/objective-functions/0{1..5}-*.{md,json} — verbatim formatted_answer + full model_dump_json for each of:
  1. Accelerometer on M23 (200 ms shock + ~10 s ringdown) — cfd30f3e
  2. High-speed / slow-mo phone video — 7d6b43bf
  3. Electrodynamic shaker transmissibility — 31126ee7
  4. Pneumatic slug-firing / gas gun — 9d74ab2e
  5. Polytec VibroFlex QTec LDV — f40e41a7
• Uniform (a–g) skeleton in every brief so they diff side-by-side when wiring bo/tensegrity_campaign.py: raw observables → BO objectives (g_max, SEA, eta, epsilon_d, transmissibility, zeta, N_reuse, …) → constraints → characterization settings (rate / mount / ASTM-ISO-JEDEC standard) → BO-campaign integration (Ax Metric shape, noise model, fidelity tier) → gotchas → numbered DOI refs.
• scripts/edison/submit_objective_functions_synthesis.py — follow-up driver that uploads every per-modality .md / .json artifact via client.upload_file() → data_entry:uuid and attaches the full set to a single JobNames.ANALYSIS create_task(files=[...]) call. Idempotent via a synthesis-SUBMITTED.json placeholder. The prompt asks for: (1) a master cross-modality objective/constraint matrix, (2) a concrete Ax MultiObjective / ObjectiveThreshold / OutcomeConstraint recipe with per-metric noise model, (3) multifidelity slotting into the MuJoCo → Newton → PolyFEM+IPC ladder from PR Add runnable tensegrity simulation demos (MuJoCo, PyBullet, PyChrono, Newton, DiffPD, PolyFEM+IPC) + 3D animated renders + Edison survey + regime-aware sweeps + PLA/TPU 85A printable-design model + BO integration bridge + PR #35 Sobol T3-prism sweep ac... #33, (4) a dedicated normalization section comparing SEA (J/g), VEA (MJ/m³), Pajunen 2019 W_min = C_min · ρ_rel, ASTM cushion-curve g_max(ρ_T), and Gibson-Ashby ρ_rel^n scaling under the PR Add T3-prism (3-strut tensegrity) parametric CAD with Bambu PETG .gcode.3mf slice + re-importable project .3mf (H2D-only, supports enabled, scale 1.5× / cable_d 4.5 mm) + PLA-cables and PLA-struts/TPU-cables MM variants (with modeled-in PLA scaffo... #35 design space that couples m_specimen and V_bb on every trial, (5) cross-modality sanity equalities, and (6) ranked open gaps for further queries. ANALYSIS task id 789de8ab-9c68-4782-a70c-0a5a4e10e268 submitted; fetch next session and write synthesis-789de8ab-*.{md,json}.
• edison-trajectories/objective-functions/README.md — index table, (a–g) recap, and a cross-modality picture assigning each sensor a BO role (destructive primary objectives / ground-truth scoring / cheap pre-screen / strain-rate extension / non-contact cross-check), with links to equipment/, bo/, and simulations/validation_experiments.md. Now also lists the synthesis row + run command for the new driver.

Driver shape

# scripts/edison/submit_objective_functions.py — idempotent resume
placeholder = OUT_DIR / f"{q['slug']}-SUBMITTED.json"
if placeholder.exists():
    task_id = json.loads(placeholder.read_text())["task_id"]
else:
    task_id = client.create_task({"name": JobNames.LITERATURE_HIGH,
                                  "query": build_prompt(q)})
    placeholder.write_text(json.dumps({"task_id": task_id, ...}))
# poll until status in {success, failed, cancelled, error, crashed}
while time.time() < deadline:
    res = client.get_task(task_id=task_id)
    if (res.status or "").lower() in TERMINAL: break
    time.sleep(30)

# scripts/edison/submit_objective_functions_synthesis.py — upload + ANALYSIS
file_uris = [client.upload_file(file_path=p, name=p.name, tags=[...])
             for p in sorted(per_modality_briefs)]   # "data_entry:{uuid}"
task_id = client.create_task({"name": JobNames.ANALYSIS,
                              "query": SYNTHESIS_PROMPT},
                             files=file_uris)
PLACEHOLDER.write_text(json.dumps({"task_id": task_id,
                                   "uploaded_files": [...], ...}))

Note: run_tasks_until_done() cannot resume by task_id — TaskRequest rejects it as extra_forbidden — so polling is the correct primitive here.

[sgbaird]

@copilot send another edison query (analysis mode) to synthesize the findings from the previous trajectories into a cohesive file, making sure to upload all of the previous artifacts. Fetch next session.

In tandem with that, I have a question: based on these results, would it make more sense during the optimization campaign to try to normalize by mass/volume, etc.? I.e., how do we normalize the results since changing the parameters in #35 changes both the mass and bounding volume. I think there might be another PR commentary somewhere that touches on this, but you might have to do a pretty in-depth, one-by-one search to find it.

[Copilot]

@sgbaird Unfortunately I hit an unexpected error while processing your comment. I've automatically reported this to GitHub.

You can ask me to try again later by mentioning me in a new comment.

If you want to contact GitHub about this error, please mention the following identifier so they can better serve you: 69483478-be78-490d-bf47-8f9f24088f90

Sorry for the inconvenience!

[sgbaird]

@copilot send another edison query (analysis mode) to synthesize the findings from the previous trajectories into a cohesive file, making sure to upload all of the previous artifacts. Fetch next session.

In tandem with that, I have a question: based on these results, would it make more sense during the optimization campaign to try to normalize by mass/volume, etc.? I.e., how do we normalize the results since changing the parameters in #35 changes both the mass and bounding volume. I think there might be another PR commentary somewhere that touches on this, but you might have to do a pretty in-depth, one-by-one search to find it.

[Copilot]

@copilot send another edison query (analysis mode) to synthesize the findings from the previous trajectories into a cohesive file, making sure to upload all of the previous artifacts. ...

Submitted Edison ANALYSIS task 789de8ab-9c68-4782-a70c-0a5a4e10e268 in c8e1ade — all 10 prior .md + .json artifacts uploaded via client.upload_file() → data_entry:… and attached to the task (see synthesis-SUBMITTED.json for the full upload manifest). Will fetch next session.

Normalization (short answer, pending the synthesis): I searched issues/PRs/branches for a prior commentary thread on this and didn't find a dedicated one; the existing per-modality briefs do discuss it implicitly, so here's the working position the synthesis prompt asks Edison to confirm against the literature.

The five PR #35 design variables (R_mm, H_mm, twist_deg, strut_d_mm, cable_d_mm, plus tiling/infill) couple both m_specimen and the bounding-cylinder V_bb on every trial, so raw (unnormalized) BO outcomes silently confound "this architecture is better" with "this cell happens to be bigger / heavier". Three commonly-cited normalizations from the briefs:

• Mass-normalize (SEA = E_abs / m, J/g) — what brief 01 already calls out as a primary objective. Defensible for the crutch-tip and lander-payload framings (mass is the budgeted resource). Bias: rewards low-mass cells even when per-unit-volume crush is worse.
• Volume-normalize (VEA = E_abs / V_bb, MJ/m³, or σ–ε integrated to ε_d) — brief 04. Right denominator when packaging volume is the budgeted resource (egg-drop bounding sphere ≤200 mm). Bias: rewards filling the box even if mass-inefficient.
• Relative-density composite — Pajunen 2019 W_min = C_min · ρ_rel (brief 04 directly cites the polymer-tensegrity target W_min < 0.21 at ρ_rel < 0.1). Gibson-Ashby E*/E_s ~ ρ_rel^n, σ_pl*/σ_ys ~ C·ρ_rel^m. Right denominator for cross-architecture comparison at fixed parent solid (PETG strut + TPU 85A tendon stays constant across PR Add T3-prism (3-strut tensegrity) parametric CAD with Bambu PETG .gcode.3mf slice + re-importable project .3mf (H2D-only, supports enabled, scale 1.5× / cable_d 4.5 mm) + PLA-cables and PLA-struts/TPU-cables MM variants (with modeled-in PLA scaffo... #35 trials, so this works).
• Cushion-curve (ASTM D1596 / D5276 / D4168) g_max vs static stress ρ_T = m_payload·g / A_footprint — collapses thickness + footprint into one curve.

Working recommendation for bo/tensegrity_campaign.py (subject to the synthesis reply):

1. Primary objectives (2–3 for qNEHVI): maximize SEA_J_per_g, minimize g_max (already intensive — m_top·a_top is independent of specimen mass), maximize eta = F_mean / F_peak (intensive by construction).
2. Promote m_specimen ≤ 500 g and V_bb ≤ V_bb_max to constraints rather than ratios, so the Pareto front isn't artificially flattered by trivially shrinking the cell.
3. As a secondary cross-architecture FoM, log Pajunen's W_min as an auxiliary Metric (Ax can carry it without it driving acquisition) so we can compare against the published W_min < 0.21 envelope.
4. For modality 4 (gas gun) explicitly do not mass-pool with modality 1 (M23) — different strain-rate regime; use a MultiTaskGP per the multifidelity section.

Re measurement cost: m_specimen is a single weigh-in (cheap, do it for every trial), and V_bb falls out of the BO parameterization directly without any post-print measurement. So normalization can be computed at every trial with one extra line. The synthesis task is asked to confirm/refine all of the above with primary citations.