10 Matching Annotations
  1. Last 7 days
    1. are merged into one plant operating cost per kg

      Claude (for David, who agrees with this split): Alex, a suggestion we'd like your view on. Instead of one merged plant operating cost, split it three ways by what drives each part:

      1. Labour per plant, depending on plant size and complexity rather than per kg.
      2. Maintenance and insurance as a share of capital (Humbird 2021: 4% and 5% a year), so a world can't pair cheap running costs with an expensive plant.
      3. Utilities and consumables per kg, varying by process, including cell-retention membranes for perfusion and continuous culture (about $4.53/kg wet in Humbird's perfusion design).

      This follows point 3 of your literature review, and would tie the input that tops 'What has to go right' (98% of minimal sets) to evidence. The main model would need the same change, since it also keeps maintenance and insurance separate from capital and has no membrane cost. Does this fit how you'd want the frontier plant modelled, and would you make the change in your version?

    2. The Learn page puts the seed train at 5–10% of production cost

      Claude: Thanks; noted for a Learn page fix, together with note 6 (its cost shares describe the cheap worlds) and the documentation fixes the literature review flags (Garrison link, maturity mean 0.5 in docs vs 0.40 in code, the toll's per-kg basis).

    3. Wet-basis costs made the answer depend on how wet the cells are.

      Claude: Agree, and it applies to the main model too, where some inputs are still converted at a fixed 20% solids. The other two fixes in this note are also worth adopting: flooring the medium-use multiplier at 1 lowers the share under $36 from 0.049% to 0.042%; centring the maturity nudge on its mean raises it to 0.055%.

    4. For usability, a very simple comparison between dry cell mass and commercially available chicken is assumed.

      Claude: In 1,000,000 draws neither engine puts any world below the $12/kg dry parity line (cheapest world: $19.4 in the main model, $15.7 here). So 'competitive' in practice means the $12–36 band, up to about three times wholesale chicken. Whether a cultured product at organic-chicken prices counts as competitive is a judgement worth putting to the workshop. For reference, the main model's $5/kg hybrid threshold equals $40/kg dry, close to this zone's upper edge.

    5. What has to go right

      Claude: A good complement to the unconditional probability rather than a replacement for it: the share of competitive worlds is the probability that every condition in at least one minimal set holds. Under $36/kg dry that is about 1 in 2,000 worlds in the main model and about 1 in 1,500 here (1,000,000 draws). Two caveats on the shares: technologies start in their unfavourable state and range inputs at their central values, so some inclusion is mechanical; and the result rests on how far the favourable tails reach and whether they co-occur, which nobody has elicited yet. That is where the workshop elicitation should point. Related: the threshold profile on the main model's sensitivity page.

    6. are merged into one plant operating cost per kg

      Claude: This is the change I would reconsider. The literature review recommends tying maintenance and insurance to capital (Humbird: 4% and 5% of capital a year), but the merged input is drawn independently of capital. It is also the top item under 'What has to go right' (98% of minimal sets), and its favourable end (about $9/kg dry, $1.80/kg wet) is, by the review, roughly labour alone. So the most decision-relevant input has the least-supported favourable tail. Worth splitting back into labour plus capital-linked costs before the workshop.

    7. There's no point imagining a frontier plant which uses a worse process, and anyway process doesn't seem to be the major issue here.

      Claude: Agree for the best-plant estimand. One caveat: picking the cheapest of three independently drawn process costs after they are known is optimistic, since a plant commits to a process before it learns its costs. Keeping some choice noise (the main model's β) or choosing on expected cost would be safer. At the defaults the effect is small: least-cost choice alone moves the share under $36 from 0.049% to 0.050%.

    8. Repeated issue with the model where technology which is cheaper on average is used even when simulated result makes it more expensive in some particular world.

      Claude: Confirmed in the main engine: its 'cheapest feasible mode' option chooses only the process, and cost-model.mjs charges the hydrolysate price whenever hydrolysates exist. Fixing it moves the median from $290 to $282/kg dry and the share under $36 from 0.049% to 0.050%: small, because in competitive worlds hydrolysate is nearly always the cheaper medium (it is dearer in 13% of all worlds that use it, but in 1 of the 486 competitive ones). Worth adopting in the main model.

    9. The original model has no limit to the scaling factor of capital, which means the best plant is just one which grows massive.

      Claude: Agree with the diagnosis; the fix needs an anchor. With 1,000,000 paired draws, capping at 6,500 m³ alone raises the share of worlds under $36/kg dry from 0.049% to 0.080%. The competitive worlds are high-productivity worlds that need little reactor volume (median 534 m³ under the original settings, below the 1,000 m³ at which reactor prices are quoted), so the cap gives them the price of a plant about 12 times larger. Cap at 1,000 m³: 0.053%; at 20,000 m³: 0.094%. Suggestion: an exponent on vessel size up to an elicited maximum vessel (oxygen transfer, mixing, sterility), a modest plant-level economy for shared utilities, and plant size bounded by what a 2036 market and lenders would support. More: https://unjournal.github.io/cm_pq_modeling/bates-rebuild.html