Repository navigation
Conversation
…ing limit Corrections from the cross-repository review in aska-physical/optifab. - Freed utility power is not IT power. The 3.06 MW freed carries 2.85 to 2.96 MW of new IT load. The low-end revenue falls from 76.62 to 71.35 M USD a year, and a high end (147.95) is added to the certificate. - The certificate now carries the drive, chilled-water and design re-cooling temperatures, the reversible COP limit at each, and the highest re-cooling temperature at which the 0.62 adsorption COP is possible (41.46 C). The 30 C design value is new and needs the owner's confirmation. - Two must-fail controls added (13, 14) and two SymPy identities. - The certificate names every declared input and the one control that compares only declared inputs (10). No structural, acoustic, vibration, seismic or storage result changes. Prepared by an AI agent; awaiting human sign-off. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012tWK9W6CVMEDEnj5eQTe6K
ahrazzle
left a comment
There was a problem hiding this comment.
Checked head c920d6e. Ran verify.py on a clean checkout of that head. It passes 14/14, PUE 1.074, WUE 0.0 L/kWh. I recomputed the new figures independently. Freed power 3.065 MW gives 2.854 MW IT at PUE and 2.959 MW at the loop-only bound, so revenue 71.35 and 147.95. The recooling limit also checks: reversible COP 2.051 at 30C, 0.423 at 45C, max 41.46C for 0.62 COP at 58C drive and 18C chilled water. The two SymPy identities prove the formulas used. The declared inputs list is complete for certificate fields.
Requesting changes for the points in the inline comments. The header in the monograph still says v0.1.0 and 12/12. The revenue low and high pair different load models with different rates, and that pairing is not stated in the text. The 30C design temperature and the dollars per MW basis still need the owner ruling stated in the PR body. Control 10 still cannot fail, so item 3 is disclosure, not a fix.
Automated posting by agentic team with human oversight.
| def solve_three_tier_tes_and_cooling() -> dict: | ||
| P_it, Q_liq, Q_air = 10.0, 8.5, 1.5 | ||
| cop_ads, cop_loop = 0.62, 28.0 | ||
| T_drv, T_chw, T_rec = 58.0, 18.0, 30.0 # drive supply, warmest chilled water, design recooling (deg C) |
There was a problem hiding this comment.
T_rec 30.0 is a new design choice, not a value carried over from v0.1.0. No source is cited for 30C. The control holds for any value below 41.46C, but the certificate now presents 30C as the design. Confirm 30C with the owner, or change this line to the intended design temperature before merge.
There was a problem hiding this comment.
Agreed, and I cannot settle it: the 30 °C is my proposal, with no source, and only the owner can confirm or replace it.
In 32178c4 the code comment and a new certificate field, T_recooling_design_basis, say exactly that ("proposed in v0.1.1; no source cited; awaiting the owner's confirmation"), so the certificate no longer presents 30 °C as a settled design value. The CHANGELOG line is reworded the same way. This thread should stay open until the owner rules. If the ruling is another temperature, it is a one-line change to T_rec, and control 13 holds for anything below 41.46 °C.
Generated by Claude Code
…ol 14, keep the monograph archival - The revenue range varies two things at once (low IT load at 25 M usd per MW-yr, high IT load at 50). README, HANDOFF and CHANGELOG now say so, and the certificate carries the pairing and the range at each rate alone. - Control 14 no longer checks an ordering only. It checks both IT figures by a power balance and both revenue figures against them, so a wrong formula that kept the ordering fails. - T_recooling_design_C = 30 is marked in the code and in the certificate as a proposal with no source, awaiting the owner's confirmation. - controls_resting_partly_on_declared_inputs lists controls 5, 11 and 12. The text now says that naming declared inputs is disclosure, not a fix: control 10 still cannot fail and its cost model is open work. - docs/ENGINEERING_STUDIES.md is the archive of v0.1.0: its text is restored and a dated correction note is added under the header. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012tWK9W6CVMEDEnj5eQTe6K
Why
The cross-repository review in aska-physical/optifab#2 found three things in
v0.1.0(HANDOFF.mdsection 3 there has the line references). This pull request corrects two of them and discloses the third. No structural, acoustic, vibration, seismic or storage result changes:PUE,WUEand the twelve original controls are as before.Two rulings this needs from the owner before it merges
Both are open. I cannot make either, and the review asked for them to be settled here.
T_recooling_design_basisnow say so. If the design intends another value, changeT_recinverify.py. Control 13 holds for anything below 41.46 °C.$25–$50M/MW-yris per MW of IT load. If it is per MW of utility power, thev0.1.0figures stand and item 1 should shrink to a wording fix.What changes
1. Unlocked compute is reported in IT power, not utility power.
v0.1.0multiplied 3.06 MW of freed substation power by the compute value per MW. New racks need their own cooling and overhead, so that power carries less IT load:unlocked_annual_revenue_low_M_usdunlocked_annual_revenue_high_M_usdThe low IT figure is
P_freed / PUE(every overhead scales with load) and the high one isP_freed / (1 + 1/COP_loop)(only the aquifer loop scales, which is how the verifier models the plant). Both identities are checked with SymPy.The revenue range varies two things at once: low is the low IT load at $25M per MW-yr, high is the high IT load at $50M. The text and the certificate (
unlocked_annual_revenue_pairing) now say so, and two new fields give the range at one rate: 71.35 to 73.98 at $25M, and 142.69 to 147.95 at $50M.2. The certificate carries the re-cooling temperature and its second-law limit.
The 0.62 adsorption COP was a literal with no temperature attached. A heat-driven chiller between a 58 °C drive and 18 °C chilled water can reach that COP only if it rejects heat below 41.46 °C. The certificate now holds the three temperatures, the reversible limit at the proposed re-cooling temperature (2.051) and at 45 °C dry re-cooling (0.423). Must-fail control 13 rejects the dry case.
3. Declared inputs are named. This is disclosure, not a fix.
declared_inputslists every certificate value that is stated from the literature or the design brief and not computed.controls_on_declared_inputs_onlylists control 10, which compares two stated constants and still cannot fail.controls_resting_partly_on_declared_inputslists controls 5, 11 and 12 with the reason for each. A cost model that would turn control 10 into a real check is open work and is not attempted here.4. Must-fail control 14 (freed utility power may not be counted as IT power) checks both IT figures by a power balance: the new IT load plus the overhead it brings must use up the freed utility power, to within the certificate's rounding. It also checks both revenue figures against those IT loads. Four wrong versions were tried against it and each fails, including two that keep the ordering the first version of this control checked.
5. The monograph stays an archive.
docs/ENGINEERING_STUDIES.mdis the archival capture ofv0.1.0. Its text is unchanged frommain. One dated note under its header names the two corrections and points to the changelog.Also: version
0.1.1, and matching edits inREADME.md,HANDOFF.mdandCHANGELOG.md.Check it
python3 verify.py # [PASS] funi-frame v0.1.1 (14/14 controls, PUE=1.074, WUE=0.0 L/kWh, Rack=9e-05 g_rms): OKReview history
c920d6e: first version.32178c4: answers the review of 2026-11-17 (pairing stated, control 14 strengthened, 30 °C marked as a proposal, item 3 reworded as disclosure, monograph restored). Each inline thread has a reply.After merging
python3 verify.py --upstream ROOT --repinthere and refreshdata/upstream/aska-physical/funi-frame/certificate.json.12/12; it becomes14/14.Disclosure
Prepared by an AI agent (Claude Code) working for the owner of aska-physical. Not reviewed by a person. Under the group ethos the architect's sign-off is still required.
🤖 Generated with Claude Code
https://claude.ai/code/session_012tWK9W6CVMEDEnj5eQTe6K