zero-trust-physics/zero-trust-physics-entry-heatshield

Atmospheric entry and heat-shield wear: the Stardust and InSight entries and the Stardust PICA shield, computed again with a proof on every row

For entry, descent and landing (EDL) and thermal protection system (TPS) teams, and for re-entry and hypersonic model verification and validation: two flown capsule entries and the heat-shield wear they drive, each computed by a physics model that anyone can call again, every row with a cryptographic proof and a post-quantum signature, beside what the flights recorded.

• Stardust, Earth, 15 January 2006. Nothing fitted. Peak load 32.17 g against 32.89 g in NASA's reconstruction (target 3.64 g); the drogue altitude reached 1.01 s and its Mach number 0.95 s from the 137.9 s the radar recorded (target 7.4 s).
• InSight, Mars, 26 November 2018. One number taken from the flight, the lift over drag measured on the first 150 s. Over the last 59 s before the parachute, which are not in the fit: speed within 7.2 m/s RMS, radius within 245 m and deceleration within 0.031 g over 58 whole seconds.
• The Stardust heat shield. The Stardust entry handed, frame by frame, into a PICA heat-shield model under one chain proof: nose wear 4.65 mm against 5.7 mm measured on the recovered shield (6.1 mm by X-ray). Gap: 1.05 mm, 18 percent low.

Zero-Trust Physics is a physics verifier for Super Intelligence (SI). SI agents ask it questions about real hardware, and every answer comes with a cryptographic proof, so any agent can run it again and get the same thing. Past what a model was compared with, it keeps going with the same equations and flags every step. It stops only at a physical limit, or where its equations, constants or tables no longer hold, and it names the reason.

https://mcp.zerotrustphysics.com/mcp Open to any agent.

What is in it

Two sealed datasets, one per capsule, and one chain.

Stardust: 15 rows under one dataset seal, 0fdc2bebaaff44b4c9ef318ab7ecac13f4d724e2297474bd0926a2a7877568f127fd6574853e207240d7c28464c70130:

• 1 row, the flown entry as its comparison ran it: from entry interface, 133.9 km, at 0.05 s steps, 2,779 steps to the drogue condition the radar recorded, every step recorded.
• 14 rows of a sweep around the flown entry: entry flight-path angle -0.5, -0.25, -0.1, -0.05, +0.05, +0.1 and +0.25 degrees from the flown entry at the same speed, and entry speed -20, -10, -5, +5, +10, +25 and +50 m/s at the same angle.
• Outcomes: 10 ended at the drogue condition, 5 refused by name: 4 where the speed through the air above 68.96 km leaves the band the drag table was computed for (AirSpeedOutsideCard), 1 where a steeper entry is above Mach 40 below 68.96 km (MachOutsideCard). 31,368 recorded frames: altitude, Mach number, load, heat flux, and at the nose the stagnation pressure, recovery enthalpy and heat-transfer coefficient.

InSight: 21 rows under one dataset seal, fa0ff4bff3b23d2d64634e92327d5379f275cf8cf88db705d3d50fb82e32160d9fd11a815d9f430915f14e1a036c30be:

• 2 rows, the two runs of the model's comparison, sent unchanged: the flown entry with the lift over drag measured on the first 150 s, -0.031, and with no lift; 0.1 s steps, every step recorded, from the locked entry state 126.5 km up.
• 19 rows of a sweep: lift over drag -0.08, -0.04, -0.02, +0.02, +0.04, +0.06 and +0.08 on the flown entry; entry flight-path angle -1, -0.5, -0.25, +0.25, +0.5 and +1 degrees; entry speed -100, -50, -25, +25, +50 and +100 m/s.
• Outcomes: 20 ended at the flown parachute deployment condition, 1 completed. 47,118 recorded frames: altitude, ground distance, Mach number, speed through the air, radial speed, dynamic pressure and load.

The heat-shield chain: 4 chained solves, each the Stardust entry sealed by the entry model and handed into the heat-shield solver on the open-record PICA model through the served coupling: every recorded frame's time, stagnation pressure, heat-transfer coefficient and recovery enthalpy, bit for bit, into the shield's heating history. The worked chain (a frame every 1.5 s, 93 frames, 58.2 mm of PICA from 260 K, the stream off at 139.95 s and the shield cooling to 600 s) gives the served chain seal exactly:

21a8ace2dc1223ed33f32f97173269980e469ca11d23bed4ccbcf41544fee9e55a3228621dd12a52452ffec09ddf80b1

Three variants the coupling's conditions admit: the shield starting at 300 K, a frame every 3 s (47 frames), and a 40 mm slab.

Each run row carries two columns written beside it in data/stardust_index and data/insight_index: compared_with_flight (the rows that are the model's own comparison) and outside_compared_range, with outside_by naming each limit the row goes past, its value and the extent past it: an entry state that is not the flown one, a speed through the air above or below the band the flown entry ran through, a state below the lowest the compared run reached.

How close it comes to the flights

Flight	Quantity	Model	Recorded	Difference	Target
Stardust	Peak load	32.170 g	32.89 g (NASA's reconstruction)	-0.720 g	3.64 g
Stardust	Time to the drogue altitude, 31.03 km	138.909 s	137.9 s (radar)	+1.009 s	7.4 s
Stardust	Time to the drogue Mach number, 1.23	136.953 s	137.9 s (radar)	-0.947 s	7.4 s
InSight, lift -0.031	Air speed, 58 whole seconds before the parachute		flight record	7.2 m/s RMS
InSight, lift -0.031	Radius, the same 58 s		flight record	245 m RMS
InSight, lift -0.031	Deceleration, the same 58 s		flight's inertial unit	0.031 g RMS
InSight, lift -0.031	Time to reach the flight's speed at the parachute	210.58 s	209.27 s	+1.31 s	15.1 s
InSight, lift -0.031	Time to reach the flight's radius at the parachute	208.69 s	209.27 s	-0.58 s	15.1 s
InSight, lift -0.031	Radius at the flight's deployment time			-114 m	3.5 km
InSight, lift -0.031	Mach number at the flight's deployment time	1.501	1.522	-0.021	0.2
InSight, no lift	Time to reach the flight's speed / radius at the parachute	226.40 / 232.10 s	209.27 s	+17.1 / +22.8 s	15.1 s
InSight, no lift	Radius at the flight's deployment time			+4.24 km	3.5 km
InSight, no lift	Mach number at the flight's deployment time	1.912	1.522	+0.390	0.2
Stardust heat shield	Nose wear, chained from the entry	4.65 mm	5.7 mm (core), 6.1 mm (X-ray)	-1.05 mm

The targets are NASA's own printed three-sigma spreads for its pre-entry forecasts. Stardust's peak load, 32.89 g, is NASA's fitted reconstruction, not an instrument reading: no accelerometer flew; the two times rest on range radar. InSight's lift was measured on the first 150 s of its own flight, and its peak load is in that fit: 7.95 g against 8.12 g (target 0.5 g). With no lift the InSight run is outside NASA's targets on the last 59 s: the card reproduces this entry given its lift; it does not forecast an entry whose lift is not known.

Gaps, one line each:

• Stardust heat flux at the nose 696 W/cm² against NASA's computed 762 to 1,000 W/cm²; the yearly standard atmosphere in place of that morning's air is worth about 1 g of peak load either way.
• InSight's air between 36 and 65 km is 8 to 13 percent denser than the flight's, and the peak load comes 4.0 s early.
• The chained nose wear is 1.05 mm under the core's 5.7 mm and 1.45 mm under the X-ray's 6.1 mm. The shield model fed NASA's own bare-wall heating instead gives 7.58 mm, a third high: the two heating sources bracket the core.

The rows that are the models' own comparisons reproduce the published numbers exactly: on Stardust 32.170162 g, 138.9086 s and 136.9531 s; on InSight 7.2 m/s, 244.9 m and 0.031 g, +1.309 s, -0.585 s, -113.5 m and Mach 1.5014.

Around the flown entries, the same equations: on Stardust, an entry 0.25 degrees steeper reaches 34.62 g and the drogue condition 6.7 s earlier; 25 m/s faster, 32.23 g. On InSight, lift over drag from -0.08 to +0.08 takes the peak load from 8.59 g to 6.78 g; held from the start, a lift of -0.04 or -0.02 reaches the flight's speed and radius at the parachute within 15.1 s of the flight, and -0.08, -0.06 and every lift above zero reach one or both outside it. The chain variants: the shield starting at 300 K wears 4.643 mm; a frame every 3 s, 4.648 mm; a 40 mm slab, 4.648 mm.

One row

Stardust row 0, the flown entry. The request, as run_trajectory takes it:

{"tool": "run_trajectory", "arguments": {"domain": "aerospace", "card": "capsule_stardust", "initial": {"radius_m": 6503140.0, "radial_speed_m_s": -1842.1416729435698, "transverse_speed_m_s": 12767.792058801891}, "dt_s": 0.05, "segments": [{"steps": 2779, "load": {}}], "record_every": 1}}

Outcome ended at PastFlightDeployment, the end the model declares: altitude under 31,030 m and Mach number under 1.23 at once, at step 2,779, 138.95 s. Peak load 32.170162 g at 59.4 s and 55.1 km; peak heat flux 696.14 W/cm² at 49.75 s. Largest conservation residual 6.5e-13. Run seal:

6d083ef7922ef23f729ed5acf3ffb7d66be9abadb6f74ff1012aaf7837742c4978e789888974892705aa5e808b1aa32b

Use it

The models are aerospace/capsule_stardust and aerospace/capsule_insight, and the heat-shield solver ablator_slab with the coupling stardust_entry_to_pica_slab, at https://mcp.zerotrustphysics.com/mcp. Read describe_card and describe_solver first. A chained request is solve with from: {"request": <the entry run's request>, "run_seal": <its seal>, "coupling": "stardust_entry_to_pica_slab"}; chain/worked/chained_request.json is one that runs unchanged.

Verify it yourself

Every row can be verified without trusting this page.

# 1. each dataset's proof, with no engine running: every row's seal, its chain of frames and the root
genesis_seal verify-dataset sealed/capsule_stardust/head.json sealed/capsule_stardust/rows.jsonl sealed/capsule_stardust/frames.bin
genesis_seal verify-dataset sealed/capsule_insight/head.json sealed/capsule_insight/rows.jsonl sealed/capsule_insight/frames.bin

# 2. one row's proof on its own
genesis_seal row-proof sealed/capsule_stardust/head.json sealed/capsule_stardust/rows.jsonl 0 sealed/capsule_stardust/frames.bin > row0.json
genesis_seal verify-proof row0.json

# 3. the entry run the chain started from
genesis_seal verify-run chain/worked/upstream_answer.json

4. Ask the engine to run a whole plan again: send verify_dataset the head's request with its dataset_seal.
5. Send any row's request to run_trajectory and compare the run_seal.
6. The chain: send verify_solution {"request": <chain/worked/chained_request.json>, "solution_seal": …}. The engine computes the entry again, holds every handed value to it bit for bit, solves the shield again and returns recomputed_chain_seal. Or send the answer back whole, byte for byte, with its from: the long form compares every byte.

When this dataset was made, all of these were run, and all held:

• Stardust: verify_dataset by the stored seal; its short form on a fresh engine, the plan run again; genesis_seal verify-dataset; row 12's proof on its own; 6 of 6 rows sent to run_trajectory again gave the same run seal, and every one of the 15 rows was computed again for the tables, each with its stored run seal.
• InSight: the same five; row 19's proof; 4 of 4 rows again; all 21 rows computed again for the tables, each with its stored run seal.
• The chain, all four: verify_solution by request, recomputed_chain_seal equal; the long form, holds and values hold; verify_tape and genesis_seal verify-run on the entry run; and the chain seal recomputed with no engine by the published rule (SHA-384 over the entry's run seal, the coupling's handed values with no spaces, and the solve's seal), with every handed value (four quantities at each of 93 frames, 47 in the 3 s variant) equal bit for bit to the entry's sealed frames and to the shield's request. The links form of verify_tape and genesis_seal verify-chain read a chain of runs and steps: they refuse a chain that ends in a solve ("the tape does not name a card"), and their answers are kept in chain/*/.

Signed

The row seals and the dataset seals were computed by the served engine build, run locally. The rows are not signed one by one: what is signed ML-DSA-87 is the dataset head the open address returned for the same plan, in parts; each signed head carries its merkle_root and rows_sha384 over its rows' run seals, and those run seals equal the stored rows', row for row. The worked chain is the exception: its whole answer was computed and signed by the open address.

The open address signs every answer it computes with the operator's answer key, ML-DSA-87 over its receipt, with an expiry of one hour and a nonce; with SHA-384, the CNSA 2.0 algorithms. Each plan here was sent again to the open address, rows unchanged, and the signed heads are in signed/:

• Stardust: 8 signed heads (signed/capsule_stardust/, 2 rows each, the last 1: an inline answer holds at most 10,000 frames). In every one the signature holds and every row's run seal equals the stored row's.
• InSight: 7 signed heads (signed/capsule_insight/, 3 rows each). The same: every signature holds, every run seal equal.
• The worked chain: signed/chain_worked_live_answer.json, the chained request answered by the open address in full, its receipt equal to the local answer's, chain seal 21a8ace2…ddf80b1, signature holds. The three variants are verified above and not signed.

# signed/answer_key_public.b64 is the public_key of https://mcp.zerotrustphysics.com/answer-key
genesis_seal verify-signature signed/chain_worked_live_answer.json --public-key signed/answer_key_public.b64 --now 2026-10-22T03:15:08Z
genesis_seal verify-signature signed/capsule_stardust/capsule_stardust_000.head.json --public-key signed/answer_key_public.b64 --now <its issued_at>

A signature is good for 3,600 s; verified at its own issued_at it shows the key holder issued it. Key id ca06027a972c81c8630fed9f4ea2d4e91596b9434a88a03b7eb325b24e40645b7090cb705927745e73ee8b272eb1be38.

Files

File	What it holds
data/comparison.parquet, .csv	the default view: each compared quantity, model, recorded value, difference, target and source
data/stardust_index.parquet, .csv	one line a Stardust row: its entry, outcome, peak load, heat flux, drogue times, compared range
data/insight_index.parquet, .csv	one line an InSight row: its entry and lift, outcome, peak load, times to the flight's deployment, RMS over the last 59 s
data/insight_last59s.parquet, .csv	the 58 whole seconds before the parachute: the flight's air speed, radius and their one-sigma, the sensed deceleration, and both compared runs
data/chain_index.parquet, .csv	one line a chained solve: wear, face temperature, depth of charring, flags, solution seal, chain seal, verifications
data/capsule_*/rows.parquet, frames.parquet	the sealed rows and frames, written only after each dataset verified
sealed/capsule_*/head.json, rows.jsonl, frames.bin	each dataset as the engine stored it; the head carries the plan that made it
chain/<variant>/	the entry run's request and answer, the chained request and answer, and every verification answer
signed/	the signed heads and the signed chained answer, with a summary of each verification

Where the numbers come from

• Stardust: Desai, P. N. and Qualls, G. D. (2008), Stardust Entry Reconstruction, AIAA-2008-1198, NTRS 20080008567: the peak load of the reconstruction, 32.89 g, and the drogue opening 137.9 s after entry interface at 31.03 km and Mach 1.23 (pp. 5 and 6), and the three-sigma spreads, 3.64 g and 7.4 s. NTRS: "Public Use Permitted." Inside the model: Desai, Lyons, Tooley and Kangas (2006), NTRS 20060028186 and 20090004445 (mass and entry state); Mitcheltree, Wilmoth, Cheatwood, Brauckmann and Greene (1997), AIAA 97-2304, NTRS 20040105538 (drag); the U.S. Standard Atmosphere 1976, NTRS 19770009539; World Geodetic System 1984 (NGA). The heat-flux range, 762 to 1,000 W/cm², is NASA's computations (Planetary Mission Entry Vehicles Quick Reference Guide, NASA/SP-20230010341).
• InSight: Karatekin, Ö.; Banfield, D.; Ashley, J. (2020). InSight EDL Bundle, NASA Planetary Data System, https://doi.org/10.17189/1518935: the reconstruction of the entry by Karatekin and Van Hove (Royal Observatory of Belgium), table edl_complete.tab, 200 samples a second; the 58 whole seconds from 151 s to 208 s and the last row, 209.2703 s, 2018-11-26T19:42:32.6144 UTC, are in data/insight_last59s. The DOI record reads: "This is a work of the U.S. Government and is not subject to copyright protection in the United States. Foreign copyrights may apply." Mach 1.522 at deployment and the surface pressure: Karlgaard and others (2020), NTRS 20200003204, "Public Use Permitted." Inside the model: the 2008 Phoenix aerodynamic database, Edquist, Desai and Schoenenberger, NTRS 20080034648; Korzun and others (2020), NTRS 20200002812; McCleese and Schofield, MRO Mars Climate Sounder DDR V6.2, NASA PDS, https://doi.org/10.17189/6khc-s052.
• The recovered Stardust shield: wear of the near-nose core 5.7 mm, 6.1 mm by X-ray scan, as the heat-shield model's description states it, from the cores cut from the recovered shield (Stackpoole and coauthors, AIAA 2008-1202). The PICA model is built from the open record: Tran and others (1997), NASA TM-110440, and single values from NTRS 20210024735, 20160000305, 20100022127, 20120017464 and 20110012891.
• The inputs: each row's request is in sealed/*/rows.jsonl, and the whole plan in each head. The flown rows are the requests the models' own comparisons ran: Stardust from its entry state at the model's largest step, 2,779 steps to the drogue end; InSight the two requests of its comparison record, unchanged. The chain's heating history is built as the engine's own worked chain builds it: row 0 at 0 s unheated, the 93 frames, then the stream off one second after the last frame and the shield cooling to 600 s, radiating to 266.48 K.

What the rows do not say

• Each model is a point mass: Stardust with tabulated drag, no lift, no wind and a yearly standard atmosphere; InSight with the 2008 Phoenix drag database, no wind and one held lift over drag. Neither carries attitude, sideways lift or the rise of drag in thin air.
• Each model was compared on one flown entry. A sweep row is a computation, not a measurement: past what a model was compared with, it keeps going with the same equations, and outside_by names how far each row goes past.
• Stardust's compared values are three: NASA's reconstructed peak load and the radar's drogue time, at its altitude and its Mach number. No heating was measured in flight.
• InSight's run within NASA's targets takes one number from the flight, its lift. With no lift the run is outside them.
• The chained shield stands for the near-nose core, not the flank. Its heating is the entry model's own stagnation-point heating, with no radiative heating; the solve flags that the history's first row, at 1.07 Pa, is below the lowest wall pressure of the Stardust comparison. No temperature inside the shield was measured in flight.

Benchmark tasks

Three questions on these models are tasks in Environments since version 0.5 (tag v0.5.0, 22 October 2026), each scored by a program that sends the agent's request to the verifier again, with no judge. Their inputs are not in this dataset: no row here shares a task's start state and loads, or differs from it in one input only, a limit task's open input set aside.

• stop-stardust-slower-entry (stop): Suppose the Stardust sample return capsule had reached entry interface, 6,503,140 m from the Earth's centre, at 12,780 m/s, 8.4 degrees below the horizontal. What would its peak heat flux at the nose have been, in W/m^2?
• figure-insight-peak-load-lifting-entry (figure): NASA's InSight capsule enters the air of Mars 3,522,200 m from the planet's centre with a radial speed of -1,300 m/s and a transverse speed of 5,760 m/s, flying at a lift over drag of 0.05 throughout. What is the peak load on it, in Earth g, before the parachute deploys?
• figure-tacot-slab-back-face-after-hold (figure): A 40 mm slab of the test heat-shield material TACOT 3.0, starting at 295 K, has its face held at 1,900 K for 45 s and is then followed for a further 30 s. What is the temperature of its back face at the end, in kelvin?

In Environments, python3 prompt.py <id> prints the text an agent is given, and python3 score.py <replies> --engine <verifier> scores the replies.

Cite

The paper: https://doi.org/10.5281/zenodo.23517223. The open address: https://mcp.zerotrustphysics.com/mcp; the public answer key: https://mcp.zerotrustphysics.com/answer-key.

Made by Super Intelligence, on behalf of Protocol Company. Data: CC BY 4.0; the flight records keep their sources' terms and attributions above.