An affective reliquary for the terminal.
Hinnom derives cryptographic keys from the pattern of a walk — where you stopped, for how long, how your body felt there, which way you turned afterwards — and lets you prove things about that walk without disclosing where it happened.
The name is taken from the sixth track of Bon Iver's Hinnom, TX: a burial ground outside Jerusalem, lifted and set down in Texas. A room that has been moved, intact, somewhere it does not belong. That is the object this program builds — a sealed room whose lock is a piece of your own geography, carried elsewhere.
Hinnom is a LANDSLIDE project. The lab builds strange things about agents and the ground under them; this one is only about the ground.
H I N N O M v0.1.0
affective trajectory keys · a room buried in a place
────────────────────────────────────────────────────────────────────────────
vault demo/vault · objects 1 · walk 15.41 km
Three claims, each implemented as code rather than as a paragraph.
- The key is a trajectory, not a location. Existing geographical-password systems ask you to choose points on a map. Hinnom never asks. It takes a recorded walk, segments it into dwell episodes, and encodes the shape of attention — relative dwell, affect, turn, spacing, whether you came back. No coordinate enters the key material.
- There is a model of how much that is worth. A conditional min-entropy estimator over a synthetic population, two attacker profiles, Shannon's guessing exponent, and a conversion to GPU-years. It reports a small, uncomfortable number and tells you how many episodes you would need.
- You can prove a property of the walk without giving the walk. A Groth16 circuit over BN254 proves "I stopped for at least N seconds inside this rectangle, and my arousal was in this band" while revealing only a MiMC commitment.
- Not Passwhere or GeoVault. Both treat remembered locations as the secret and study the entropy of choosing points (Corn & Podržaj 2026). Hinnom's input is involuntary: dwell durations and affect transitions that nobody publishes.
- Not GeoCrypt. That line of work encrypts files to a geofence and releases the key when you are physically inside it (GeoCrypt 2026). Hinnom has no unlock-on-arrival behaviour; the walk is the secret, not the gate.
- Not a drift app. The Naked City and similar tools give you prompts and send you walking (The Naked City). Hinnom is silent during the walk and does its work afterwards.
One screen, no chrome. The route is a braille plot laid over the place catalogue
it moved through, the episode ledger sits beside it, and the selected episode
carries the only ember on the page. These are frames rendered by
hinnom tui --snapshot 118x40 at 118×40, captured in Cascadia Mono.
The solitary dots are places from the catalogue; anything outside the frame is
simply clipped. They are ground, not evidence: nothing in the key is drawn from
them, and the map says so in its corner. A trajectory that has a places.json
beside it brings its catalogue along; otherwise --places names one.
Two pages sit over the map. e runs the entropy model
against a population and reports what the walk is worth; a checks a
Groth16 attestation against the verifying key in the directory.
The same walk at 80×24. The ribbon, the map and the masthead give up width before anything is dropped.
| attacker | knows | is defeated by |
|---|---|---|
| cross-walker | the population prior, the segmenter, every other user's corpus | affect and dwell distributions that are only weakly determined by place class and hour |
| companion | the above, plus your disposition at places you have visited together | the blend toward a per-walker posterior; measured at 45.2 bits per reference walk |
| device thief | the vault directory, including manifest, salt, and ciphertexts | Argon2id over the descriptor, then XChaCha20-Poly1305 |
| verifier | one attestation | the proof reveals a commitment, not coordinates |
Out of scope: an attacker who records your walk with their own sensors, and an attacker who compels you to reproduce it. Hinnom is a lock, not a witness.
A walk is segmented into dwell episodes by a two-pass rule: cluster fixes within a radius, then split at the largest internal gap. Each episode is reduced to a two-byte record in the HTJ1 encoding.
| field | width | meaning |
|---|---|---|
| dwell | 3 bits | dwell ÷ walk median dwell, 8 buckets |
| arousal | 2 bits | circumplex arousal, 4 buckets |
| valence | 2 bits | circumplex valence, 4 buckets |
| bearing | 3 bits | compass sector from the previous stop |
| distance | 2 bits | spacing ÷ median spacing, 4 buckets |
| inflection | 1 bit | affect change point detected inside the episode |
| returned | 1 bit | the same place appears earlier in the walk |
Three properties matter here.
Nothing absolute. Coordinates, timestamps, and absolute durations are absent. A GPS fix wanders by tens of metres between runs; a median-relative dwell does not.
Nothing chosen. You cannot pick a memorable route and get a weak key, because you are not choosing. You would have to consciously control how long you paused at each corner and how your heart rate responded, for an entire walk, and then remember it.
Nothing visible. The descriptor is derived from the vault's own salt plus the walk. The vault stores only a fingerprint of the result, sixteen hex digits, and only after a successful open.
The record stream is fed to Argon2id with a 32-byte vault salt, then split by
HKDF-SHA256 into three subkeys under distinct labels: vault, index, and
attest. The index key salts the MiMC commitment in the circuit, so a proof
made for one vault does not transfer to another.
| profile | memory | passes | lanes | use |
|---|---|---|---|---|
interactive |
64 MiB | 2 | 1 | demos, tests |
standard |
256 MiB | 3 | 1 | default |
paranoid |
1 GiB | 4 | 2 | high-value vaults |
hinnom entropy builds a synthetic corpus — 8 walkers, 6 walks each, 7
episodes per walk, 336 episodes, 54 candidate places — and estimates
conditional min-entropy per field, conditioned on place class and hour bucket,
with Laplace(1) smoothing. The companion attacker is a Dirichlet posterior
centred on the population prior. Real output:
field raw cross companion guess
arousal 2.00 0.67 0.57 0.63
valence 2.00 0.66 0.57 0.63
dwell 3.00 1.44 1.20 1.23
bearing 3.00 3.00 3.00 2.17
distance 2.00 1.20 1.20 0.88
inflection 1.00 0.73 0.73 0.48
returned 1.00 0.26 0.26 0.22
per episode cross 7.97 bits · companion 7.53 bits (raw 14)
reference walk cross 47.8 bits · companion 45.2 bits
MAP probability 4.05e-15 · 2.46e-14
guessing work 37.4 bits · log10 GPU-years 0.5 · -0.2
to reach 128 17 episodes by min-entropy · 21 by guessing work
location channel 5.75 bits per stop over 54 candidate places
the key ignores it entirely
Read that honestly. A six-episode walk carries about 45 bits against someone who has walked beside you. That is not a 128-bit key, and the program says so instead of rounding up. The last line is the point of the whole design: a location channel of 5.75 bits is what existing systems put their trust in.
The estimator is calibrated on synthetic data. It is an argument about the shape of the distribution, not a measurement of any city.
hinnom setup # Groth16 parameters, written to ./params
hinnom prove demo/walk.jsonl \
--episode 8 --radius 140 --min-dwell 600 \
--arousal 10-30 --valence 45-75 \
--out attestation.json
hinnom verify attestation.json
The circuit has 2,564 constraints and proves four things at once: the quantised coordinate lies inside a rectangle (bit decomposition with a booleanity check, so no field-wrap), the dwell exceeds the threshold, arousal lies in a band, and valence lies in a band — all under a MiMC-5 commitment that binds the five witness values to the vault's index key. The proof is 128 bytes; its public inputs are the commitment and nine small integers.
claim a quiet place · dwell ≥ 600s · arousal 10–30 · valence 45–75
commitment 1d4c0acd4f46e1d83a555a6a
proof 128 bytes
proved in 0.05 s
That is the prover alone. The whole command — reading 2,222 fixes, segmenting, running Argon2id at 64 MiB with two passes, then proving — takes about 0.3 s on a laptop.
The verifier learns that the claim is true. It learns nothing about the coordinates beyond their membership in a rectangle you supplied, and nothing about the walk beyond the two bands.
vault/
manifest.json salt, KDF profile, object list (id, kind, label, digest, sizes)
objects/<id>.hnm sealed envelope: "HNM1" | 24-byte nonce | ciphertext
XChaCha20-Poly1305 with the entry's id, kind, and plaintext digest as
associated data, so an attacker who swaps two objects in the manifest gets an
authentication failure rather than a wrong plaintext. Labels, sizes, and digests
are public by design; coordinates never leave the envelope.
| command | does |
|---|---|
hinnom demo --out demo |
builds a full demo: trajectory, places, vault, keys, attestation, entropy report |
hinnom init --vault v |
creates an empty vault |
hinnom walk walk.jsonl --vault v |
segments a trajectory and seals it |
hinnom unlock walk.jsonl --vault v --ledger |
opens every object this walk unlocks |
hinnom inspect --vault v |
prints the manifest without decrypting |
hinnom entropy |
runs the psychogeographic model |
hinnom setup / prove / verify |
Groth16 parameters, attestation, check |
hinnom tui --walk walk.jsonl --vault v |
the interface; places.json beside the walk is drawn under the trace |
hinnom tui ... --places FILE |
the same, with a catalogue named explicitly |
hinnom tui ... --snapshot 118x40 |
renders one frame and exits |
hinnom tui ... --snapshot 118x40 --ansi --overlay entropy |
the same frame with true colour, opened on a page |
Trajectories are JSONL or CSV. The loader accepts t, lat, lon, and
optionally hr, rmssd, audio_db, audio_centroid, cadence, valence,
poi.
cargo build --release
cargo test --release # 36 unit tests, 6 end-to-end
Requires Rust 1.89. The release profile uses thin LTO; the whole binary is a few megabytes and has no runtime dependencies.
The interface is specified in DESIGN.md: bone/ash/ember/river, no
rounded corners, no cards, one ember element on screen, braille as the only
raster, one slow pulse as the only motion. The catalogue is drawn in field, a
shade below ash, so the ground is legible without competing with the route.
Hinnom is built in the open under LANDSLIDE, a group that makes strange things and mostly makes them about agents. The kinship here is one of notation rather than plumbing.
- MONOLITH reads a repository as a geologic body: commits as sediment, reverts as erosion, a slow collapse observed and never repaired. Hinnom reads a body as a geologist would read a valley — a cross-section, taken afterwards, with no attempt to correct it.
- Tiltmeter gives an agent skill a provenance chain. An attestation here is the same move for a human claim: a document that says who proved what, under which circuit, against which verifying key.
- saltglass keeps knowledge that expires unless it is reverified. A Hinnom proof is bound to one vault salt and one circuit, so it cannot be replayed into another context; nothing in it survives the salt being discarded.
There is no wire between these programs yet, and none is claimed. What is shared is a discipline: deterministic artefacts, a single frame, no dashboards, and a willingness to print an uncomfortable number rather than round it up.
The one seam worth naming is Sphinx, the approval plane that stops dangerous actions at a human. An approval is only as good as the evidence that a human was there to give it. A Hinnom attestation is evidence of a body in a place, but not yet at a time: the circuit proves dwell inside a rectangle, not dwell inside a window. A time-window claim is the first thing this circuit would need before an agent could ask it for anything.
- Segmentation is part of the key. Changing the dwell radius or the gap threshold changes the descriptor and therefore the key. There is no versioned migration yet; a real deployment needs one, and needs to treat the segmenter as frozen infrastructure.
- Without a passphrase, the trajectory is the only secret. An attacker who obtains both the vault and the walk can open it. Argon2id buys time, not entropy.
- Affect is a proxy. Valence and arousal are inferred from HRV, cadence, and audio; they measure the body's state, not the feeling.
- The setup is local. Groth16 needs a per-circuit trusted setup. The parameters here are generated on your machine and are not auditable by a third party. A deployed system needs a ceremony or a universal setup.
- The circuit is narrow. Region membership, a dwell threshold, two affect bands. Time-window claims ("in the last month") and path-shape claims are designed but not implemented.
- The corpus is synthetic. The entropy figures establish an argument, not a measurement. A field study is the obvious next step.
- No sensor capture. Ingest is a file. Phone and watch capture is where the reproducibility question actually gets decided.
- No remote storage. The optional encrypted-IPFS path is not implemented; a vault is a directory. The object format already supports it — the envelope is self-contained and the manifest only needs the digest — but pinning, retrieval and a second copy of the manifest are not written.
1. Corn, M.; Podržaj, P. GeoVault: Leveraging Human Spatial Memory for Secure Cryptographic Key Management. Mathematics 2026, 14(10), 1653. https://www.mdpi.com/2227-7390/14/10/1653
2. GeoCrypt: Location-based Secure File Encryption and Access Control with GPS Spoofing Detection. IEEE, 2026. https://ieeexplore.ieee.org/document/11382983
3. GeoCrypt: A Location-Derived Multi-Stage Encryption Model for Adaptive and Resilient Secure Communication. IEEE, 2026. https://ieeexplore.ieee.org/document/11416151
4. Passwhere — a geographical password system. https://www.passwhere.com
5. The Naked City — a dérive app after Guy Debord. https://walklistencreate.org/walkingpiece/the-naked-city/
- Debord, G. Théorie de la dérive. Internationale Situationniste 1956, 2.
- Debord, G. Guide psychogéographique de Paris (The Naked City), 1957.
- Russell, J. A. A circumplex model of affect. J. Pers. Soc. Psychol. 1980, 39(6), 1161–1178.
- Massey, J. L. Guessing and entropy. Proc. IEEE Int. Symp. Inf. Theory 1994, 204.
- Dodis, Y.; Reyzin, L.; Smith, A. Fuzzy extractors: how to generate strong keys from biometrics and other noisy data. EUROCRYPT 2004, 523–540.
- Krumm, J. A survey of computational location privacy. Pers. Ubiquitous Comput. 2009, 13(6), 391–399.
- RFC 9106 — Argon2 Memory-Hard Function for Password Hashing and Proof-of-Work Applications. IRTF, 2021.
- Groth, J. On the size of pairing-based non-interactive arguments. EUROCRYPT 2016, 305–326.
- Albrecht, M.; Grassi, L.; Rechberger, C.; Roy, A.; Tiessen, T. MiMC: efficient encryption and cryptographic hashing with minimal multiplicative complexity. ASIACRYPT 2016, 191–219.
- XChaCha: eXtended-nonce ChaCha and AEAD_XChaCha20_Poly1305. draft-irtf-cfrg-xchacha, IRTF.



