Skip to content

Releases

Pharos v3.6.1

Choose a tag to compare

@github-actions github-actions released this 15 Oct 09:44

Ten commits since v3.6.0. v3.6.0 came from fuzzing the verdicts; this release
came from taking the 50 properties that audit produced, re-checking every one
against the code as it now stands, and putting each survivor through three
adversarial readers. Twenty-two of the fifty did not survive that - several
because the property was never a rule this project made - which is the point
of running the panel at all.

An ordinary roam was billed as a PMKID orphan. ph_settle() took no clock,
so it could not tell a message 1 that was two milliseconds old from one that
was ten seconds old, and the lens calls it twenty times a second. A client's
message 1 was retired before its message 2 arrived, and nothing decrements the
counter afterwards. A PMKID in message 1 is how fast roaming works, so two
ordinary roams on an ordinary WPA2 network reached SUSPECTED at 46/96 with "A
PMKID was requested and never completed".

A working building accused itself of being jammed. frames_total * 1000u
is a 32-bit product and wraps at 4,295,000 frames - about half an hour of
camping on a dense channel. The wrapped value decides barren, and a loud
channel reading barren is DENIAL. 4294 identical dwells grade CONGESTED at 42;
the 4295th, bit-identical, grades DENIAL at 74.

A block of flats reported a beacon flood. The novelty rate divided a
LIFETIME count of names by the caller's twelve-second window, so the number
stopped falling as the same names spread over a longer time: forty names over
sixty seconds and over six hundred both read 200.0 a minute.

Two strangers could be merged into one person. Probe's link search took the
first candidate it met and stopped, so when two tracked devices both matched,
array order decided which person a new address joined. Fingerprints are shared
by every phone of a chipset and the sequence test accepts a 64-wide window, so
this is not rare. A link is now claimed only on a unique match, and the refusal
is counted rather than silent.

One device showed two opposite claims on one screen. The roster row said
UPLOADING while the expansion said "nothing exposed on air" with a green ok,
because the headline had no branch for that exposure. And "uploading
continuously" did not mean continuously: one frame every ninety seconds
cleared every threshold at 100% duty over four hours.

A band plan scored as suspicious behaviour. Twin's "alone on this channel"
is true of whoever was picked as suspect when every radio has its own channel,
which is the ordinary 5 and 6 GHz layout.

A device was expired while it was still transmitting, and seven
vocabularies were cut in the field they are drawn in
- including all eleven
CWE references, where a reference cut in half names a weakness nobody can look
up. Where a slot had slack it was widened rather than the words shortened.

Two notes shared a bit in Vigil, latent only because the engine never set
one of them yet.

AND THE SUITE ITSELF HAD A FALSE GREEN. make compares modification times with
one-second granularity, and the whole host suite is a single cc invocation, so
an edit landing in the same second as the previous build ran the OLD binary and
printed "0 failures" about code that was not on disk. That happened three times
in one session, twice right after restoring a file to check a test really
catches its bug. The binary now always relinks.

Every fix was verified by reverting it and watching the tests name the failure.
Three of the new assertions were wrong on the first attempt and passed against
broken code; each was caught the same way.

15,227 host checks, 515,084,483 parser post-conditions and 50,704,513 verdict
invariants, all clean under ASan and UBSan. All 21 lenses photographed off the
panel, a nine-minute soak with zero heap drift and zero fence trips, and all
nine armed watches visited across eight tower rotations.

Flash it

From a browser - Chrome or Edge, no tools to install:
https://at0m-b0mb.github.io/Pharos-ESP32S3/

Or with esptool, using the single merged image:

pip install esptool
esptool.py --chip esp32s3 write_flash 0x0 pharos-v3.6.1-esp32s3.bin

Upgrading from a build older than v1.4.0? Erase first - the partition
layout and PSRAM settings changed:
esptool.py --chip esp32s3 erase_flash

What is attached

file what it is
pharos-v3.6.1-esp32s3.bin the whole thing, flashes at 0x0
pharos-esp32s3.bin same image at a version-less name, for scripts
pharos-v3.6.1-esp32s3.zip the above plus component images and FLASH.md
SHA256SUMS.txt checksums for all of it

Audited in CI against this exact ELF

  • Transmit fence - no reachable path to a transmit primitive. This
    firmware receives only; the fence is enforced at link time, not by
    convention.
  • Lens linkage - every lens survived the link. A lens registers from
    a constructor, so the linker will happily discard one and still
    produce a clean, empty-feeling firmware.
  • Display path - the panel is driven via Waveshare's BSP.

A serial console runs at 115200 - type help. See
docs/HARDWARE.md.

Pharos v3.6.0

Choose a tag to compare

@github-actions github-actions released this 15 Oct 04:15

Ten commits since v3.5.1. v3.5.0 and v3.5.1 came out of an adversarial audit
reading the code. This release comes from asking a different question: not "is
this line wrong" but "what must be true of every input, and is it?"

Two new harnesses answer it. The parser fuzzer from v3.5.1 proves the engines
survive hostile bytes; a second one drives the SCORING engines with random
event streams and checks the invariants after every step, with time
deliberately misbehaving - out of order, repeated, jumping by an hour - because
the Wi-Fi callback, the NimBLE host task and the dwell ring are three producers
with their own clocks reaching one consumer. 515,084,483 parser post-conditions
and 50,665,338 verdict invariants, all under ASan and UBSan.

Five bytes in front of a tracker made it invisible. The worst of them, in
the lens whose whole job is to find a tracker somebody planted on a person.
pv_classify() returned the first AD structure it recognised, and an
advertisement holds several that whoever is in radio range chooses. So a
five-byte Eddystone frame in front of a real tag classified the frame as a
proximity beacon - and Vigil rightly refuses to table a beacon. The tracker was
never entered at all: score 0, band CLEAR, zero tags, for a tag heard across
three places by someone who had moved. This file had already paid for this bug
once with an UNRECOGNISED prefix; it came straight back through a recognised
one, and every decoy fixture in the tests used the unrecognised kind. The walk
now sees every structure and keeps the most alarming kind it found. A
beacon-only advertisement is still a beacon, and still refused.

And the separated verdict came from the wrong byte. In a Find My element
the status sits at p[4]. An element too short to hold one was read at
p[plen - 1] instead - the length byte - so the six bytes 05 FF 4C 00 12 04
classified as SEPARATED. That is remembered permanently and scores +6, which
means six bytes anybody can transmit permanently escalated a device.

A device was expired while it was still transmitting. Every ingest path
stamps last_us so rd_expire() means something; rd_observe_traffic did
not, and admit() only stamps first_us. So a device known only from its
traffic was dropped on the next sweep, and one heard earlier aged out
mid-stream. The device transmitting most was the likeliest to vanish - which
inverts the point of watching traffic, since the upload-duty measure exists to
find a camera that streams continuously.

The 802.11 random-address bit means nothing to Bluetooth. admit() guesses
whether an address is the manufacturer's or the privacy layer's from the
802.11 locally-administered bit. Bluetooth marks privacy with the top two bits
of the most significant octet, and the bit being read is one of the 46 random
ones below them - so half of every privacy address on the air read as burned-in
hardware and had a manufacturer looked up in three random bytes. A phone hiding
itself could be published as a Hikvision camera while the row beside it
correctly declined to call it trackable.

Twin's verdict depended on the order the scan filled its table. Three
decisions pick one member out of a group, and all three were strict
comparisons, so a tie fell to whichever access point sat earlier in an array
filled in whatever order the scan found things. One unchanged room could score
78 on one sweep and 88 on the next, and across a band boundary that changes the
headline word. Ties break on the lowest address now.

Aegis would publish certainty if a contributing engine regressed. It takes
each detector's score and ceiling on trust, and the overall ceiling is a
MINIMUM - so an impossibly high contributor never lowered it and went straight
to the glass as 100. Clamped at the door, and the aggregate ceiling is bounded
by PA_CEILING_MAX, which is 99 and not 100 on purpose.

An unmeasured window was read as a ten-second one. window_ms of 0 means
the caller never timed the sweep; it was substituted with ten seconds and used
as the denominator of a churn rate. The same twenty-four networks score 58 with
an unmeasured window and 51 with a measured 24-second one - so failing to time
the sweep made the detector MORE confident.

And the second advice row was budgeted at the first row's width. On a round
panel the lower row sits on a shorter chord: 27 characters against 20. Any
second line between the two was budgeted as if it fitted and then silently
ellipsized by the label underneath. The splitter also filled line one to the
brim, stranding single words - "access point - tap for the" over "list". Each
row now carries its own cap and the break is balanced between them.

And ten of fifty-six vendor prefixes named the wrong company. Nobody had
checked k_ouis against the IEEE registry. Hikvision and Dahua were each
other, so each vendor's cameras were reported as the other's. 00:05:5D is
D-Link not Netgear, 00:0F:7C is ACTi not Axis, CC:F7:35 is Amazon not Roku,
30:05:5C is Brother not HP. C8:3F:26 belongs to Microsoft and was labelled
Fitbit; D8:8C:79 belongs to Google and was labelled Garmin, and both carried
RD_WEARABLE, so the device class was derived from a vendor that was not
there. 00:62:6E is in no registry at all, and DA:A1:19 cannot be in one - DA
has the locally-administered bit set, so the table was naming a RANDOMISED
address as a Google media device.

tools/check_ouis.py keeps it honest, and is worth running before trusting a
new row: 57 entries, 57 confirmed. The idea came from SquachWatch, which
published the same audit of its own table and found a prefix labelled
"Vigilant" for eleven releases that is registered to Sonos. Checked against
the registry, not against their table.

Every fix was verified by reverting it and watching the tests name the failure.
Two of the new harness's own invariants were wrong and were caught the same way

  • by checking the engine rather than believing the test.

Flash it

From a browser - Chrome or Edge, no tools to install:
https://at0m-b0mb.github.io/Pharos-ESP32S3/

Or with esptool, using the single merged image:

pip install esptool
esptool.py --chip esp32s3 write_flash 0x0 pharos-v3.6.0-esp32s3.bin

Upgrading from a build older than v1.4.0? Erase first - the partition
layout and PSRAM settings changed:
esptool.py --chip esp32s3 erase_flash

What is attached

file what it is
pharos-v3.6.0-esp32s3.bin the whole thing, flashes at 0x0
pharos-esp32s3.bin same image at a version-less name, for scripts
pharos-v3.6.0-esp32s3.zip the above plus component images and FLASH.md
SHA256SUMS.txt checksums for all of it

Audited in CI against this exact ELF

  • Transmit fence - no reachable path to a transmit primitive. This
    firmware receives only; the fence is enforced at link time, not by
    convention.
  • Lens linkage - every lens survived the link. A lens registers from
    a constructor, so the linker will happily discard one and still
    produce a clean, empty-feeling firmware.
  • Display path - the panel is driven via Waveshare's BSP.

A serial console runs at 115200 - type help. See
docs/HARDWARE.md.

Pharos v3.5.1

Choose a tag to compare

@github-actions github-actions released this 15 Oct 00:30

Four commits since v3.5.0. The adversarial audit that drove v3.5.0 confirmed
twenty-three findings; v3.5.0 closed the sharp ones and left a tail. This
closes the rest of what was worth closing.

Rival latched a BLE-spam alarm and nothing could clear it. The
advertisement-rate ring counts distinct advertisers per second across eight
slots, and it only cleared the slot it was moving INTO - and it only moved when
an advertisement arrived in a new second. So the instant BLE went quiet, the
remaining slots kept whatever burst had filled them and the peak was read
straight back out of them for ever. One burst, one permanent alarm. The pairing
counter two fields away already stamped each slot with the second it belonged
to; this one did not, and now does.

Sentinel published an access point before it had filled it. The table
handed back &s_aps[s_n_aps++] for the caller to populate. on_event runs on
the analytics core and takes no lock; sentinel adopt and the tick read the
table from other tasks and do take one - so the lock protected the readers from
each other and nothing at all from the writer. In the window between the
increment and the fill, a reader saw a record with a BSSID and nothing else: no
name, channel zero, no RSN.

That is worse here than it sounds, because what the reader is usually doing is
adopting a baseline - the thing every later sweep is compared against. A
half-written access point in the baseline is a permanent, quiet false "changed"
for the rest of the session. The count is the publication point, so it moves
last now, behind a release fence.

rotate drove the panel from the console's task. Its only caller is the
rotate command, and it re-oriented the panel while the LVGL task was free to
be flushing a frame out of the same bus - the same class of fault as the
console re-initialising the Wi-Fi driver on its own task, fixed the same way.
It takes the display lock now, for the rotation only; the NVS write that
remembers the choice stays outside it, because holding a display lock across a
flash write is how a frame gets dropped.

And the release page says what changed. The release job waits for the
maintainer to write notes and drops a one-line placeholder if they do not
appear - which is what the entire v3.0.0 through v3.5.0 range shipped with. The
notes are composed from this file now, every run, with how to flash it and what
CI proved about the exact image attached. Hand-written notes still win; this
only decides what a release says when nobody writes one.

8901 host checks, 0 failures - and 0 under UBSan and ASan. Fence, lens and
source-wiring audits intact. All twenty-one faces photographed off the glass.

Flash it

From a browser - Chrome or Edge, no tools to install:
https://at0m-b0mb.github.io/Pharos-ESP32S3/

Or with esptool, using the single merged image:

pip install esptool
esptool.py --chip esp32s3 write_flash 0x0 pharos-v3.5.1-esp32s3.bin

Upgrading from a build older than v1.4.0? Erase first - the partition
layout and PSRAM settings changed:
esptool.py --chip esp32s3 erase_flash

What is attached

file what it is
pharos-v3.5.1-esp32s3.bin the whole thing, flashes at 0x0
pharos-esp32s3.bin same image at a version-less name, for scripts
pharos-v3.5.1-esp32s3.zip the above plus component images and FLASH.md
SHA256SUMS.txt checksums for all of it

Audited in CI against this exact ELF

  • Transmit fence - no reachable path to a transmit primitive. This
    firmware receives only; the fence is enforced at link time, not by
    convention.
  • Lens linkage - every lens survived the link. A lens registers from
    a constructor, so the linker will happily discard one and still
    produce a clean, empty-feeling firmware.
  • Display path - the panel is driven via Waveshare's BSP.

A serial console runs at 115200 - type help. See
docs/HARDWARE.md.

Pharos v3.5.0

Choose a tag to compare

@github-actions github-actions released this 14 Oct 22:30

Twenty-nine commits since v3.4.0. Where v3.4.0 audited the engines against the
published research, this release audits the device against itself - and the
tool that made it possible was teaching Pharos to send back a picture of its
own screen.

screen dump already existed. It was lying. It emitted bare base64, so any log
line that landed mid-transfer was spliced into the payload by the decoder and
every pixel after it was invented; it also held the LVGL lock for the whole
six-second transfer, starving the UI task, whose lock-failure log then landed
inside the payload it was corrupting. Every line is marked now, the stream is
checksummed, the lock is released before the wire work, and the host refuses
to write a file it cannot verify
. tools/screen/sweep.py photographs all
twenty-one faces into a contact sheet and bounds-checks each real capture
against the safe radius - the same check tools/render does on the simulation,
but on what the panel actually drew.

Nearly every user-visible fix below was found by looking at the result.

The second advice line was never built. s_l_action2 was declared, cleared
on the idle path, written to on the live path - and mk_label was never called
for it. Both setters begin if (!o) return;, so every write was a silent
no-op, and the device had never once shown the half of the advice that says
what to DO. That is the exact failure the two-line layout was introduced to
fix.

The ceiling tick pointed 135 degrees wrong. The arc carries the gauge's
start angle as its LVGL rotation, so angles handed to it afterwards are already
relative - and the tick added the start a second time. Because the arc opens at
the bottom, mid-range ceilings landed in that opening and printed across the
advice. Measured on the glass: a ceiling of 56 drew at 62 degrees when it
belonged at 286.

The advice did not fit, and most of it was discarded before the UI saw it.
The display contract carried 96 characters against a median engine string of 93
and a longest of 525 - it truncated 48 of the 101 strings in the tree. The
contract is 256 now, and the whole text has a page of its own across the wide
middle of the circle. The bottom strip has three stops: the verdict, then what
to do, then the numbers.

The home face drew names through their own dots. The layout placed dots at
radius 168 and the renderer drew them at 178, so every clearance the host tests
proved was about a ring that is not on the glass. The scorer also never checked
the dot at all - a label shares its dot's angle, so at the sides of the ring it
needs half its own width in radial clearance - and it asked whether the WIDEST
possible name would fit, for every name.

What the air could say that it should not have

A crafted deauthentication frame could name any router. The WPS, RSN and
vendor-element parsing was gated on a test that answers "is this body long
enough to hold its fixed parameters" - which, for every subtype that has no
fixed parameters, is unconditionally true. So it ran on deauth frames, whose
two-byte reason code became the first tag/len pair. One frame from anyone in
range made the roster assert a make, model and an exposed WPS PIN against any
BSSID the sender chose, and suppressed that AP's genuine model for the rest of
the capture.

Two radios were pushing to a ring that takes one producer. Rival, Roster
and Vigil each handed the same ring to the Wi-Fi start and the BLE start, so
two tasks on two cores pushed to a ring that is single-producer by
implementation. The payload is a union, so a torn write could pair a BLE event
type with a length byte taken from an attacker's 802.11 frame, against a
31-byte buffer. Each has its own ring now, and the radio layer refuses a shared
one rather than documenting that it must not be.

An unauthenticated timestamp overflowed the clock fit. A beacon's TSF is
sixty-four bits and nobody signs it; the skew engine stored it, subtracted two
of them and multiplied by a million, all signed. Anyone in range could
manufacture a "the crystal behind this name changed" finding. The suite now
runs clean under -fsanitize=undefined,address, which is how it was pinned
down - and how two more were found.

Somebody else's text reached the console raw. SSIDs and BLE names were
drawn and printed exactly as received. A byte the font lacks renders as an
empty box, and an ESC byte starts an ANSI sequence on the operator's terminal -
a receive-only device as the delivery mechanism.

Telling the truth about what it heard

A four-byte decoy hid a tracker from Vigil. The classifier abandoned the
whole advertisement on any Apple manufacturer element that was not Find My,
and anything classified unknown is dropped - so a scrap of Apple-shaped data in
front of the real element made a Flipper, or a SmartTag's own service UUID,
invisible. The same classifier called every Samsung manufacturer advertiser a
SmartTag: company 0x0075 is Samsung, carried by every Galaxy phone, every pair
of Buds and every Watch, all of which were entering a stalking lens as
trackers.

Roster threw away every advertisement it was handed. It passed null for the
name and zero for the appearance, so the name-leak exposure was unreachable and
every BLE-only device fell through to "wearable" - printers and speakers
included.

Every BLE address was stored backwards. NimBLE hands them over
least-significant byte first. The OUI table read the device half, the
random-address class was read off a device byte, and the last three bytes shown
to tell one device from another were its vendor bytes. Confirmed against live
air before changing anything: six of seven Apple advertisers in the room
carried the resolvable-private-address signature in the last byte, none in the
first.

The battery cried wolf. The AXP2101's counter restarts after a reset and
re-learns, so a charging pack read 43%, then 0%, then climbed - and the device
drew that in alarm red. A reading the pack voltage contradicts is marked
approximate now, and nothing raises an alarm on a number flagged that way.

Ward discarded fifteen of the sixteen networks it had learned. Its eviction
set the list length to the faintest slot's index rather than replacing it, so
everything above that slot fell outside the list, and "adopt nearest" then
guarded the nearest of whatever survived.

squall camp had never done anything. The command had advertised it since
it was written - it parsed, reported success, and surveyed anyway.

Underneath

The console was re-initialising the Wi-Fi driver on the REPL task, racing the
UI task's own lens switch, in violation of the rule its own header states. The
LVGL widget tree was sitting in the only memory the display flush and the Wi-Fi
driver can do DMA out of. And CI had been red since v3.4.0 without anybody
seeing a build fail: the build job is gated behind the audit job, and the audit
had been failing on a missing no-panel stub, so the firmware build step had not
actually run for weeks.

8898 host checks, 0 failures - and 0 under UBSan and ASan. Fence and lens ELF
audits intact. All twenty-one faces photographed off the glass, with nothing
crossing the safe radius.

Flash it

From a browser - Chrome or Edge, no tools to install:
https://at0m-b0mb.github.io/Pharos-ESP32S3/

Or with esptool, using the single merged image:

pip install esptool
esptool.py --chip esp32s3 write_flash 0x0 pharos-v3.5.0-esp32s3.bin

Upgrading from a build older than v1.4.0? Erase first - the partition
layout and PSRAM settings changed:
esptool.py --chip esp32s3 erase_flash

What is attached

file what it is
pharos-v3.5.0-esp32s3.bin the whole thing, flashes at 0x0
pharos-esp32s3.bin same image at a version-less name, for scripts
pharos-v3.5.0-esp32s3.zip the above plus component images and FLASH.md
SHA256SUMS.txt checksums for all of it

Audited in CI against this exact ELF

  • Transmit fence - no reachable path to a transmit primitive. This
    firmware receives only; the fence is enforced at link time, not by
    convention.
  • Lens linkage - every lens survived the link. A lens registers from
    a constructor, so the linker will happily discard one and still
    produce a clean, empty-feeling firmware.
  • Display path - the panel is driven via Waveshare's BSP.

A serial console runs at 115200 - type help. See
docs/HARDWARE.md.

Pharos v3.4.0

Choose a tag to compare

@github-actions github-actions released this 06 Sep 20:25

Eight commits since v3.3.0. Where v3.3.0 added new engines, this release
audits the ones already shipping - each lens read against the working
implementations and published research for the same attack, then fixed where
ours came off worse.

The pattern that kept recurring: a detector was making an argument it was not
entitled to make, or was blind to a whole variant of the attack it was named
after.

Probe - a link across a long gap is a coincidence. The de-randomisation
test matched an IE fingerprint plus a sequence counter within 64, with no bound
on WHEN. A 12-bit counter has 4096 values, so an unrelated device lands in that
window 1.6% of the time by chance - a steady trickle of false links across a
long session. And a device that last probed ten minutes ago has counted on by
thousands of steps, so a near-match then is coincidence rather than
continuation. A false link asserts that two people are one person.

Mirage - a hand-built beacon is underdressed. Volume, ephemerality and
synthetic addressing all need TIME, and ephemerality is an absence claim
besides. A real access point's beacon carries a dozen elements or more; a
flooding tool writes the three that make a phone list the name. That shows in a
single frame.

Census - transition mode graded an A while its handshake stayed catchable.
WPA3-Personal transition mode advertises SAE and PSK, so it read as WPA3 and
graded 90/A-. A client can still be pushed down the PSK path and its handshake
captured - which is the attack pharos_harvest grades from the other end. Capped
at B, the same argument the MFP ceiling already makes.

Sentinel - evidence about an impersonator may only ever add. A returning
BSSID whose vendor prefix had changed was scored as less suspicious, on the
grounds that a different manufacturer is unlikely. But the OUI is three bytes an
attacker chooses, so cloning the vendor made the alarm quieter. Forgeable
evidence may raise suspicion and may never lower it.

Twin - a beacon built unlike its siblings. A roaming group is one
deployment configured once, so its beacons are structurally identical. A member
whose frame is built differently is not the same hardware - positive evidence
about a frame that arrived, where posture and vendor are heuristics an attacker
controls.

Locate - the needle kept pointing at a memory. If the target stopped
transmitting, the averages held their last values and the trend reported them
forever. last_us was recorded and never read. This is the lens you follow
WITH YOUR FEET: a frozen HOTTER is worse than no reading.

Vigil - a full tag table was a first-come lottery. When the table filled it
set full and returned, so the first thirty-two addresses kept their slots for
the session - and a tag that started following you after that could never be
recorded. Eviction is now ranked by evidence, and silence is only the
tie-break: BLE-Doubt measured paired AirTags going quiet for up to an hour,
and that persisted after separation from the owner.

Harvest - one radio shopping is not a crowd of clients. The modern PMKID
attack is clientless, so it produced no forced cycles and was invisible to the
breadth family. touch_and_go counts pairs and cannot tell one radio touching
six networks from six phones touching one each.

8103 host checks, 0 failures. Every fix carries the negative it must refuse to
fire on.

Flash it

From a browser - Chrome or Edge, no tools to install:
https://at0m-b0mb.github.io/Pharos-ESP32S3/

Or with esptool, using the single merged image:

pip install esptool
esptool.py --chip esp32s3 write_flash 0x0 pharos-v3.4.0-esp32s3.bin

Upgrading from a build older than v1.4.0? Erase first - the partition
layout and PSRAM settings changed:
esptool.py --chip esp32s3 erase_flash

What is attached

file what it is
pharos-v3.4.0-esp32s3.bin the whole thing, flashes at 0x0
pharos-esp32s3.bin same image at a version-less name, for scripts
pharos-v3.4.0-esp32s3.zip the above plus component images and FLASH.md
SHA256SUMS.txt checksums for all of it

Audited in CI against this exact ELF

  • Transmit fence - no reachable path to a transmit primitive. This
    firmware receives only; the fence is enforced at link time, not by
    convention.
  • Lens linkage - every lens survived the link. A lens registers from
    a constructor, so the linker will happily discard one and still
    produce a clean, empty-feeling firmware.
  • Display path - the panel is driven via Waveshare's BSP.

A serial console runs at 115200 - type help. See
docs/HARDWARE.md.

Pharos v3.3.0

Choose a tag to compare

@github-actions github-actions released this 06 Sep 03:31

Fourteen commits since v3.2.0. The thread running through the new engines is
one idea: an attacker writes the fields, but they do not write the physics.
Every field in a frame can be set to whatever the sender likes. An oscillator's
drift rate, a transmitter's signal level and a corrupted checksum cannot be.

New engines

  • Attribution (pharos_attrib) - a deauthentication frame's source address
    is forged by definition, so it is ignored. The RSSI is not: the radio that
    sent it was at some real distance, radiating real power. Forged frames are
    matched against every device in the room by signal level, which can name the
    radio behind a spoofed address.
  • Clock skew (pharos_skew) - every access point counts microseconds on
    its own crystal, and no two crystals run at the same rate. If a BSSID's rate
    CHANGES, the radio behind that name changed. Uses the minimum offset per
    bucket, not the mean, because reception delay is a one-sided tax and
    averaging it measures how busy the CPU was rather than the oscillator.
    States plainly what it cannot do: a determined attacker can synchronise TSF,
    so a match is never reported as innocence.

Detection improvements, each measured against the field

  • Karma gains IMPOSTOR. PiKarma, the best-known detector of this attack,
    flags any radio answering for more than one SSID - which accuses every
    corporate AP carrying guest, staff and IoT. Pharos already scored those
    zero. But its three families all rested on an ABSENCE, the weakest claim a
    hopping receiver can make. Answering for a name a DIFFERENT radio actually
    beacons is a contradiction between two things heard, so it raises its own
    ceiling to 88 past the hopping bound.
  • Vigil gains Google's Find My Device network - Pebblebee, Moto Tag, Eufy,
    Chipolo Point - roughly half the trackers a person meets, previously
    invisible. FMDN shares service UUID 0xFEAA with ordinary Eddystone beacons,
    so shop and museum beacons are told apart by frame type and explicitly
    BARRED from the tag table. Everything in that table is a candidate for
    "travelling with you", and accusing a supermarket shelf is the worst
    sentence this lens has available.
  • Squall gains BROKEN. Its other families infer denial from what is
    missing; a failed checksum is direct evidence of interference. A congested
    channel is loud and its frames decode, a jammed one is loud and its frames
    shatter.
  • Harvest now needs TWO unanswered PMKIDs, and only from clients it has
    actually heard transmit. One is routine fast roaming, and message 2 comes
    from the client - which a one-antenna receiver often cannot hear at all.
    This was reported from hardware as SUSPECTED 46/96 on an idle network.
    "not offered" is now a stated result rather than an ambiguous zero.

Fixes found on the glass

  • Camping revoked what the lens asked to hear. The camp plan carried
    want_mgmt alone, so squall camp stopped delivering the data frames its
    retry family is computed from. 0% retries camped, 3% once fixed.
  • The START button could not start anything - drawn at +102..+158, while
    the zone that launches a lens stopped at +95. The two never overlapped.
  • Press-and-hold did not exist. Only LV_EVENT_CLICKED was registered, so
    there was no way off a page with a finger - and the guide taught the gesture
    anyway.
  • Every sensor's detail rows lost their first characters, because labels
    were centre-aligned by bounding box and grew symmetrically off the glass.
  • Whisper's verdict flickered ~450 times in 30 seconds; a confirm counter plus
    a band kerb brings it to 4.
  • Settings looked like every other list. Controls now carry a stripe, a
    chevron and an accent value; readings keep their tone.
  • The first-run guide is reachable again from Settings.

Honesty notes

Clock skew and Squall's BROKEN family are wired and host-tested but not
demonstrated on hardware
: neither an evil twin nor a jammer was available to
make them fire. They are defensible, not proven.

Flash it

From a browser - Chrome or Edge, no tools to install:
https://at0m-b0mb.github.io/Pharos-ESP32S3/

Or with esptool, using the single merged image:

pip install esptool
esptool.py --chip esp32s3 write_flash 0x0 pharos-v3.3.0-esp32s3.bin

Upgrading from a build older than v1.4.0? Erase first - the partition
layout and PSRAM settings changed:
esptool.py --chip esp32s3 erase_flash

What is attached

file what it is
pharos-v3.3.0-esp32s3.bin the whole thing, flashes at 0x0
pharos-esp32s3.bin same image at a version-less name, for scripts
pharos-v3.3.0-esp32s3.zip the above plus component images and FLASH.md
SHA256SUMS.txt checksums for all of it

Audited in CI against this exact ELF

  • Transmit fence - no reachable path to a transmit primitive. This
    firmware receives only; the fence is enforced at link time, not by
    convention.
  • Lens linkage - every lens survived the link. A lens registers from
    a constructor, so the linker will happily discard one and still
    produce a clean, empty-feeling firmware.
  • Display path - the panel is driven via Waveshare's BSP.

A serial console runs at 115200 - type help. See
docs/HARDWARE.md.

Pharos v3.2.0

Choose a tag to compare

@github-actions github-actions released this 25 Aug 17:20

The gauge has a channel now

A score of 12 used to draw a lone green stub at the bottom-left of the face,
floating in black, and it read as a rendering fault rather than as a small
number. The track behind it was 6 px of near-black under a 14 px bright arc —
so there was nothing for the reading to be a fraction of.

The track is the same weight as the score now, and lifted out of the shadows.
QUIET 12 / 96 finally looks like twelve out of ninety-six.

Nothing sits on top of the gauge any more

PR_SAFE_R is where the glass stops. On a face with a score arc, the arc
stops sooner — and everything was being fitted against the glass. The four
evidence chips were laid out to ±190 px while the arc crosses their row at
±178, so the outer two were drawn on top of it.

There is a second working radius now, PS_INNER_R, derived from the ring's
own geometry rather than typed in again, and every horizontal budget on a gauge
page measures against it.

The words that mattered were the ones getting cut

Fitting to the gauge made the why line narrower, which exposed something
worse: its vocabulary never fitted in the first place. pw_forgery_name() was
returning 25–31 characters into a 24-character band, and the cut landed on
exactly the word carrying the meaning —

sequence counter went backwards → sequence counter went.

Truncation is the right behaviour for an unbounded string and the wrong
outcome for a fixed vocabulary somebody chose. So the vocabulary was
shortened — counter went backwards, signal level is wrong,
unprotected on MFP net, source never beaconed — and a host test now walks
the whole table.

Polish

  • A smoother aura. Seven layers instead of five; five still stepped
    visibly across a 300 px disc.
  • The ribbon flows. Older seconds are dimmer, so the sixteen-second
    timeline has a direction and the bright end is now.
  • Lit chips get a hairline edge as well as a fill — fill alone is
    ambiguous against a coloured aura.
  • The boot screen earns its second and a half: the ring sweeps closed once,
    green for a clean fence, red if it could not be proven.
  • Each tool gets its own rim on the browse card, so twenty-one tools have
    an identity you register before reading the name.
  • The receive-only pip breathes. "Is it frozen?" was a real question
    asked of the old face, because a quiet room and a crashed device look
    identical. Ten pixels, four seconds, and it never touches the reading.

7883 checks. Fence, lens, display and sources audits pass. missed=1 across
500 paints on hardware — a single miss during the splash sweep, not a cost.

Flash it

From a browser - Chrome or Edge, no tools to install:
https://at0m-b0mb.github.io/Pharos-ESP32S3/

Or with esptool, using the single merged image:

pip install esptool
esptool.py --chip esp32s3 write_flash 0x0 pharos-v3.2.0-esp32s3.bin

Upgrading from a build older than v1.4.0? Erase first - the partition
layout and PSRAM settings changed:
esptool.py --chip esp32s3 erase_flash

What is attached

file what it is
pharos-v3.2.0-esp32s3.bin the whole thing, flashes at 0x0
pharos-esp32s3.bin same image at a version-less name, for scripts
pharos-v3.2.0-esp32s3.zip the above plus component images and FLASH.md
SHA256SUMS.txt checksums for all of it

Audited in CI against this exact ELF

  • Transmit fence - no reachable path to a transmit primitive. This
    firmware receives only; the fence is enforced at link time, not by
    convention.
  • Lens linkage - every lens survived the link. A lens registers from
    a constructor, so the linker will happily discard one and still
    produce a clean, empty-feeling firmware.
  • Display path - the panel is driven via Waveshare's BSP.

A serial console runs at 115200 - type help. See
docs/HARDWARE.md.

Pharos v3.1.1

Choose a tag to compare

@github-actions github-actions released this 25 Aug 16:49

A row's value could still become two lines

Reported from a photograph of the PROBE page: the grade unclassified
rendered as unclassifi with a lone e floating outside its card.

Two faults compounding, and both are fixed.

The label wrapped. LV_LABEL_LONG_DOT is documented as "break the text and
write dots in the last line" — and break means wrap. With the height left
to grow, an over-long value simply became two lines and dotted the second. Both
row columns are pinned to a single line height now, which leaves LONG_DOT no
room to wrap into, so it truncates horizontally as intended. It also restores
the rule that a cell's height can never change, and therefore can never move
the invalidated region.

Three strings never fitted in the first place. struct pharos_lens_row
carries right[12] — eleven printable characters — and pp_place_name() was
returning shop or cafe (12), carrier hotspot (15) and unclassified (12),
so snprintf silently cut them on the way into the row before the UI ever
saw them. They are now shop/cafe, carrier and unknown, and a host test
walks the whole table against the contract — because the next place name
somebody adds will be long too, and nothing else would have noticed.

Which column loses, when one has to

The narrowest card is the bottom one, where the chord has shrunk toward the
rim, and it has about 310 px for both columns. The value takes 112 of it
and is never cut: it is the judgement, and a grade that reads A when it meant
A+ is a wrong answer, not a cosmetic loss. The label is the column that
degrades — 19 characters, which covers every row the firmware currently
produces, with an honest ellipsis beyond that.

That trade is now an assertion rather than an accident.

7875 checks.

Flash it

From a browser - Chrome or Edge, no tools to install:
https://at0m-b0mb.github.io/Pharos-ESP32S3/

Or with esptool, using the single merged image:

pip install esptool
esptool.py --chip esp32s3 write_flash 0x0 pharos-v3.1.1-esp32s3.bin

Upgrading from a build older than v1.4.0? Erase first - the partition
layout and PSRAM settings changed:
esptool.py --chip esp32s3 erase_flash

What is attached

file what it is
pharos-v3.1.1-esp32s3.bin the whole thing, flashes at 0x0
pharos-esp32s3.bin same image at a version-less name, for scripts
pharos-v3.1.1-esp32s3.zip the above plus component images and FLASH.md
SHA256SUMS.txt checksums for all of it

Audited in CI against this exact ELF

  • Transmit fence - no reachable path to a transmit primitive. This
    firmware receives only; the fence is enforced at link time, not by
    convention.
  • Lens linkage - every lens survived the link. A lens registers from
    a constructor, so the linker will happily discard one and still
    produce a clean, empty-feeling firmware.
  • Display path - the panel is driven via Waveshare's BSP.

A serial console runs at 115200 - type help. See
docs/HARDWARE.md.

Pharos v3.1.0

Choose a tag to compare

@github-actions github-actions released this 25 Aug 16:28

The guide: the device teaches itself

Twenty-one tools, four verdict colours, and a touch surface with no visible
buttons. The controls were explained in the console banner — which is no help
at all to somebody holding the device and looking at the glass.

So the glass explains itself now. Nine steps on first boot, remembered in NVS,
replayable forever with the new guide command, and left in a single press
by anybody who already knows the device.

Each step shows the gesture rather than describing it: the outline of a
zone appears and a fingertip pulses inside it, travelling between the two side
zones for the step that has two. You copy what you just watched instead of
skimming a sentence about it.

The order is deliberate — controls first, because you cannot explore without
them; then the colour key, because it is the key to every other screen; then
the two pieces of vocabulary that are genuinely unusual: the ring of dots and
the evidence chips.

Every detail row was losing its first word

Reported from a photograph of the Rival page, and it affected every lens:

shown on hardware should have read
p models / advs popup models / advs
resses used addresses used
ed senders faked senders

The row labels were children of the page, positioned with
lv_obj_align(LV_ALIGN_CENTER, -w/2 + 18, dy). That aligns a label's bounding
box centre at that x — and an auto-sized label grows symmetrically about
it, so half of any long string extended further left, off its card and off the
glass. LV_TEXT_ALIGN_LEFT does not help: it aligns lines within the box and
says nothing about where the box sits.

Rows are children of their card now, with explicit widths, so text is
clipped by the card and cannot reach the panel edge however long it gets.
LONG_DOT ellipsises the few labels that genuinely do not fit — a legible
failure instead of a silent amputation.

The column budget is taken from the narrowest card, the bottom one, where
the chord has shrunk toward the rim. A budget measured on the middle rows
leaves the last row two characters short; a host test asserts this and found
exactly that.

Layout is still arithmetic

The first draft of the guide put its body text at +132/+162 and its hint at
+198 — fine offsets on a rectangle, hopeless on a circle, where the chord at
+198 holds about thirteen characters. tools/render caught ten escaped
primitives
before any of it reached the device. The guide's bands are named
tokens now, with host tests behind them.

7851 checks. Display, sources, fence and lens audits pass.

Flash it

From a browser - Chrome or Edge, no tools to install:
https://at0m-b0mb.github.io/Pharos-ESP32S3/

Or with esptool, using the single merged image:

pip install esptool
esptool.py --chip esp32s3 write_flash 0x0 pharos-v3.1.0-esp32s3.bin

Upgrading from a build older than v1.4.0? Erase first - the partition
layout and PSRAM settings changed:
esptool.py --chip esp32s3 erase_flash

What is attached

file what it is
pharos-v3.1.0-esp32s3.bin the whole thing, flashes at 0x0
pharos-esp32s3.bin same image at a version-less name, for scripts
pharos-v3.1.0-esp32s3.zip the above plus component images and FLASH.md
SHA256SUMS.txt checksums for all of it

Audited in CI against this exact ELF

  • Transmit fence - no reachable path to a transmit primitive. This
    firmware receives only; the fence is enforced at link time, not by
    convention.
  • Lens linkage - every lens survived the link. A lens registers from
    a constructor, so the linker will happily discard one and still
    produce a clean, empty-feeling firmware.
  • Display path - the panel is driven via Waveshare's BSP.

A serial console runs at 115200 - type help. See
docs/HARDWARE.md.

Pharos v3.0.0

Choose a tag to compare

@github-actions github-actions released this 25 Aug 15:59

LUMEN: the face, rewritten from scratch

The old face was an aircraft-instrument pastiche — a 24-tick bezel, hairline
arcs, text down to 12 px. It photographed well and read badly in a hand, which
is the only place it is ever used. It is gone.

Nothing on the glass is under 14 px now. At 466 px across a 1.75 inch
circle this panel is 266 ppi, so the old 12 px body text was about 1.1 mm of
cap height — legible in a screenshot, mush at arm's length. Half the supporting
text was set in it. Montserrat 28/32/36 were added to the build because the old
scale jumped 26 → 48 with nothing usable in between, which forced every verdict
word down to 26 while numbers sat at 48.

Every screen now reads in one order: colour → word → number → evidence →
action. A soft aura behind the centre carries the verdict before you have read
a single glyph, so a glance from across a room gets the answer and can stop.
The number is now set smaller than the word it explains — the old face had
that exactly backwards, which is why it read as a number generator rather than
as a judgement.

The evidence chips are named. RATE SHAPE FORGE AFTER, not four
anonymous dots. "Why does it think so" is answerable from the glass instead of
from the manual.

The home ring lost its labels, and that is the fix

Fourteen names never fitted a 466 px circle at any radius; twenty-one lenses
made it worse. Names were being drawn through the headline, behind a rule about
which ones were "worth naming". The names were never the point: you answer is
anything wrong
by counting the dots that are not green. Which watch it is
matters only once you have decided to look, and then it is one name in the
middle — always legible, because there is only ever one of it.

Removing the labels removed the clipping bug, the capacity rule and the
staggered-radius machinery in one go. The dots now sit along the gauge's own
270°, leaving the bottom notch for the clock.

Measured on the device

before after
frames painted=257 missed=20 (~7% dropped) painted=983 missed=0
internal DMA RAM free 25 KB 59 KB
esp_lv_adapter_lock failures repeated, every boot none

Twenty-four bezel ticks that meant nothing, and sixteen curved lv_line bars
whose bounding boxes each spanned most of the panel, were being redrawn
underneath everything that did mean something. LUMEN draws nothing that does
not carry information.

Two real bugs this exposed

Locate painted the glass red as you walked toward your target. Its score is
closeness, not threat, but it never set has_alert, so the threat palette
was applied to it — "found it" rendered as "danger". Harmless behind a thin
arc; not behind an aura.

Advice was being cut to its first sentence, throwing away the half that
says what to do: "Broad, spoofed deauth." survived, "Preserve the log." did
not. Advice now runs to two fixed lines.

The parts that did not change

The four verdict colours are frozen and asserted byte-for-byte in
test_style.c. A red on Census means what a red on Watch means; that contract
outranks any visual refresh. The transmit fence, the confidence ceilings and
every detection engine are untouched.

Layout is arithmetic, not opinion: pharos_style.c is free of LVGL and
ESP-IDF, so the type scale, the chip row and the card stack are all proved on a
laptop — 514 new checks, 7831 total. tools/render now draws the LUMEN
face from the same tokens and bounds-checks every primitive against the safe
radius; it caught five escapes on the first pass, four of them real engine
advice lines that would have run off the edge of the glass.

Fence and lens audits pass. Verified on hardware.

Flash it

From a browser - Chrome or Edge, no tools to install:
https://at0m-b0mb.github.io/Pharos-ESP32S3/

Or with esptool, using the single merged image:

pip install esptool
esptool.py --chip esp32s3 write_flash 0x0 pharos-v3.0.0-esp32s3.bin

Upgrading from a build older than v1.4.0? Erase first - the partition
layout and PSRAM settings changed:
esptool.py --chip esp32s3 erase_flash

What is attached

file what it is
pharos-v3.0.0-esp32s3.bin the whole thing, flashes at 0x0
pharos-esp32s3.bin same image at a version-less name, for scripts
pharos-v3.0.0-esp32s3.zip the above plus component images and FLASH.md
SHA256SUMS.txt checksums for all of it

Audited in CI against this exact ELF

  • Transmit fence - no reachable path to a transmit primitive. This
    firmware receives only; the fence is enforced at link time, not by
    convention.
  • Lens linkage - every lens survived the link. A lens registers from
    a constructor, so the linker will happily discard one and still
    produce a clean, empty-feeling firmware.
  • Display path - the panel is driven via Waveshare's BSP.

A serial console runs at 115200 - type help. See
docs/HARDWARE.md.