seismo-relay/docs/micromate_protocol_reference.md

# Micromate Protocol Reference — Thor / Micromate Series IV, live wire protocol

Sibling to [instantel_protocol_reference.md](instantel_protocol_reference.md)
(Series III, "the Rosetta Stone") and
[idf_protocol_reference.md](idf_protocol_reference.md) (Series IV *file*
format).  This document covers the Series IV **live device protocol** — what
the unit says over the wire, as opposed to what it writes into a `.IDFW`.

**Status (2026-09-23): opening session.** Everything below was established in
a single bench session against one unit.  Treat it as a strong start, not a
settled spec — in particular, everything here comes from **one unit, over USB,
with no events stored**.

---

## The headline

**A Micromate running the *Blastware* firmware answers Series III command
frames.**

An unmodified Series III `POLL` (`SUB 0x5B`), built by
`minimateplus.framing.build_bw_frame` with no changes at all, produced a
complete two-step probe/data cycle.  Ten Series III read commands were then
tried and **all ten answered**, every one obeying the Series III response-SUB
rule.

⚠ **That qualifier is load-bearing, and it was discovered after the fact.**
Instantel ships the Micromate in two firmware lines:

| firmware | Instantel's own description |
|---|---|
| `11.0CB` | "Utilize with **Blastware**" |
| `11.0BD` | "Utilize with **THOR, Vision, Vision II**" |

The bench unit reports **`11.0CB`** — the Blastware build.  So the clean
Series III behaviour above is very likely *because the unit is in Blastware
mode*, not because the Micromate natively speaks Series III.  **Nothing here
should be assumed to hold on a `11.0BD` unit until tested.**

This reframes the project.  The question is no longer only "what is the
Series IV protocol" but **"which firmware line do we target, and does one of
them let the existing Series III stack drive the whole fleet?"**

---

## Firmware — the variable nobody knew was a variable

This explains a production problem TMI has lived with: two ACH servers, two
machines, and units that will not cross over.  It is not a misconfiguration.
**Instantel ships different firmware for different host software**, and the
wire protocol differs with it.

### Fleet audit (physical, 2026-09-22/23)

| unit | firmware | line | location |
|---|---|---|---|
| UM11719 | `11.0CB` | Blastware | Ped Bridge Loc 2 |
| UM6047 | `11.0CB` | Blastware | Brookville Loc 9 |
| UM12947 | `11.0CB` | Blastware | **bench** |
| UM14133 | `11.0CB` | Blastware | Pitt-Music Bldg Loc 1 |
| UM11402 | `11.0BD` | Thor | Ped Bridge Loc 1 |
| UM20147 | `11.0BD` | Thor | **bench** |
| UM13981 | `11.0AK` | pre-split | RKM Loc 1 |
| UM20146 | `11.0AK` | pre-split | Karns Loc 2 |
| UM12420 | `10.90GC` | pre-split | RKM Loc 2 |

**4 Blastware / 2 Thor / 3 pre-split.**  The Blastware line is already the
plurality, which makes "standardise on Blastware" less disruptive than it
first appeared.

A store-derived audit (firmware is recorded in every `.sfm.json` sidecar as
`extensions.idf_report.version`) agreed with the physical audit on **7 of 9**.
The two that differed — UM6047 and UM14133 — are the most recently deployed
units, reflashed after their last stored event.  Useful technique: the fleet's
firmware history is reconstructable from the store without touching a unit,
but it lags reality by one deployment.

Firmware is **not stable per-unit over time** — five of nine have been
reflashed at least once.  Any fleet-wide claim needs a fresh audit.

### ⚠ Retraction: firmware line does NOT determine Thor compatibility

An earlier draft of this document suggested that UM12947's trouble with Thor
was explained by its being on the Blastware build.  **That is not supported.**

Ped Bridge runs UM11402 (`11.0BD`) and UM11719 (`11.0CB`) side by side, both
deployed 2026-04-20, and **both call Thor successfully** — UM11719 has 331
Thor-collected events in the store while on `11.0CB`, through 2026-08-23.

So a Blastware-line unit does feed Thor.  Whatever the CB/BD split changes, it
is not "which host software can collect from it", and UM12947's specific
problem remains unexplained.

**What is actually established:** a `11.0CB` unit answers Series III command
frames.  Whether a `11.0BD` unit does is **untested** — and UM20147 (`11.0BD`)
is on the bench, which makes that a direct A/B away.

The version is readable over the wire from `SUB 0x01` as two separate ASCII
runs — `"0CB"` then `"11"` — with no single concatenated string.

### The strategic fork

- **Option A — standardise the fleet on the Blastware line.**  Every unit,
  MiniMate and Micromate alike, then speaks Series III, and the existing
  `minimateplus/` stack drives all of it.  One protocol, one call-home
  receiver.  Dramatically cheaper *if* it holds up.
- **Option B — reverse-engineer the Thor line (`11.0BD`) and support both.**

Option A is the shortcut, but it is unproven and carries real unknowns, all
of which are cheap to answer on the bench and expensive to discover later:

1. **What file format does a `11.0CB` unit produce?**  If it emits Blastware
   binaries rather than `.IDFW`/`.IDFH`, the (now exact) Series III decoder
   applies and the IDF codec becomes a legacy path.  Not a loss — we have
   both — but it changes what the ingest pipeline sees.
2. **Does Thor still work with a `11.0CB` unit?**  If not, flashing a
   production unit breaks data collection until the replacement path exists.
3. **Are any Micromate-specific capabilities lost** on the Blastware line?
4. **Is the flash reversible in the field**, and what does it cost in downtime?

## How this was obtained

No Thor, no modem, no Windows machine.  The Micromate exposes its protocol on
a **USB CDC-ACM virtual serial port**:

```
ID 2504:0300 Instantel Inc. MICROMATE COM PORT
driver: cdc_acm          ATTRS{serial}=="V1.00"
→ /dev/ttyACM0
```

Plain CDC ACM, so no vendor driver and no proprietary USB layer — any host
that can open a serial port can talk to the unit.

**Baud is irrelevant over USB.**  Identical byte-for-byte responses at 38400
and 115200; CDC-ACM ignores the line rate.  TMI provisions Micromate *modem*
links at **115200** (Series III uses 38400) — that matters for the cellular
path, not for USB.

**The device never speaks first.**  20 s of passive listening on an idle open
port produced zero bytes.  It is strictly request/response.

---

## Physical + framing layer

### Requests — Series III format, unmodified

Every frame in this session was produced by `build_bw_frame(sub, offset)` with
no Series IV changes, and the device accepted all of them:

```
[ACK 0x41] [STX 0x02] [10 10] [flags 00] [SUB] [00] [00] [offset] [params×10] [chk] [ETX 0x03]
```

⚠ Only the doubled `BW_CMD` (`10 10`) form has been exercised.  Whether other
literal `0x10` bytes inside params require stuffing is **untested** — none of
the probes sent carried one.

### Responses — Series III *minus the DLE prefix*

```
Series III:   [DLE 0x10] [STX 0x02] … [chk] [ETX 0x03]
Micromate:              [STX 0x02] … [chk] [ETX 0x03]      ← no leading DLE
```

This single byte matters operationally: Blastware's parser locates frames by
scanning for `DLE+STX`, so it will **never find a frame boundary** in Micromate
traffic no matter what else is correct.  That is a structural reason a
Micromate cannot call into a Blastware ACH server, independent of any baud
mismatch.

### Response payload header

```
[0] CMD    0x00          same as Series III
[1] flags  0xC5          ← Series III uses 0x10.  Constant across all 10 SUBs.
[2] SUB    0xFF − request_SUB
[3] PAGE_HI
[4] PAGE_LO
[5+] data
```

### Checksum — the DLE-aware variant

```python
chk = sum(b for b in payload if b != 0x10) & 0xFF
```

Confirmed on every frame captured.  The `POLL` probe response contains no
`0x10` and so cannot distinguish plain SUM8 from the DLE-aware form; the
`POLL` **data** response contains a `0x10` at payload offset 42, and only the
DLE-aware rule matches there.  This is the same checksum Series III uses for
its `5A` bulk-stream and write frames — not the plain SUM8 of ordinary
Series III reads.

### The probe response carries the data length

Series III hardcodes `DATA_LENGTHS` per SUB.  On the Micromate the **probe
response tells you**, as a **uint16 BE at `payload[8:10]`**:

⚠ **Corrected 2026-09-23.**  An earlier draft read this as a single byte at
`payload[9]`.  That is right only while the high byte is zero, and it is
catastrophically wrong for `SUB 0x1A`, whose real length is `0x082C` = **2092**
— read as a byte it gives **44**, a 47× under-read.  Always read the pair.

| SUB | command | `payload[9]` | Series III constant |
|---|---|---|---|
| `0x15` | serial number | `0x0A` | `0x0A` ✓ |
| `0x01` | device info | `0x98` | `0x98` ✓ |
| `0x1C` | monitor status | `0x2C` | `0x2C` ✓ |
| `0x06` | storage range | `0x24` | `0x24` ✓ |
| `0x2C` | call-home config | `0x7E` | `0x7C` ✗ **differs by 2** |
| `0x08` | event index | `0x5A` | — |
| `0x1E` | event header | `0x08` | — |
| `0x1A` | compliance config | `0x082C` (2092) | — |
| `0x0A` | waveform header | `0x00` | — (no event context) |
| `0xFE` | full config | `0x00` | — (see note) |

Four of four known Series III lengths match exactly.  **Read the length from
the probe rather than hardcoding it** — it is free, and it already caught the
call-home divergence.

---

## Confirmed commands (read-only)

All ten below answered with a correct `0xFF − SUB` response.  Nothing that
writes, erases, or changes monitoring state has been sent to a unit.

| SUB | RSP | Command | Data proven |
|---|---|---|---|
| `0x5B` | `0xA4` | POLL / handshake | yes — ID block |
| `0x15` | `0xEA` | Serial number | yes — `UM12947` |
| `0x01` | `0xFE` | Device info | yes — 152 B |
| `0x2C` | `0xD3` | Call-home config | yes — 126 B |
| `0x1C` | `0xE3` | Monitor status | yes — 44 B |
| `0x06` | `0xF9` | Event storage range | yes — 36 B |
| `0x08` | `0xF7` | Event index | probe only |
| `0x1E` | `0xE1` | Event header / first key | probe only |
| `0x0A` | `0xF5` | Waveform header | probe only |
| `0x1A` | `0xE5` | Compliance config | probe only |

### Decoded so far

**`SUB 0x15` — serial.** ASCII, null-terminated: `UM12947`.

**`SUB 0x5B` / `0x01` — identification strings.**
`Instantel\0` and `MM/ISEE/S/IO` (MicroMate / ISEE standard).  `0x01` also
carries eight consecutive `3f 80 00 00` float32 values (= `1.0f`) — almost
certainly per-channel calibration/scale factors, by analogy with Series III's
`geo_hardware_constant`.  **Unverified.**

**`SUB 0x1C` — monitor status.  Series III field offsets apply unchanged:**

| field | offset | read |
|---|---|---|
| battery × 100 | `payload[-10:-8]` uint16 BE | `0x017D` → **3.81 V** |
| memory total | `payload[-8:-4]` uint32 BE | 15,000,000 |
| memory free | `payload[-4:]` uint32 BE | 15,000,000 (empty) |
| date | `payload[18:22]` | day 23, month 9, year `0x07EA` = 2026 |

The battery reading independently corroborates: Thor's own event reports for
these units print `BatteryLevel : 3.8 volts`.

**`SUB 0x2C` — call-home config.** Contains the ASCII string `RADIO RING`.
Worth flagging: that is the *exact* string seen in the RV50 `ALEOS_SERIAL`
debug during the BE12599 incident —
`'ATQ1^MATE0^MATS0=2^M^MRADIO RING^M'`.  So this block holds the modem dial /
answer strings, and it is the most directly relevant command to the
call-home-receiver goal.  Field layout **not yet mapped**; Series III's map
(`raw[5]` enabled, `raw[6:46]` dial string) is a starting hypothesis only, and
the length already differs (`0x7E` vs `0x7C`).

**`SUB 0x06` — storage range.** All zeros on this unit, consistent with
`memory free == memory total`.  Series III reads first/last event keys from
the final 8 bytes; untestable until the unit holds events.

---

## The event chain — walked end to end (2026-09-23, 5 events)

With 5 events on the bench unit (4 waveform + 1 histogram), the Series III
browse walk works unmodified:

```
1E (all-zero params)        -> first key + size
  0A (key)                  -> partial record, histogram only
  0C (key)                  -> 210-byte waveform record
1F (all-zero params/browse) -> next key + size
  ... repeat ...
1F                          -> all-zero key = NULL SENTINEL, chain ends
```

The sentinel terminated correctly after exactly 5 events.

### Event keys are sequential, not addresses

```
055d4a81  055d4a82  055d4a83  055d4a84  055d4a85
```

**This is a real divergence.** Series III keys are flash-buffer *addresses*
(`01110000`, `011121F2`, …) that advance by the event's byte length, which is
why its 5A chunk walk is address-arithmetic. Micromate keys are a plain
incrementing counter. Any port of the Series III download walk must not
assume key arithmetic means anything.

### The 4 bytes after the key are the event's size

`1E`/`1F` return `[key 4B][size 4B]`. Series III uses that slot as an offset
to the next key; here it is a byte count:

| key | size | kind |
|---|---|---|
| `055d4a81` | 4,076 | histogram |
| `055d4a82` | 11,032 | waveform |
| `055d4a83` | 11,502 | waveform |
| `055d4a84` | 13,424 | waveform |
| `055d4a85` | 8,746 | waveform |

Consistent with real file sizes (corpus `.IDFH` ≈ 3.7–25 KB, `.IDFW` ≈
8.6–15.8 KB), and the histogram is unmistakably the small one.  ⚠ Inferred,
not proven: the sizes sum to 48,780 while monitor status reports 57,344 bytes
used, so ~8.5 KB of overhead is unaccounted for.

### `SUB 0x0C` — waveform record, and it carries the job metadata

**Length `0xD2` = 210 bytes — identical to Series III.**  Contents confirmed
across all 5 events:

- the event key, echoed
- date + time (`17 09 07 ea` → 23 Sep 2026, then `10 21` → 16:33 — matching
  the actual bench recording time)
- title note `"Location"`
- **the project string** — `"Univ of Pitt-1st Yr Housing-Loc1 Ruskin"`
- serial `"UM12947"`
- channel labels `Tran` / `Vert` / `Long` / `Mic` — the same labels Series III
  uses, and the same label-relative float32 layout
- per-event float32 peaks: 3.5152, 1.3720, 2.3542, 3.5152, 0.4227 in/s
  across the five events (varied deliberately during recording)

**This closes the biggest open question for the call-home-receiver goal.**
The job identity strings (`project` / `client` / `operator` / `setup`) that
today arrive only via Thor's `.txt` sidecar — and which no amount of sample
decoding can reconstruct — are **available over the wire from `0x0C`**.  A
direct-to-SFM event need not arrive with blank metadata.

### `SUB 0x0A` — partial record, histogram only

`0x0A` returned `len = 0x1E` (30 B) for the histogram and `len = 0x00` for all
four waveforms.  The histogram payload carries **two timestamps** and the
ASCII string `"\r Vert: 0.300 in/s"` — structurally the Series III
**monitor-log partial record** (`0x2C` type), which likewise holds a start/stop
pair and a `"Geo: <float> in/s"` trigger string.

So on the Micromate the division of labour is: `0x0A` describes interval-style
records, `0x0C` describes triggered events.  Series III uses `0x0A`'s
*response length* (`0x46` vs `0x2C`) to tell real events from boundaries;
that discriminator does not apply here.

### DLE stuffing in responses — confirmed present

Earlier marked untested.  The `0x0C` timestamp field contains `10 10`, which
destuffs to a single `0x10` and yields a sensible clock reading.  **Responses
are DLE-stuffed**, so a parser must destuff before applying field offsets.

---

## A/B: Blastware build vs Thor build (2026-09-23)

UM12947 (`11.0CB`) and UM20147 (`11.0BD`) were each put on the bench and given
the identical read-only sweep.  **Both answer Series III command frames.**

| | UM12947 `11.0CB` | UM20147 `11.0BD` |
|---|---|---|
| POLL answers | ✓ | ✓ |
| All 10 read SUBs answer | ✓ | ✓ |
| `response_SUB = 0xFF − req` | ✓ | ✓ |
| DLE-aware checksum valid | ✓ | ✓ |
| Two-step probe/data read | ✓ | ✓ |
| ID string | `MM/ISEE/S/IO` | `MM/ISEE/S` |
| **flags byte** | **`0xC5`** | **`0x03`** |
| **`0x1C` length** | **`0x2C`** | **`0x30`** |

**The firmware line does not change the wire protocol.**  One protocol stack
can drive the whole fleet regardless of which build a unit is on.  This is the
single most consequential finding so far: "standardise the fleet on one
firmware" becomes an *optional* convenience rather than a prerequisite for
building a call-home receiver.

### The two differences that do exist

**1. The flags byte identifies the build.**  Response `payload[1]` is `0xC5`
on the Blastware line and `0x03` on the Thor line, constant across all ten
SUBs on both units.  That makes firmware line detectable from *any* response,
without reading device info.  ⚠ Two units, one each — treat as a strong
hypothesis, not a proven encoding.

Note `0x03` is ETX, so on Thor-line units it arrives DLE-escaped as `10 03`.
A parser that fails to destuff will mis-locate every field by one byte on
exactly half your fleet.

**2. `SUB 0x1C` (monitor status) is 4 bytes longer on the Thor line** —
`0x30` vs `0x2C` — with four extra trailing bytes (`0f a0 00 00`, purpose
unknown).

⚠ **This breaks relative-to-end parsing.**  Series III reads battery and
memory from the *end* of the `0x1C` block (`[-10:-8]`, `[-8:-4]`, `[-4:]`).
Those offsets are correct on `11.0CB` and wrong on `11.0BD` — applying them
blindly to UM20147 yields a battery reading of **577.92 V**.  Parse forward
from the declared length instead of backward from the end.

With the offsets shifted by 4, UM20147 reads correctly: battery **3.81 V**,
memory 15,000,000 total and free (no events stored).

## Divergences from Series III (running list)

1. **No `DLE` prefix on responses** — bare `STX`.
2. **Response flags byte is `0xC5`**, not `0x10`.
3. **Call-home config is 126 bytes**, not 124.
4. **Data lengths are discoverable** from the probe response at `payload[9]`.
4b. **Event keys are a sequential counter**, not flash addresses.
4c. **`1E`/`1F` return the event size**, where Series III returns an offset.
4d. **`0x0A` vs `0x0C` split by record type**, not by the `0x46`/`0x2C`
    length discriminator Series III uses.
4e. **Response `payload[1]` (flags) encodes the firmware line** — `0xC5`
    Blastware, `0x03` Thor — where Series III has a constant `0x10`.
5. **Modem serial rate is 115200**, not 38400 (per TMI provisioning practice;
   not independently verified here).
6. **Reads are single-step** — Thor asks for `offset = 0xFFFF` and gets the
   whole block; Series III always probes first.  (2026-09-24)
7. **Setup writes are preceded by `SUB 0xDA`**, which names the target `.MMB`
   file.  Series III has no equivalent — it has one config, not named files.
8. **The compliance block is written in ONE `0x71` frame**, not Series III's
   three chunks.
9. **`0x69` / `0x74` (waveform data write) are absent** from a setup push.
10. **Six channel blocks** (`Tran`/`Vert`/`Long`/`Mic`/`LMic`/`SMic`) on a
    48-byte stride, against Series III's four.
11. **The notes block is four fixed-width `[label:22][value:42]` entries** on a
    64-byte stride, with different labels — not Series III's `label: value`
    scan targets.
12. **The geophone scale factor is `3.10308`**, exactly half Series III's
    `6.206053`, with an ADC of 10,000 counts per volt.
13. **The monitoring flag is `0x0E`**, where Series III uses `0x10` — and it
    sits at `SUB 0x1C` `data[12]`.  A second indicator lives at `SUB 0x49`
    `data[11]` (`0x02` monitoring), which Series III has no equivalent of.
14. **Setup files are enumerated** with a `0x3F`/`0x40` first/next walk.
    Series III has one config and nothing to enumerate.
15. **`SUB 0x1C` carries the device clock** (day/month/year/h/m/s at
    `data[13:21]`).  Nothing else read so far reports the unit's own time.
18. **Deletion is per-event** (`0xA8` + a key, then `0xAA`), where Series III can
    only erase everything (`0xA3`/`0xA2`).
19. **`SUB 0x93` arms each event** before `1E`/`1F`, replacing Series III's
    `1E(token=0xFE)` — no token, no params.
20. **Event keys are a sequential counter** (`055D4A81…86`), not flash addresses.
21. **The ACH enable is `0x05`/`0x04`**, not Series III's `0x01`/`0x00` — bit 0
    is the flag, bit 2 is set in both states.
16. **There is a generic file transfer addressed by full path** — `0x94`/`0x48`
    read, `0x8D`/`0x8E` write.  Series III has nothing comparable; its config is
    reachable only through dedicated commands.
17. **The scheduler is a separate file**, `\system\schedule\schedule.dat`,
    and a schedule entry names a **setup file** to load.

---

## The firmware images are unencrypted — and they document themselves

`ref-stuff/micromate-firmware/MICROMATE(CB).BIN` and `MICROMATE(BD).BIN`,
~2.77 MB each and within 192 bytes of one another.

- **Entropy 6.08 bits/byte** — neither encrypted nor compressed.  Plain code
  and data.
- Header is a **big-endian vector table**, handlers at `0x4010_30xx`.
- **~16,700 extractable strings**, including the developers' own debug
  `printf` format strings with function names intact.

This is a legitimate interoperability reference for hardware TMI owns, and it
short-circuits work I had scoped as "only answerable from a live modem
capture".

### The call-home state machine, verbatim

```
ACH_NOT_STARTED → ACH_IDLE → ACH_INITIALIZING → ACH_CONNECTING
                → ACH_CONNECTED → ACH_TRANSFER_DATA
                → ACH_RETRY / ACH_QUITTING          (also ACH_STARTED)
```

Supporting strings:

```
ACH: Entry StartCallHome()
ACH: CallHome_task ; CheckAliveTime  CANCEL ; TimeBetweenRetries = %d
ACH: CallHome_task ;  !ExpectedCommunicationsDetected() CANCEL ; TimeBetweenRetries = %d
ACH: CallHomeCommectionCompleteProcessing() ACH=%s EAMWC=%s
ACH: %s() three attempts and it's over
ACH: %s() Send CMD_START_MONITOR
ACH: %s() Send CMD_STOP_MONITOR
ACH: Start Ignore request, Call Home is in progress
```

What this tells us without a single captured packet:

1. **Retry limit is three** — "three attempts and it's over".
2. **`ExpectedCommunicationsDetected()` gates the session.**  If the host does
   not say something the unit recognises, the call is *cancelled* and
   rescheduled after `TimeBetweenRetries`.  A homebrew receiver must satisfy
   this check or units will retry forever — which is exactly the failure mode
   seen on BE12599.
3. **The unit stops monitoring to call home and restarts afterwards**
   (`Send CMD_STOP_MONITOR` / `CMD_START_MONITOR`).  Relevant to any
   wedged-unit rescue: the monitoring state around a call is the device's own
   doing, not ours.
4. **Calls are not re-entrant** — "Call Home is in progress" is ignored.

### Internal command table

`CMD_CALLHOME`, `CMD_CALLHOME_CANCEL`, **`CMD_CALLHOME_CONNECTION_CONFIRMED`**,
`CMD_CALLHOME_CONNECTION_COMPLETE`, `CMD_CALLHOME_SET_SCHEDULE`,
`CMD_CALLHOME_CLEAR_SCHEDULE`, `CMD_STOP_CALLHOME_FILETRANSFER`,
`CMD_DUTYCYCLE_AUTOCALLHOME`, `CMD_PURGE_EVENT_FLASH`.

`CONNECTION_CONFIRMED` as a distinct state from `CONNECTION_COMPLETE` implies
a **handshake the host must complete before data flows** — the concrete shape
of `ExpectedCommunicationsDetected()`.

### Event delivery — the mechanism, probably

```
All Events Uploaded
Mark/Unmark File          Delete Marked Events        Marked Events were Deleted
MONITOR::MESG PURGE_EVENT_FLASH BEGIN / END
```

A **marking** mechanism exists, alongside a distinct "all uploaded" terminal
state.  🔶 **Inferred:** events are *marked* as transferred rather than
deleted on send, and purging is a separate explicit act.  If so, a receiver
that fails to mark would see the same events re-offered every call — the
question that gates a safe homebrew receiver.  **Not yet confirmed**; needs
either a live call-home capture or disassembly around these strings.

### A full user manual is embedded

The firmware carries its own HTML help, which documents configuration we would
otherwise have to infer:

- **Modem mode**: `Generic` (through a modem) vs `USB to PC`.
- **Modem baud**: 9600 / 19200 / 38400 / 57600 / 115200 / 230400 — "must match
  the expected rate of the PC or modem".  Confirms 115200 is a *setting*, not
  a fixed rate.
- **Modem relay + warmup** (0–300 s), auxiliary mode, warning/alarm hold.
- **Record modes**: Waveform, Waveform Manual, Histogram, Histogram-Combo;
  sample rates 1024 / 2048 / 4096.
- **A scheduler downloaded from THOR** that can start/stop monitoring, change
  record mode, trigger a call home, or run a self check on a daily/weekly
  schedule.  Pairs with `CMD_CALLHOME_SET_SCHEDULE`.

### Still worth doing

Diffing the two images should isolate exactly what the CB/BD split changes —
we know the wire protocol is not it, and the flags byte (`0xC5` vs `0x03`)
gives a concrete anchor to search for.

## `SUB 0x5A` — bulk download.  It streams the `.IDFW` file verbatim.

**The complete read path works with no Instantel software in the loop.**

### It needs no arming sequence

Series III ignores a `5A` probe unless preceded by
`1E → 0A → 1E(token 0xFE) → 0C → 1F(token 0xFE) → POLL × 3`.  The Micromate
answers a **bare `5A` request** with nothing before it.  That whole ritual is
gone.

### The offset word is a LENGTH, not a position

This is the key divergence.  Series III walks chunks by absolute flash
address, stepping `0x0200` per request.  On the Micromate the offset word
requests *how much to send*:

```
offset_word = 0x1000 + 2 × pages        pages = ceil(event_size / 512)
```

| `offset_word` | pages | destuffed file bytes |
|---|---|---|
| `0x1002` | 1 | 518 |
| `0x1004` | 2 | 1,030 |
| `0x1006` | 3 | 1,541 |
| `0x102C` | 22 | **11,033 — the whole event** |

`event_size` comes from the chain walk (the 4 bytes after the key in
`1E`/`1F`).  **One request returns the entire event**; there is no chunk loop,
no `STRT` end-offset parsing, and no `TERM` frame.  Over-requesting is safe —
`0x1030` (24 pages) returned exactly the same bytes as `0x102C`, so the device
caps at the real size.

Params are the Series III *probe* form: `[0x00][key4][6 × 0x00]`.

### The payload is the `.IDFW` file, byte for byte

```
[18-byte frame header] [ .IDFW file ] [chk] [ETX]     ← raw wire
                        ^ destuffed offset 16
```

The file begins `00 12 01 00 00 00 "Instantel\0"` — `_THOR_PREFIX` +
`_INSTANTEL_TAG` from `micromate/idf_file.py`.  The first 32 bytes are
**identical to a production `.IDFW`** pulled from the store.

⚠ Responses are DLE-stuffed.  Destuff before locating the file, or the raw
byte count overshoots (11,781 raw → 11,049 destuffed for an 11,032-byte event).

### End-to-end proof

Event `055d4a82` downloaded over USB and fed straight to `read_idf_file()`:

```
serial     UM12947
timestamp  2026-09-23 16:33:19
samples    Tran 3072   Vert 3072   Long 3072   MicL 3072
peaks      Tran 0.2433   Vert 1.3706   Long 0.2672  in/s
```

All four channels equal length, and the timestamp matches the `0x0C` record
for the same key.  **Independent cross-check:** `0x0C` reports a stored peak
of **1.3720** for this event; the decoded samples give **1.3706** — two
unrelated paths agreeing to 0.1%.

**Consequence:** no new codec work is needed.  The bytes off the wire are the
same bytes `thor-watcher` forwards today, so `/db/import/idf_file` ingests a
directly-downloaded event unchanged.  Everything the IDF decoder already does
per-sample-exact applies.

### What a full read now looks like

```
1E                     → first key + size
  0C(key)              → project/client/operator, timestamp, peaks
  5A(key, 0x1000+2×ceil(size/512))  → the whole .IDFW
1F                     → next key + size   (until null sentinel)
```

## Setups are FILES, not a config block

Series III has one compliance config you overwrite.  Series IV keeps **named
setup files on an on-device filesystem**, with a pointer to the current one.
From the firmware:

```
csetup.MMB              the current setup
factory.MMB             Factory Default Setup File
callhome.MMB            call-home config is a file too
"Current Setup File:  "         "Can Not Delete Active Setup File"
GetSelectedSetupFilePathName()  CSelectSetupFiles  CSaveSetupFile
```

Names are up to 20 characters and may contain spaces, hyphens, underscores.
The unit's help text describes selecting, renaming and deleting them, and the
event list records which setup file produced each event.

Filesystem primitives exist internally (`NS_ReadFile_internal`,
`NS_WriteFile_internal`, `NS_SeekFile_internal`).

> ⚠ **RETRACTED 2026-09-24.**  This section originally concluded "**no generic
> file-transfer command is exposed on the wire**", reasoning from the absence of
> firmware strings.  **That was wrong.**  `SUB 0x94` / `0x48` / `0x8D` / `0x8E`
> are exactly that — a read/write file transfer addressed by **full filesystem
> path** — and they were caught in the open on the first capture that touched
> the scheduler.  See *The scheduler, and a generic file transfer* below.
>
> The lesson is the usual one: absence of a firmware string is not absence of a
> command.  The dispatch is a 68K jump table and the commands carry no strings
> of their own.
>
> The *setups* half of the original claim survives: setups are **not** pushed as
> raw `.MMB` blobs.  They go through `0xDA` + the ordinary config block.

### `SUB 0x1A` reads the whole active setup — 2,092 bytes

Structurally close to Series III's ~2,126-byte compliance block, and it
carries everything a setup consists of:

- **the setup file name** — `Univ of Pitt-1st Yr. Housing-Loc1 Ruskin.MMB`
- all four title note/value pairs — `Location`, `Client`, `Company`,
  `General Notes`, with their strings
- the sensor location string (`Loc 1`)
- per-channel labels *and units*: `Tran in./s.`, `Vert in./s.`,
  `Long in./s.`, `Mic psi (L)`, `LMic psi (L)`, `SMic (A)`

Note `LMic` / `SMic` — linear and sound-level microphone variants that
Series III does not have.

This is the **read half of setup management**, and it means a setup can be
round-tripped: read the active config, modify, write it back.  ✅ **The write
half was observed on 2026-09-24** — see *The write path* below, which confirms
the round-trip: the written block is the read block, 91% byte-identical.

## The write path — observed end to end (2026-09-24)

⚠ **We have still never sent a write command to a unit.**  Thor did every write
below; seismo_lab's TCP bridge sat between Thor and the unit and recorded both
directions.  Capture: `bridges/captures/be12599-diag/9-24-26 - micromate2/`
(`raw_bw_*` = Thor, `raw_s3_*` = unit), UM12947 on the USB **PC** port, relayed
to TCP by `socat` on mint-mac.  Operation: push a setup named `TEST1.mmb`.

**All 12 device responses checksum-validate and every write is acked.**

### The sequence

```
Thor:   5B  │ 41 │ 08 │ 2E │ 1A │ DA │ 68 → 73 │ 82 → 83 │ 71 → 72
unit:   A4  │ BE │ F7 │ D1 │ E5 │ 25 │ 97   8C │ 7D   7C │ 8E   8D
        └──────── reads ────────┘   └──────────── writes ────────────┘
```

Series III, for comparison: `68→73 │ 71×3→72 │ 82→83 │ 69→74→72`.

Same write SUBs, and **every response SUB matches the Series III table
exactly** (`68`→`97`, `73`→`8C`, `82`→`7D`, `83`→`7C`, `71`→`8E`, `72`→`8D`).
Three differences:

- **`SUB 0xDA` is new** and comes first — see below.
- **`0x71` is a single write, not three chunks.**  Series III splits the
  compliance block into 1027 + 1055 + remainder; the Micromate takes all
  2,090 bytes in one frame.
- **`0x69` / `0x74` (waveform data write) do not appear at all.**

Every write ack is a 16-byte zero-data frame, same shape as Series III's.

### Three new read commands

| SUB | rsp | payload | what it carries |
|---|---|---|---|
| `0x41` | `0xBE` | 271 B | **the active setup's file name**, e.g. `TEST1.mmb` |
| `0x2E` | `0xD1` | 44 B | trigger-config block; mirrors what `0x82` writes |
| `0xDA` | `0x25` | 272 B (write) | **declares the target setup file** — see below |

### `SUB 0xDA` — name the setup file you are about to write

Data is exactly **256 bytes: the file name, null-padded, nothing else.**

```
54 45 53 54 31 2e 6d 6d 62 00 00 …      "TEST1.mmb" + 247 × 0x00
```

This is the missing link in *setups are files*: Thor names the target file, then
writes the ordinary config block into it.  The unit acks with `0x25` before any
config bytes are sent.

Note `0xDA` takes a **bare filename** null-padded to 256 bytes.  That is a
different mechanism from the path-addressed file transfer (`0x94` / `0x8D`,
which carry `\system\schedule\schedule.dat` unpadded) — setups do not go through
the file transfer, and the file transfer is not how setups are written.

### Reads are single-step — no probe

Series III sends every read twice (probe at `offset=0x00` to learn the length,
then a data step).  Thor **skips the probe** and asks for
**`offset = 0xFFFF`**, getting the whole block in one response:

```
41 02 10 10 00 1a 00 ff ff 00 …        SUB 0x1A, offset 0xFFFF
→ 2,103-byte response
```

`POLL` is the exception — it uses `offset = 0x0030` (48), its data length.

This does not contradict the probe response documented above; the probe still
works and still reports its length at `payload[8:10]`.  Thor simply does not
need it.

### Write-frame destuffing — `10 XX` → `XX`, uniformly

Only the leading `BW_CMD` is doubled (`10 10`); after that **every `10 XX` pair
on the wire destuffs to `XX`**, including `10 03`.

This was settled by checksum, not by assumption.  Four candidate rules were
tested against all four data-carrying write frames; **only this one makes all
four checksums validate**:

| rule | `0xDA` | `0x68` | `0x82` | `0x71` |
|---|---|---|---|---|
| **`10 XX` → `XX`** | ok | **ok** | ok | **ok** |
| only `10 03` → `03` | ok | BAD | ok | BAD |
| `10 10`→`10`, `10 03`→`03` | ok | BAD | ok | BAD |
| nothing collapses | ok | BAD | ok | BAD |

`0x71`'s data contains **4 literal `0x03` bytes**, escaped as `10 03` on the
wire.  A writer that does not escape `0x03` will emit a frame the device
terminates early — this is the same defensive ETX escaping Blastware does, and
it is mandatory, not optional.

Checksum is plain SUM8 of the destuffed payload; the DLE-aware form gives the
same answer once destuffing is correct, so it does not discriminate.

### The `offset` field is a per-command constant

| SUB | data bytes | `offset` | note |
|---|---|---|---|
| `0xDA` | 256 | `0x0100` = 256 | = the name-field size |
| `0x68` | 88 | `0x005A` = 90 | **identical to Series III's documented value** |
| `0x82` | 28 | `0x001C` = 28 | **identical to Series III's documented value** |
| `0x71` | 2090 | `0x082C` = 2092 | = the length `SUB 0x1A` reports |

⚠ **There is no single length formula** — two are `len`, two are `len + 2`, and
Series III's `data[1] + 2` rule does not reproduce either.  Treat these as
observed constants the device accepted.  Pinning the actual rule needs a second
capture whose payloads differ in size.  (This document has already been wrong
once by inferring a length field; do not infer this one.)

### The write body **is** the read body

The `0x71` write payload and the `0xE5` read response align at a fixed
**11-byte shift** with **1902/2090 bytes equal (91.0%)** — the remaining 9% is
exactly what was edited.

**So a setup is read-modify-write**, the same shape as Series III, and a write
client does not need to synthesise a config block from scratch.

### Field map, from the read/write diff

Twelve differing regions, all accounted for.  Offsets are into the `0x71`
**data** section (= the read response's data + 11).

| offset | size | field | this capture |
|---|---|---|---|
| `0x0000` | 3 | block length echo | `08 2a` (=2090) → zeroed on write |
| `0x0009` | 1 | unidentified | `3c` → `02` |
| `0x0013` | 1 | unidentified | `00` → `40` |
| `0x002A` | 44 | **setup file name** | → `TEST1.mmb` |
| `0x0090` | 42 | note 1 value — `Location` | → `Test Location 1 - 1234 electric boogaloo` |
| `0x00D0` | 42 | note 2 value — `Client` | → `TMI` |
| `0x0110` | 42 | note 3 value — `Company` | → `ServersDownLabs` |
| `0x0150` | 42 | note 4 value — `General Notes` | → `Hopefully this works!` |
| `0x0428` | 16 | sensor location | → `Test Config Send` |
| `0x06CA` | 4 | **Tran trigger level** float32 BE | `0.3` → `0.5` in/s |
| `0x06FA` | 4 | **Vert trigger level** | `0.3` → `0.5` in/s |
| `0x072A` | 4 | **Long trigger level** | `0.3` → `0.5` in/s |

**Notes block: four entries on a 64-byte stride**, each `[label: 22][value: 42]`,
labels at `0x008A + 64n`.  The labels are `Location`, `Client`, `Company`,
`General Notes` — **not** Series III's `Project:` / `Client:` / `User Name:` /
`Seis Loc:`, and they are fixed-width fields, not `label: value` pairs.

**Channel blocks: six, 48 bytes each**, from `0x06BC` — `Tran`, `Vert`, `Long`,
`Mic`, `LMic`, `SMic`.  Trigger level sits at **label + 30**.

### The geophone scale factor is in the config block — and it confirms our LSB

Each geo channel block carries a float32 BE at **label + 24**:

```
40 46 98 dd  =  3.10308003…
```

Which closes a loop from the file-decode work:

```
3.10308003… / 10000        =  0.000310308      = _GEO_LSB_IPS, to 8 figures
10.0 / 3.10308003… × 10000 =  32226.046        = the 32226.05 full scale
```

`_GEO_LSB_IPS` was derived statistically — intersecting 991,415 rounding
constraints from Thor's own CSV exports (see `idf_protocol_reference.md`).
**The unit reports the constant directly**, and it agrees.  That upgrades the
value from a fit to a reading, and explains the odd full-scale count: the
Micromate's ADC is **10,000 counts per volt**, and 3.10308 in/s per volt is
**exactly half** Series III's documented `6.206053` (ratio 1.99997).

Do **not** retune `_GEO_LSB_IPS` — this is corroboration, not a correction.

### Still unknown on the write path

- **What `0x68` and `0x82` actually contain.**  Both were written with
  near-zero payloads here and nothing in them changed, so no field is located.
  Series III maps backlight/power-save/LCD-cycle into `0x68`; unverified here.
- **The three header bytes** at data `0x0000`/`0x0009`/`0x0013`.
- **Scheduler / call-home writes.**  `callhome.MMB` is a file too, so it may go
  through the same `0xDA` + block-write shape with a different target name.

### ✅ `0xDA` CREATES setup files (confirmed on the device, 2026-09-24)

`TEST1.mmb` did not exist on the unit before the push.  Afterwards it is
**present in the unit's setup list and selected as the active config** —
verified on the Micromate's own screen, not inferred from the ack.

So the four commands below are the complete setup-management path, and a
homebrew client needs no file-transfer primitive and no pre-existing target:

```
0x41  read the active setup's name
0x1A  read its config block          ─┐ read
0xDA  name the target .MMB           ─┘ modify
0x71 → 0x72  write the block back      write  (creates the file if absent)
```

That closes the last open question about whether Series IV setup management is
reachable from outside Thor.  It is.

### ✅ Overwriting is protocol-identical to creating (2026-09-24)

The firmware carries strings that suggest an overwrite handshake:

```
Overwrite File                     MFS FILE EXISTS
Cannot be Overwritten              Can Not Delete Active Setup File
```

**There is no such handshake on the wire.**  A second push to `TEST1.mmb` — a
name that now existed, and which `SUB 0x41` confirmed was the *active* setup —
produced a byte-for-byte identical command sequence:

| | create | overwrite |
|---|---|---|
| SUBs, and their order | `5B 41 08 2E 1A DA 68 73 82 83 71 72` | **identical** |
| every `offset` field | — | **identical** |
| `0xDA` / `0x68` / `0x82` data | — | **0 bytes differ** |
| `0x71` data | — | 18 bytes differ = the one edited note |
| all seven write acks | 11 zero bytes | **identical, still all-zero** |

No extra command, no confirm step, no error status, and **no dialog on Thor**.
Those firmware strings belong to the on-device Save screen (the
`CSaveSetupFile` UI class), not to the protocol.

**The overwrite was verified on the unit itself** — the edited `General Notes`
string is present in the setup on the device, so the write applied, silently and
in place.

This is the case that mattered most, and it landed the right way round: the
target was the **active** setup, which is what a real remote config push would
hit.  Being active bought it **no protection** — it was overwritten directly.
A writer therefore needs no exists-check and no overwrite negotiation.

⚠ That cuts both ways.  A remote push to the active setup of a **monitoring**
unit changes the config it is recording with, with no prompt, no warning and no
distinguishable ack.  Whatever SFM eventually exposes should gate this on the
operator, not on the protocol — the device will not stop anyone.

⚠ Two overwrite cases remain untested: a **non-active** setup file, and
`factory.MMB`, which `Cannot be Overwritten` probably guards.  Neither blocks a
writer — the active setup is the one worth pushing to.

⚠ Note also that the write acks are **all-zero in every case observed**, across
a create and an overwrite.  We have never seen this protocol report a *failed*
write, so **do not treat a zero ack as proof a write was applied.**  Read the
config back and compare; `SUB 0x41` plus `SUB 0x1A` make that cheap.

## The scheduler, and a generic file transfer (2026-09-24)

Capture: `bridges/captures/9-24-26 - micromate2/*_read_scheduler.bin`.
25 request frames, 25 responses, every checksum valid.

Four operations in one capture, segmented by Thor's `POLL` preamble (marks are
not written into the raw `.bin` — in TCP mode seismo_lab logs them to the
on-screen log only):

| frames | operation |
|---|---|
| 0–4 | read the schedule off the unit |
| 5–18 | push a setup — the sequence already documented above, unchanged |
| 19–21 | write the schedule back |
| 22–24 | enable the scheduler |

### 🔑 `SUB 0x94` / `0x48` / `0x8D` / `0x8E` — file transfer by path

**The unit will read and write files addressed by full filesystem path.**

```
0x94  <path>   → 0x6B    open for READ
0x48           → 0xB7    read next page; repeat until an all-zero response
0x8D  <path>   → 0x72    open for WRITE
0x8E  <body>   → 0x71    write the file body
```

The path is plain ASCII, **unpadded**, with `offset` = its exact length:

```
5c 73 79 73 74 65 6d 5c 73 63 68 65 64 75 6c 65 5c 73 63 68 65 64 75 6c 65 2e 64 61 74
\system\schedule\schedule.dat                        29 bytes, offset = 0x001D
```

`0x94` and `0x8D` sent **byte-identical** 29-byte payloads — the same path, only
the command distinguishing read from write.

The `0x48` read is paged, and the page number comes back in the response header
at `payload[3:5]`:

| call | page | payload | content |
|---|---|---|---|
| 1st | `0x0000` | 15 B | descriptor — `04 00 … 01 00 00 00 00 00 01` |
| 2nd | `0x0002` | 535 B | the file body |
| 3rd | `0x0000` | 11 B | all zeros = **end of file** |

⚠ This is the single most consequential find of the session, because it is not
specific to the scheduler.  **`callhome.MMB` is a file too**, and the call-home
config is the last unsolved piece of the project.  Reading it should be a
matter of pointing `0x94` at the right path.  That is a *lead*, not a result —
no path other than `schedule.dat` has been tried.

### The schedule file itself

The write body is the read body minus an 11-byte response prefix (524 vs 535),
so both describe the same structure:

```
03 00 00 00 0f 03 02 [namelen] [setup-file name]   …zeros…   27 03 10
```

- **`[namelen]` is a length prefix**: `0x28` = 40 for the 40-character name read
  off the unit, `0x09` = 9 for `TEST1.mmb` written back.  Confirmed both ways.
- The trailing `27 03 10` sits at `0x108` in the write body and `0x113` in the
  read — the same offset once the 11-byte prefix is accounted for.

**A schedule entry names a setup file.**  That is how the help text's "change
the record mode" works: an entry says *at this time, load this setup*.  It also
means the scheduler and the setup list are coupled — deleting a setup that a
schedule references is a foot-gun worth checking before we ever expose either.

✅ **The entry internals ARE decoded** — see *The schedule record format* below.
The file turned out to hold **two** 260-byte records, not one: `27 03 10` is the
second record's time/day fields, not a trailer.  `1/2h` is a half-hour slot and
both times check out against what the operator entered.

### `SUB 0x47` — the scheduler enable, probably

Step 4 is two bare `0x47` frames with no data, differing only in `params[7]`:

```
req  params = 00 00 00 00 00 00 00 [01] 00 00   → rsp 04 00 … 00 [01] 00 00 00 00 00 01
req  params = 00 00 00 00 00 00 00 [03] 00 00   → rsp 04 00 … 00 [03] 00 00 00 00 00 01
```

The response echoes the selector and ends `01`.  Its 15-byte shape is identical
to the descriptor `0x48` returns on page 0.

⚠ **Whether `0x47` sets or merely reads is genuinely undetermined.**  Both calls
returned the same trailing `01`, and there is no before/after to compare — the
scheduler was enabled in this same capture, so no "disabled" reading exists.
`params[7]` is also the token position Series III uses, which is suggestive but
not evidence.  **Do not implement an enable against this until a
disable-then-enable capture settles it.**

### The scheduler enable is NOT in the setup config block

A prediction made before this capture — that the `Scheduler On/Off` switch would
appear as a byte in the `0x71` block, because the unit's help text lists it in
the same edit screen as Record Mode and Sample Rate — **did not hold**.

The `0x71`, `0x68` and `0x82` payloads are **byte-identical** to the previous
capture's, all three, zero differences.  The scheduler was enabled without any
of them changing.

⚠ The test is weaker than it looks: the operator re-sent the *same* config, so
a byte-identical block is also what "nothing changed" looks like.  What it does
establish is that enabling the scheduler did **not** require a config write —
whatever `0x47` does, it does alone.

## Monitoring control and the setup list (2026-09-24)

Capture: `bridges/captures/9-24-26 - micromate2/*_turn_on_monitormode_*`.
Thor started monitoring, listed the unit's setups, and stopped monitoring.
**40 request frames and 40 responses, every checksum valid.**

### Thor's per-operation preamble

Thor re-runs this before *every* operation — three times in this one capture:

```
POLL (0x5B)  →  SERIAL (0x15)  →  0x49  →  POLL (0x5B)  →  <the operation>
```

⚠ **Whether the unit requires this is untested** — see *Thor's conventions vs
the protocol's requirements*.  Thor sends it before trivial reads too, so it may
be habit rather than handshake.  Do not assume it is mandatory.  Note `POLL` here carries `offset = 0x0030` (its data
length), not `0xFFFF` — `POLL` is the one read Thor still addresses by length.

### `SUB 0x96` / `0x97` — start and stop monitoring ✅

Identical to Series III, including the acks:

| request | ack | effect |
|---|---|---|
| `0x96` | `0x69` | **start monitoring** |
| `0x97` | `0x68` | **stop monitoring** |

Both are bare frames — no params, no data — and both ack with the usual 16-byte
zero-data response.  Thor follows each with a `SUB 0x1C` status read to confirm.

### `SUB 0x1C` — monitor status, 55-byte data

```
data[12]        monitoring flag   0x0E monitoring / 0x00 idle
data[13]        day
data[14]        month
data[15:17]     year, uint16 BE
data[17]        ⚠ unidentified — 32 while monitoring, 100 when idle
data[18]        hour          ─┐ device clock, verified against the capture's
data[19]        minute         │ own wall time (19:12:25 → 19:13:34 EDT)
data[20]        second        ─┘
data[-8:-4]     memory total, uint32 BE   = 15,000,000 bytes exactly
data[-4:]       memory free,  uint32 BE
```

⚠ **The monitoring flag is `0x0E`, not Series III's `0x10`.**  Do not reuse the
Series III constant.  Only five bytes differ between the monitoring and idle
responses: the flag, `data[17]`, the clock, and the memory-free field.

Memory free dropped by exactly **4,096 bytes** across the ~70-second monitoring
session — monitoring allocates as it runs, so free memory is not a stable value
to compare against.

The device clock is free here, on a command Thor already sends.  That is worth
having: nothing else read so far reports the unit's own time.

### `SUB 0x49` → `0xB6` — a cheap state check

16-byte data, in Thor's preamble before every operation:

```
05 00 00 00 00 00 00 00 00 00 00 [ST] e8 00 0b 00
                                   ↑ data[11]: 0x02 monitoring, 0x00 idle
```

**A second monitoring indicator, in a 21-byte response instead of `0x1C`'s 60.**
Different encoding from `0x1C`'s flag (`0x02` vs `0x0E`), so they are separate
fields, not the same byte read twice.  For a polling client this is the cheaper
of the two, and Thor evidently treats it as the routine one.

### `SUB 0x3F` / `0x40` — walking the setup-file list ✅

The same first/next shape as Series III's `1E`/`1F` event walk, applied to
setup files:

| request | ack | meaning |
|---|---|---|
| `0x3F` | `0xC0` | **first** setup record |
| `0x40` | `0xBF` | **next** setup record; repeat until the name is empty |

Every record is 266 bytes of data and carries **nothing but the name**:

```
ff 00 00 00 00 00 00 00 00 00 00   <null-terminated name>   00 …
└──────────── 11-byte header ─────┘
```

The walk ends on a record whose name is empty — 24 records for 23 setups.
`factory.MMB` comes first, from `0x3F`.

**There is no active-setup flag in the list.**  The header is byte-identical on
every record including the terminator, and the tail is all zeros.  The active
setup is identified only by `SUB 0x41`, whose response uses this exact record
format.  (On the unit's own screen the active setup is marked with a trailing
asterisk — that is a UI decoration, not a field.)

`TEST1.mmb`, created over the wire earlier the same day, appears in the list and
is what `0x41` reports as active — independent confirmation that a written setup
becomes a real, enumerable file.

### Thor refuses to send a setup while a unit is monitoring

Send-to-unit is **greyed out in Thor's UI** when the unit is monitoring, and
nothing is transmitted — this capture contains no `0xDA` or `0x71` at all.

That is a Thor-side policy, not a device refusal: nothing observed suggests the
Micromate would reject the write.  It squares with the earlier finding that a
push to the **active** setup overwrites silently — Thor is preventing exactly
the footgun the protocol leaves open.

The *reason* behind the rule is sound and SFM should honour it — but not
necessarily by copying the greyed-out button.  Stop → push → restart as a single
operation is what an operator usually wants, and `0x49` data[11] (or `0x1C`
data[12]) makes the state check cheap either way.

## Static analysis of the firmware (2026-09-23, solo session)

### Architecture

**ColdFire / 68K, big-endian, Freescale MQX RTOS** — not ARM as the vector
table first suggested.  The giveaway is the function epilogue/prologue
`4E 5E 4E 75 4E 56` = `UNLK A6` / `RTS` / `LINK A6`, littered through both
images, plus an `MQX_OK` assertion string.

### CB vs BD: the same source, ~10 lines apart

A byte diff is useless — **68% of bytes differ** because the two are separately
linked builds with everything relocated.  A *string-set* diff is
position-independent and tells the real story: **17,128 strings shared**, and
almost every "unique" string is the same message with a different source line
number:

```
CB:  MONITOR[3268]: STATUS_BATTERY_LOW
BD:  MONITOR[3258]: STATUS_BATTERY_LOW      ← consistently 10 lines apart
```

The offset is exactly 10 across `STATUS_BATTERY_LOW`, `STATUS_BATTERY_CRITICAL`,
`Battery Critical Exit Monitor`, `histogram interval size of 0` and
`Offsets by channel` — so one ~10-line block differs in the monitor module and
essentially nothing else.  The only functional string unique to either build is
`CITIZEN` (a receipt-printer brand) in BD.

**This corroborates the bench A/B from the other direction:** the CB/BD split
is a tiny code delta, not two protocol stacks.  Whatever drives Instantel to
ship two downloads, it is not a different wire protocol.

⚠ The `SUB` dispatch is a 68K switch jump table (`CMPI.L` bounds check →
`MOVE.W (table,PC,Dn)` → `JMP (d8,PC,Xn)`).  Byte-pattern hunting will not
isolate the write opcodes — that needs a real disassembler.

### Call-home config field names, from the firmware's own debug dump

```
CallHome.Enable               = %s
CallHome.DialString           = "%s"
CallHome.Retries              = %d
CallHome.SessionTimeout       = %d
CallHome.WaitForConnection    = %d
CallHome.WarmupTime           = %d
CallHome.PowerSave            = %s
```

Seven fields, which is what the 126-byte `SUB 0x2C` block has to encode.  Note
`SessionTimeout` and `PowerSave` have no Series III equivalent, and Series III's
scheduled-time fields (`time1/time2 hour/min`) are absent here — consistent
with Series IV moving scheduling into the THOR-downloaded scheduler instead.

Also present: `AT+CSQ` (signal quality), so the firmware talks AT to the modem
directly.

## All five bench events, downloaded and decoded

Read-only, over USB, no Instantel software:

| key | declared size | got | decoded |
|---|---|---|---|
| `055d4a81` | 4,076 | 4,076 | histogram, 1 interval, 16:33:16 |
| `055d4a82` | 11,032 | 11,032 | waveform, 3072 × 4 ch, 16:33:19 |
| `055d4a83` | 11,502 | 11,502 | waveform, 3072 × 4 ch, 16:33:27 |
| `055d4a84` | 13,424 | 13,424 | waveform, 3072 × 4 ch, 16:33:34 |
| `055d4a85` | 8,746 | 8,746 | waveform, 2048 × 4 ch, 16:33:36 |

Every event arrived at exactly its declared size, every channel came out equal
length, and the timestamps are sequential across the recording session.

### Record type + filename: generate it, don't detect it

`read_idf_file()` decides waveform vs histogram from the **filename suffix** —
and there is no filename when downloading over the wire.

⚠ Worth correcting a natural assumption: **Series III does not detect this from
content either.**  `event_file_io.derive_record_type_from_filename()` reads the
last character of the extension (`M529LKIQ.G10H` → `H` → Histogram).  Nothing
in the codebase infers record type from file content, for either family.

And there is no obvious type field to find.  The first 64 bytes of a histogram
and a waveform are byte-identical; they diverge at ~`0x0947` into wholly
different structures rather than differing by a flag.

**The answer is the Series III pattern — generate the name.**  Series III has
`blastware_filename()`, which builds a name from serial + timestamp + type.
Series IV needs the same thing, and its convention is far simpler:

```
<serial>_<YYYYMMDDHHMMSS>.IDF{W,H}          e.g. UM12947_20260923163319.IDFW
```

versus Series III's `<letter><serial3><4-char base-36 stem><AB0T ext>`, where
the stem is base-36 of seconds-since-1985 ÷ 1296.

All three inputs are already available on a direct download:

| input | source |
|---|---|
| serial | `extract_binary_metadata()` — decoded from the IDF header |
| timestamp | `extract_binary_metadata()` — same |
| **type** | **the chain walk** — `SUB 0x0A` length `0x1E` = histogram, `0x00` = waveform |

Verified against all five bench events: the generated names match the
convention of real files in the production store byte for byte.  A directly
downloaded event can therefore be filed under exactly the name Thor would have
given it, and `/db/import/idf_file` needs no change at all.

⚠ The type still comes from the *protocol*, not the payload — so a downloader
must carry it out of the chain walk.  Losing it means losing the ability to
name the file correctly.

### ⚠ Unresolved: the `0x0C` peak float

The float32 extracted from `0x0C` runs 2–5% above `max(Tran, Vert, Long)` from
the decoded samples:

| key | `0x0C` float | max channel |
|---|---|---|
| `…81` | 3.5152 | 3.4419 |
| `…82` | 1.3720 | 1.3706 |
| `…83` | 2.3542 | 2.2227 |
| `…84` | 3.5152 | 3.4419 |
| `…85` | 0.4227 | 0.4198 |

It is not peak vector sum either (computed PVS runs *higher* than both).  The
field may not be the peak at all — its offset was inferred from a byte marker,
not established.  **Do not rely on it** until it is pinned properly.

Worth noting the histogram (`…81`) and the loudest waveform (`…84`) report
*identical* peaks to four decimals, in both measures.  That is self-consistent:
the histogram's single 1-minute interval spans the whole thumping session, so
its maximum should equal the loudest event in it.

## Event download, per-event delete, and ACH config (2026-09-25)

Three captures with operator-supplied ground truth, including Thor screenshots of
the event list and the schedule.  All Thor-originated; we still send nothing.

### ✅ `Day` is the day of week — and `[0]` is the schedule type

A **weekly** schedule settles both fields.  Thor's screen: *"multiday, Repeat
Weekly: Disabled, Start Monitoring at 8:00 AM every day, alternating
TEST1 / test2."*  The file is **7 × 260 + 4 = 1,824 bytes**, and every record
matches:

| rec | `1/2h` | time | `Day` | | `[6]` | name |
|---|---|---|---|---|---|---|
| @0 | 16 | 08:00 | **0** | Sun | 2 | `TEST1.mmb` |
| @260 | 16 | 08:00 | **1** | Mon | 2 | `test2.mmb` |
| @520 | 16 | 08:00 | **2** | Tue | 2 | `TEST1.mmb` |
| @780 | 16 | 08:00 | **3** | Wed | 2 | `test2.mmb` |
| @1040 | 16 | 08:00 | **4** | Thu | 2 | `TEST1.mmb` |
| @1300 | 16 | 08:00 | **5** | Fri | 2 | `test2.mmb` |
| @1560 | 16 | 08:00 | **6** | Sat | 2 | `TEST1.mmb` |

**`Day` = 0 Sunday … 6 Saturday**, and the alternating setup names line up with
the screen row for row.

`[0]` on record 0 read **4** here, against `2` and `3` in the daily schedules —
so it carries the schedule *type* as well as Repeat:

| `[0]` | meaning |
|---|---|
| `2` | Daily, repeat off |
| `3` | Daily, repeat on |
| `4` | **Weekly, repeat off** |
| `5` | Weekly, repeat on — ⚠ predicted, not observed |

Bit 0 is Repeat; `2` and `4` are the Day/Week bases.  Same bitmask style as
`[6]`.

⚠ In *daily* schedules `Day` reads `3` in every record of every capture.  With
`[0]` already saying "daily", the field is presumably ignored there — but that is
inference, and `3` is unexplained.

### Event download — `SUB 0x93` arms each event

Thor's download of six events:

```
POLL 0x15 0x49 POLL 0x1C 0x15 0x49 0x06
    0x93 0x1E 0x0C  0x5A×4          ← event 1
    0x93 0x1F 0x0C  0x5A×11         ← event 2
    0x93 0x1F 0x0C  0x5A×12         ← event 3
    0x93 0x1F 0x0C  0x5A×14         ← event 4
    0x93 0x1F 0x0C  0x5A×9          ← event 5
    0x93 0x1F 0x0C  0x5A×6          ← event 6
    0x93 0x06 0x93 0x1E  0x0A×9     ← re-walk the list
```

**`SUB 0x93` → ack `0x6C`** is sent before every event, with empty params and an
all-zero ack.  It is the Series IV analogue of Series III's `1E(token=0xFE)` arm
step, and it is simpler: no token, no params.

**Event keys are a plain sequential counter.**  The six events walked as
`055D4A81 … 055D4A86`, consecutive, at `data[11:15]` of the `0x1E`/`0x1F`
response, with the event's byte size at `data[17:19]`.  Nothing like Series III's
flash addresses.

`SUB 0x0A` walks the event list returning **30-byte records** carrying start and
stop timestamps (`[11]` day, `[12]` month, `[13:15]` year BE …), terminating on
an all-zero record.  The dates match Thor's event list exactly (five events on
09/23/2026, one on 09/24/2026).

### 🔑 Delete is PER-EVENT — `SUB 0xA8` + `0xAA`

```
0xA8   params[0:4] = <event key>   → ack 0x57
0xAA   params = zeros              → ack 0x55
```

The operator deleted the **top row of Thor's list** — the newest event — and
`0xA8` carried `05 5d 4a 86`, the **highest** key from the walk.  Thor lists
newest-first, so newest = highest key.  Confirmed end to end.

**This is the last piece a homebrew ACH receiver was missing.**  The manual
established that collection state is server-side bookkeeping and that "delete
events from unit" is a separate server-issued action; this is that action's
opcodes.

Two things worth noting against Series III:

- Series III erases **everything** (`0xA3 → 0x1C → 0x06 → 0xA2`).  Series IV
  deletes **one named event**, which is strictly safer — a receiver can delete
  exactly what it has confirmed it stored.
- The opcodes are different.  Do not reach for `0xA3`/`0xA2` here.

⚠ `SUB 0x06` (storage range) read **identically before and after** the delete in
this capture, so it is not a quick way to confirm a deletion landed.  Re-walk the
event list instead.

### ✅ ACH config — `0x2C` / `0x7E` / `0x7F`, exactly Series III

Thor enabling then disabling Auto Call Home, twice through the same sequence:

```
POLL → 0x15 → 0x49 → POLL → 0x2C (read) → 0x7E (write, 126 B) → 0x7F (confirm)
```

Same SUBs and the same acks as Series III — `0x2C`→`0xD3`, `0x7E`→`0x81`,
`0x7F`→`0x80`.  The write payload is **126 bytes** with `offset = 0x007E`, and
the `0x2C` read returns those same 126 bytes behind an **11-byte prefix**, so
`read[n + 11]` is `write[n]`.

**The enable flag is `write[5]`** (= `read[16]`):

| value | state |
|---|---|
| `0x05` | Auto Call Home **enabled** |
| `0x04` | **disabled** |

Bit 0 is the enable — the same bit position Series III uses at `raw[5]`, but with
bit 2 set as well in both states.  Do not test for `0x01`/`0x00`.

`write[6:]` holds the dial string, null-padded — `"RADIO RING"` on this unit,
matching Series III's `raw[6:46]`.

#### ⚠ `write[118:120]` — a field the UNIT maintains

Three samples:

| when | enable | field |
|---|---|---|
| 00:50, idle (schedule session) | `0x04` | `43 23` |
| 01:16, ACH read #1 | `0x04` | `43 23` — **unchanged after 26 min** |
| 01:16, ACH read #2, after a config write | `0x05` | `2e 5e` |

Thor **echoes back whatever it last read**, including a value that no longer
matches the config it is sending — and the write is accepted anyway.

What that rules out:

- **Not a clock or timer** — identical across 26 minutes of idle.  Only a config
  write moved it.
- **Not a counter** — it went *down*, 17187 → 11870.
- **Not computed by Thor** — Thor demonstrably sends a stale value.
- **Not a standard CRC16.**  Brute-forced all 65,536 polynomials × init
  {`0x0000`, `0xFFFF`} × all four reflection combinations, over four candidate
  regions (`[0:118]`, `[0:118]+[120:]`, `[5:118]`, and the payload with the field
  zeroed).  **No match.**  Recorded so nobody repeats the sweep.

It behaves like a hash or checksum the unit computes over the stored config: a
one-byte input change (`0x04`→`0x05`) scattered the output completely, XOR
`0x6D7D`, where a sum would have moved by 1.  But two samples cannot separate
that from a nonce regenerated on each write.

✅ **Operationally this does not matter.**  The unit does not validate the field
on input — Thor proved that twice in one capture.  **Read the config, change your
field, send everything else back byte-for-byte.**  Never synthesise or zero it.
Same read-modify-write shape as the setup block.

Resolving it properly needs more samples: several config writes with the times
noted, which is cheap to collect during any future ACH capture.

Neighbouring bytes: `write[117]` is `0x03`, lining up with Series III's
`num_retries` of 3, and `write[120:122]` is `00 3c` = 60.

⚠ Beyond the enable byte and the dial string, **the field map is not
established** — only one setting was varied.  Series III's offsets are a starting
hypothesis, not a transfer.

## What the THOR manual settles about the ACH session (2026-09-24)

Source: `manuals/723U0201 THOR Operator Manual Rev 08.pdf` §6.2, §3.6.
⚠ **This is vendor documentation, not observed bytes.**  It tells us the shape of
the session and the vocabulary; it does not give opcodes.  Treated as a strong
prior, not as confirmed protocol.

### 🔑 A unit does NOT learn that an event was accepted — the server decides

This document previously listed, as the thing *gating any homebrew receiver*:

> how a unit announces itself, and **how it learns an event was accepted so it
> stops re-sending it.**

**That was the wrong question.**  There is no acknowledgement mechanism to
discover, because the unit is not tracking what has been collected.  Per §6.2.2.2
the ACH session is a list of **server-chosen actions**, and two of them are
independent:

| ACH action | what it does |
|---|---|
| **Copy events** | downloads events — *"Only applies to events not previously downloaded"* |
| **Copy monitor log** | downloads the monitor log |
| **Delete events and Logs from Unit** | **explicitly** deletes them from the unit |
| **Set Date/Time** | unit synchronises its clock **from the computer** |
| Send Events / Schedule to Vision | Instantel cloud, not relevant to us |

"Not previously downloaded" is **computer-side bookkeeping**.  The unit keeps its
events until a server tells it to erase them, and the manual's own warning proves
the two are decoupled:

> ⚠ *"If you enable 'Delete Events and Logs from Unit' but disable 'Copy Events'.
> The events and logs will be deleted without being uploaded."*

A server that never issues the delete simply re-reads the same events forever.

**This is exactly the model our Series III ACH server already implements** —
`ach_state.json` with `downloaded_keys` / `max_downloaded_key`, and erase as a
separate deliberate step.  No new mechanism is needed for Series IV.

So a homebrew receiver needs: accept the connection, identify the unit, walk the
events (already solved — `0x08`/`0x1E`/`0x0A`/`0x5A`), keep our own high-water
mark, and *optionally* erase.  **The erase opcodes are the only genuinely missing
piece**, and they remain on the unsafe list.

### The session is server-driven

Which matches the firmware's state machine and its
`CMD_EXIT_CALL_HOME_DELETE_EVENTS_START_MONITORING` /
`Call Home Deleting Events` strings: the unit dials in, then waits to be told
what to do, and returns to monitoring when the server is finished.

Scheduled ACH and event-triggered ACH behave differently (§3.6.1):

- **Event-triggered** (ACH enabled on the unit, an event occurred) — *"Records
  events + transfer data, no stopping to monitor."*  Governed by **Monitoring
  While Calling Home**; with MWCH enabled *"the monitor log will not be copied,
  events will not be deleted, and time will not synchronize"*.
- **Scheduled** — *"Stop monitoring + transfer data (or delete events, or
  synchronize)."*  Monitoring stops and in-progress events complete first.

⚠ Worth noting for SFM: with MWCH on, an event-triggered session **cannot**
delete or sync.  A receiver that relies on erase to avoid re-reading will
silently never erase on those units.  Our high-water mark must be the primary
mechanism, with erase as an optimisation — which is how Series III already does it.

### Session Time Out is unit-side only

§6.2 is explicit that THOR configures ACH on both ends *"with one exception; the
Session Time Out.  This must be configured directly on the unit."*  That matches
the firmware's `CallHome.SessionTimeout` field, and means it lives in
`callhome.MMB` rather than anywhere THOR reaches — one more reason to read that
file with `0x94`.

### Unit identification is by serial number, with wildcards

§6.2.2.2: filters match against the serial number the unit presents, `*`
wildcards allowed, best-match wins — `UM*` all Micromates, `MP*` all Minimate
Pros, `BE`/`BC` Minimate Plus variants.

So **the unit announces its serial early enough for the server to route on it.**
`SUB 0x15` returns the serial and is in Thor's standard preamble, which is
consistent, though the inbound direction has still never been observed.

⚠ The manual's worked example contradicts its own table — it says *"To filter
Micromate units type MP*"* and *"To filter Minimate Pro units type UM*"*, which is
backwards.  The table is right.  Noted so nobody copies the error.

### Thor requires Idle for configuration

§6.2.1 step 3: *"The Monitoring Mode must be 'Idle'."*  Independent confirmation
that the greyed-out send observed on the bench is deliberate Thor policy, applied
to ACH setup as well as compliance setup — and still not evidence the device
refuses.

### The schedule's actions, from the UI side

§3.6 lists exactly **four** actions THOR exposes: Start monitoring, Stop
monitoring, Self-check, Auto Call Home.  The firmware has **five** — it also has
a setup-less `DUTYCYCLE_START_MONITOR`.

§3.6.2 step 8 confirms the coupling is Thor's own requirement: *"Select the
appropriate Action and the Unit Setup. (A Unit Setup must exist…)"* — while §3.6.2
step 1 notes *"The unit will execute any actions in a schedule using its current
settings."*  The two statements sit a paragraph apart and pull in opposite
directions, which is consistent with `START_MONITOR` existing and THOR never
emitting it.

Also from §3.6: schedules are **Day or Week** ("Select Day… Select Week"), have a
**Repeat Daily / Repeat Weekly** flag, carry a Name, Description and Unit Type,
and *"Saving a schedule will only store it on the computer, it must still be sent
to the unit and enabled"* — which is why the capture shows `0x8E` (send) and
`0x47` (enable) as separate steps.

Those are the fields to look for when the schedule entry is finally decoded:
**action, setup name, time, day-or-week, day selection, repeat.**  The one
captured entry (`03 00 00 00 0f 03 02 …`) has seven bytes before the name, which
is the right order of magnitude for that field list.

## What THOR's "status check" actually is (2026-09-25)

Captured with `bridges/mm_link.py` standing in for the modem, THOR configured
`Communication: TCP`, **Check connection every 5 s**, **Check status every 5 s**.

### It is eleven commands, not one

```
POLL → DEVICE_INFO → 0x49 → 0x5C → MONITOR_STATUS → SETUP_NAME_READ
     → STORAGE_RANGE → 0x02 → OPERATOR → 0x47 → CALLHOME_CFG
```

Each check opens a **new TCP connection**, runs all eleven exchanges in ~300 ms,
and closes it.  Measured over 21 consecutive checks:

| | |
|---|---|
| request payload | **236 B** |
| response payload | **936 B** |
| TCP connect + teardown | ~1,000 B (fresh handshake every time) |
| **per check** | **~2.2 KB** |

### The cost

| cadence | checks/day | per unit |
|---|---|---|
| **every 10 s (as observed)** | 8,640 | **18.8 MB/day — 563 MB/month** |
| every 60 s | 1,440 | 3.1 MB/day — 94 MB/month |
| every 15 min | 96 | 0.2 MB/day — 6 MB/month |

⚠ **On a metered cellular plan this is real money, and most of it is waste.**
A status check re-reads the call-home config, the operator name, the active setup
name and the full device info block **every ten seconds** — none of which changes
between checks.  `SETUP_NAME_READ` alone returns 274 bytes each time.

A genuine liveness + state check is **`POLL` + `MONITOR_STATUS`**: two commands,
131 bytes of response.  Same information about whether the unit is alive and
whether it is monitoring, for a ninth of the traffic.

### There are TWO checks, and only one of the dials is honoured

The eleven-command sequence is the **status check**.  There is also a much
cheaper **connection check** — three commands:

```
connection check   POLL → SERIAL → 0x49
status check       POLL → DEVICE_INFO → 0x49 → 0x5C → MONITOR_STATUS
                   → SETUP_NAME_READ → STORAGE_RANGE → 0x02 → OPERATOR
                   → 0x47 → CALLHOME_CFG
```

⚠ **With both dials equal, the connection check never runs separately at all** —
every connection observed was the full eleven.  It only appears once the two
intervals differ.  That alone explains a lot of "changing the settings does
nothing": at equal values you only ever get the expensive one.

With *connection check = 10 s* and *status check = 30 s*, the steady state over
14 consecutive cycles is:

```
FULL    at T
short   at T + 10.1
short   at T + 25.3      (+15.2)
FULL    at T + 30.4      (+5.1)
```

| dial | set to | observed | |
|---|---|---|---|
| status check | 30 s | **30.4 s** | ✅ honoured |
| connection check | 10 s | **15.2 s** | ❌ 52% slow |

27 consecutive connection-check gaps, all 15.1–15.3 s.  That is systematic, not
jitter.

> ⚠ **A "half the status period" hypothesis was raised here and REFUTED.**  It
> predicted that with status at 60 s the connection check would run at 30 s.
> Measured: **60.5 s**.  Recorded rather than deleted — the prediction was sharp
> and it was wrong.

### The connection dial sets a COUNT, not a period

Three configurations, each measured over many cycles:

| status | connection | status observed | connection observed | separate conn checks per status cycle |
|---|---|---|---|---|
| 10 s | 10 s | 10.1 s ✅ | — | **0** (none ever ran) |
| 30 s | 10 s | 30.4 s ✅ | 15.2 s ❌ | **2** |
| 60 s | 30 s | 60.5 s ✅ | 60.5 s ❌ | **1** |

**The status dial is honoured in all three** — within about 1%.

**The connection dial is not honoured in any of them.**  What actually holds is:

```
separate connection checks per status cycle  =  (status ÷ connection) − 1
```

1 → 0 checks, 3 → 2 checks, 2 → 1 check.  So the field does not set a period; it
sets how many checks fit inside a status cycle, and it is off by one.

Consequences worth naming:

- **Setting the two dials equal yields zero connection checks.**  Every
  connection is then the expensive eleven-command status check.  That is the
  default-looking configuration, and it is the worst one.
- **"Every 30 s" with status at 60 s gives one check per minute** — half the
  advertised rate.
- The observed cadences are not multiples of the connection setting at all
  (10 s → 15.2 s; 30 s → 60.5 s), so no simple scale factor describes it either.

⚠ The *timing within* a cycle is still not fully explained.  At status 30 /
connection 10 the two short checks landed at T+10.1 and T+25.3, where an evenly
divided cycle would put them at T+10 and T+20.  The count rule holds across all
three configurations; the phase does not follow from it.

**For SFM:** this is the concrete argument against exposing a dial whose
behaviour is not what it says.  If a control sets an interval, it should set that
interval — and if the effective rate is derived from something else, the UI
should show the derived value, not the input.

### Traffic at each configuration

| status / connection | per minute | per day | per month |
|---|---|---|---|
| 10 s / 10 s | ~13 KB | **18.8 MB** | **563 MB** |
| 60 s / 30 s (current) | ~3.4 KB | **4.9 MB** | **147 MB** |
| a minimal check at 60 s | ~0.2 KB | **0.3 MB** | **9 MB** |

The last row is `POLL` + `MONITOR_STATUS` only — enough to answer "alive?" and
"monitoring?".

### ⚠ What this means for the "won't stay connected" failure

Because idle polling **opens a fresh connection each time**, a silently-dead link
is *less* dangerous here than expected — a dead socket fails at connect and the
next cycle simply tries again.  The earlier hypothesis (THOR wedged on a
half-open socket during polling) is weakened by this.

The exposure is during **operations**, not polling: THOR was observed holding one
connection from **00:30 to 00:47 — seventeen minutes** — while downloading events
and pushing setups.  A link that dies silently mid-operation leaves THOR waiting
on a socket the OS will not fail for roughly the default keepalive (~2 h).  That
remains the best candidate for a unit that will not come back and where refresh
does nothing.

**Testable with `mm_link.py`:** set `blackhole` *during* a download or config
push rather than while idle, and watch whether THOR ever gives up, whether
refresh emits any bytes, and whether it recovers when `pass` is restored.

### Design notes for SFM

Three things this argues for, all cheap:

1. **Separate liveness from inventory.**  Poll `POLL` + `MONITOR_STATUS`
   frequently; read config, setup name and call-home settings only when
   something says they changed, or on demand.
2. **Show the connection attempt.**  Every check here has a visible outcome —
   connected, frames exchanged, closed.  THOR surfaces none of it, which is why
   a failing unit is undiagnosable from the UI.  The log this section is built
   from took one afternoon to produce and answers questions THOR cannot.
3. **Bound every operation with its own timeout**, independent of TCP's.  Do not
   rely on the socket to report a dead peer.

## Reproduced: THOR stops polling after a connection drops mid-download

**2026-09-25, one trial, on the bench link.**  This is the failure the operator
described in the field on UM12947 — *"just wouldn't stay connected… clicking the
little refresh button did nothing… no way to view the status of a connection
attempt."*

### What was done

THOR was downloading events.  Partway through the bulk transfer the link was
faulted, the connection died, and the link was then fully restored.

⚠ **What THOR actually experienced was a TCP close mid-download**, not the silent
link that was intended — `mm_link.py` had a bug (`socket.timeout` subclasses
`OSError`, so 200 ms of quiet was mistaken for a closed socket).  Fixed, but the
run stands as a **drop-mid-download** test, which is arguably the more realistic
case: a modem losing its tower often does produce a real close.

### What happened

```
13:02:49.0   THOR connects, bulk download running (165 BULK_DOWNLOAD frames in)
13:02:50.0   link faulted -> connection dies mid-transfer
13:03:10.4   THOR reconnects once, 20 s later, sends SUB_1F (advance event
             pointer) — an attempt to resume
13:03:10.8   that attempt dies too
13:03:33.3   link fully restored and healthy
             ... 48 s of complete silence ...
13:04:21.4   operator clicks Refresh -> full 11-command check, every response
             correct, unit perfectly healthy
             ... 2 min 6 s of complete silence ...
13:06:27     still nothing.  That refresh is the ONLY connection since 13:03:10.
```

Before the fault THOR was connecting every 30 s without a miss for over an hour.

### What this establishes

1. **One retry, then give up.**  THOR retried once after 20 s and stopped.  No
   backoff, no further attempts.
2. **The automatic poll loop dies with it.**  Not just the download — the 30 s
   connection check and 60 s status check both stopped entirely.
3. **It does not recover when the link returns.**  Three minutes of healthy link
   produced zero connection attempts.
4. **Refresh works, but only once.**  The manual check succeeded completely —
   all eleven commands, correct responses.  It did **not** restart the automatic
   loop; two minutes later there was still nothing.

⚠ **Point 4 is the dangerous one.**  Refresh makes the UI report a healthy unit
while nothing is actually watching it.  The operator sees a successful check and
moves on; the background loop stays dead.  That is worse than a refresh button
that plainly fails, because the failure is *silent and looks like success*.

### 🔑 "Not polling" and "unreachable" are different states — and THOR conflates them

Demonstrated directly.  At **13:10:44**, six and a half minutes after THOR's last
connection of any kind:

```
THOR connections since 13:04:22 : 0
independent POLL through the same relay, same moment : succeeds, correct response
```

**The unit is reachable.  The link is fine.  THOR is simply not asking.**

Whatever THOR's UI reports in that window, it is wrong:

| if it shows | why it is false |
|---|---|
| connected / OK | nothing has been checked for minutes |
| disconnected / unreachable | it **is** reachable — THOR stopped trying |

The true state is *"I have given up checking this unit"*, and no display in THOR
expresses it.  So an operator cannot separate **the unit is down** from **the
poller is asleep** — and those demand completely different responses: a site
visit versus a mouse click.

This also settles a question the previous section left open.  It does not much
matter whether stopping after one retry is an intentional design choice or a
defect: **the reporting is wrong either way**, because both plausible displays
misrepresent reality.

It explains, too, why the field symptom resisted characterisation.  The unit is
perfectly reachable throughout.  Poke it and it answers.  Nothing is wrong with
the unit, the modem, or the link.

### Design consequences for SFM

1. **Retry must be unbounded with backoff**, never one-and-done.  A transient
   link fault should not take a unit out of service until a human notices.
2. **A manual check must restart the automatic loop**, or the UI must state
   clearly that automatic checking is stopped.  A green tick that means "it
   worked when you asked" is a trap.
3. **Surface the poll loop's own state** — last successful check, last attempt,
   next scheduled attempt, consecutive failures.  All four were invisible here,
   and every one of them would have made this diagnosable in seconds.
4. **Distinguish "unit unreachable" from "we stopped checking."**  Proven above
   to be different states that THOR cannot tell apart.  A unit's displayed state
   should be one of: *checks passing*, *checks failing* (with the attempt count
   and last error), or *not being checked* (with why, and a way to resume).
   Collapsing the last two into one indicator is the root of this whole failure
   being undiagnosable.

### Caveats

- **One trial.**  Not yet repeated, and not yet tried with a true silent link
  (`blackhole` with the fixed tool) or with a clean `drop`.
- THOR's UI state during the dead window was not recorded — worth capturing on a
  repeat, since what the operator *sees* is half the problem.
- Whether a THOR restart is required, or whether it eventually recovers on a
  longer timescale, was not tested.

## Thor's conventions vs the protocol's requirements

**SFM is not meant to reimplement Thor.**  Thor is the only available teacher of
the wire protocol, but almost nothing about *how* it sequences its work has been
shown to be required by the device.  Those are two different things and this
document should not blur them.

The distinction matters because several observations above were written as
"a client should do X" when the honest statement is "Thor does X, and we have not
checked whether the unit cares."

| Thor does this | required? | what it means for SFM |
|---|---|---|
| `POLL → 0x15 → 0x49 → POLL` before **every** operation | **unknown** | Series III needed `POLL`×3 before `5A` *specifically*, so a preamble requirement is plausible — but Thor sends this before trivial reads too.  **Testable:** issue one operation cold and see if it answers. |
| Reads with `offset = 0xFFFF`, skipping the probe | **no — the probe works** | We have a genuine choice, and the probe is arguably better: it reports the length instead of making us trust a fixed one.  This is the one place Thor's shortcut is probably worse. |
| Sends `0x68` + `0x82` in every setup push | **unknown** | In both setup captures these carried near-zero payloads and **changed nothing**.  If they are optional, a setup write is 3 frames instead of 7, with less to get wrong.  **Worth testing before we build the writer.** |
| Rewrites the whole 2,090-byte config for a one-field change | **unknown** | No narrower write has been observed.  Read-modify-write is safe and known; a targeted write would be nicer but is unevidenced. |
| Reads `0x41` twice in a row (scheduler capture, frames 6 and 8) | **no** | Plainly redundant.  A reminder that Thor's sequence is not a minimal one. |
| Greys out send-to-unit while monitoring | **Thor policy** | The *reason* is real — a push to the active setup overwrites silently.  But "the button is grey and you figure it out" is a UX choice, not the only answer.  SFM could offer stop → push → restart as one operation, which is what an operator actually wants. |
| Names the target with `0xDA` before a config write | **almost certainly required** | The unit has to know which file to write.  Closest thing here to a genuine protocol requirement. |

### The schedule↔config coupling is Thor's, not the protocol's

In Thor, putting "start monitoring" in a schedule forces you to attach a setup,
and sending the schedule pushes that setup at the same time — overwriting
whatever already has that name.  **The protocol does not require any of this.**

The evidence, from the scheduler capture:

- **The two writes are separate operations**, not one transaction.  The config
  write ends at frame 18 (`0x72`), then Thor sends a fresh `POLL` preamble, and
  only then opens the schedule for writing at frame 20.  Different commands,
  different paths, no shared state:

  ```
  config    0xDA → 0x68/0x73 → 0x82/0x83 → 0x71/0x72
  schedule  0x8D → 0x8E
  ```

- **The schedule stores a NAME, not a config.**  A length-prefixed filename is
  all an entry carries.  It is a *reference*, and references do not require the
  referent to be rewritten.

- **The config Thor pushed was already on the unit, unchanged.**  Its 2,090-byte
  `0x71` payload is **byte-identical** to the previous capture's — zero
  differences.  Thor spent a full block write re-sending a setup the device
  already had, purely to satisfy its own coupling.

So the sequencing is a Thor UI decision, and a costly one.

**What SFM can do instead, with today's protocol and nothing new:**

1. Enumerate the unit's setups with `0x3F` / `0x40` — cheap, read-only, and it
   yields the exact names a schedule may legally reference.
2. Write **only** the schedule (`0x8D` / `0x8E`) when the referenced setup is
   already present.
3. Push a config only when it is actually missing, or when the operator
   deliberately edited it — and say so explicitly rather than silently.

That removes the write entirely from the common case: scheduling against a setup
that already exists becomes a 524-byte schedule write instead of 524 bytes plus
a 2,090-byte config overwrite.

⚠ **And it removes a real hazard.**  Because the reference is by name, and
because a same-name write overwrites silently with an indistinguishable ack,
Thor's pattern means *scheduling* something can quietly rewrite a setup that
other schedules — or the operator's own saved work — depend on.  Nothing in the
protocol or the ack reports that this happened.  Validating the reference
instead of rewriting the referent avoids the whole class of problem.

### ✅ The schedule record format — DECODED (2026-09-24)

A debug printf in the scheduler names the fields outright:

```
SCHEDULER : _ReadRecord(%d) -> %s (Action=%u, 1/2h=%u, Day=%u, Setup="%s") [WDAY=%d]
```

The captured file is **two 260-byte records plus four zero bytes** — 524 bytes
exactly.  Every non-zero byte in the file falls inside those two records.

```
offset  field          size  notes
[0]     schedule flags 1     record 0 ONLY — 3 = repeat on, 2 = repeat off
[1:4]   padding        3     zero in every record observed
[4]     1/2h           1     half-hour slot, 0–47
[5]     Day            1     3 in every record observed — undetermined
[6]     Action         1     bitmask: 2 start, 4 stop, 8 self-check, 16 ACH
[7]     name length    1
[8:260] Setup name     252   null-padded; empty for actions that take no setup
```

⚠ `[0]` and `[6]` were **mislabelled twice** before a five-entry schedule
separated them — see *CORRECTED* below.

Decoded against the operator's own description of what they entered —
*"start monitoring TEST1 at 7:30 AM, Auto Call Home at 7:30 PM"*:

| | record @0 | record @260 |
|---|---|---|
| `Action` | **3** | **0** |
| `1/2h` | 15 → **07:30** ✅ | 39 → **19:30** ✅ |
| `Day` | 3 | 3 |
| `[6]` | 2 | 16 |
| name length | 9 | **0** |
| Setup | `TEST1.mmb` | *(none)* ✅ |

**Five independent confirmations**, no fitting required:

1. Slot 15 = 07:30 and slot 39 = 19:30, both matching the operator's stated
   times, on a 30-minute grid — and 39 − 15 = 24 slots = exactly 12 hours.
2. The name-length byte reads 9 for `TEST1.mmb`, 40 for the 40-character name in
   the read capture, and **0** for the Auto Call Home entry.
3. **Auto Call Home carries no setup name** — which is what it should do, and
   direct proof that a schedule entry can exist with no setup attached.
4. The 260-byte stride lands the second record's `1/2h` exactly at `[264]`.
5. `2 × 260 + 4 = 524`, the whole body, with nothing left over.

⚠ **Correction:** the `27 03 10` bytes at `[264]` were previously recorded here as
a possible trailer.  They are **record 2's `1/2h`, `Day` and `[6]` fields**.  Not
a trailer — a second record hiding behind an assumption that the file held one.

Action codes, so far:

| value | action |
|---|---|
| `0` | **Auto Call Home** |
| `3` | **Start monitoring, with setup** |

### ⚠ CORRECTED — `[6]` is the Action, and it is a bitmask (2026-09-25)

A five-entry schedule settled this, and it **overturns the two previous
readings of this record.**  The operator's Thor screen, captured alongside:

```
actioncheck — Micromate — Repeat Daily: Disabled
  7:30 AM   Start Monitoring   TEST1
  8:00 AM   Stop Monitoring
  8:30 AM   Self Check
  9:00 AM   Start Monitoring   test2
  7:30 PM   Auto Call Home
```

Five entries, and the file holds exactly five 260-byte records:

| rec | `[0]` | `1/2h` | `Day` | **`[6]`** | len | name |
|---|---|---|---|---|---|---|
| @0 | **2** | 15 = 07:30 | 3 | **2** | 9 | `TEST1.mmb` |
| @260 | 0 | 16 = 08:00 | 3 | **4** | 0 | |
| @520 | 0 | 17 = 08:30 | 3 | **8** | 0 | |
| @780 | 0 | 18 = 09:00 | 3 | **2** | 9 | `test2.mmb` |
| @1040 | 0 | 39 = 19:30 | 3 | **16** | 0 | |

**`[6]` tracks the action exactly, and the values are powers of two:**

| `[6]` | action |
|---|---|
| `2` | Start Monitoring |
| `4` | Stop Monitoring |
| `8` | Self Check |
| `16` | Auto Call Home |

Bits 1–4 of a bitmask.  **Bit 0 (value `1`) is unobserved** — a natural home for
the setup-less `DUTYCYCLE_START_MONITOR`, but that is a guess, not a finding.

**Two retractions:**

1. `[6]` was recorded as *"the one unidentified field"* and predicted to be the
   Repeat flag.  It is the **Action**.
2. `[0]` was labelled **Action**, then *"Action with repeat folded in"*.  Both
   wrong.  `[0]` is **non-zero only on record 0** — including here, where record
   0 and record 3 are the same action with different `[0]` values.  It is a
   **schedule-level field carried in the first record**, and it holds the Repeat
   flag: `3` repeat on, `2` repeat off, matching *"Repeat Daily: Disabled"* on
   the screen.

The earlier repeat capture was consistent with both readings because it had one
start entry and changed one byte.  A single-variable test is not always enough —
it took a schedule with four distinct actions to separate the two fields.

`Day` is `3` on all five records across every capture.  Still undetermined.

### ✅ `SUB 0x47` is the scheduler enable — confirmed by Thor's own notifications

Previously recorded as *"genuinely undetermined"* whether `0x47` sets or reads.
Thor's notification pane timestamps it:

```
00:51:29   schedule write (0x8E) completes
00:51:31   Thor: "actioncheck has been successfully SENT to ... UM12947"
00:51:31   → 0x47  params[7] = 0x01
00:51:33   → 0x47  params[7] = 0x03   (on the wire as 10 03 — DLE-escaped)
00:51:35   Thor: "actioncheck was successfully ENABLED on ... UM12947"
```

The only frames between *sent* and *enabled* are the `0x47` pair.  **`0x47`
performs the enable.**

⚠ The meaning of `params[7]` ∈ {1, 3} is still open — a single `0x47` with
`params[7] = 3` also appears at *session start*, where nothing is being enabled,
so the byte is more likely a selector than a value.  Note it must be
**DLE-escaped**: a bare `0x03` in params truncates the frame at what the device
reads as ETX.

### 🔑 Setups ARE written as raw `.MMB` files — second retraction

This document twice stated that setups are *not* pushed as raw `.MMB` blobs and
go only through `0xDA` + the config block.  **Wrong.**  In this session Thor did
both, choosing by whether the setup is the active one:

| setup | how it was written |
|---|---|
| `TEST1.mmb` (**active**) | `0xDA` → `0x68`/`0x73` → `0x82`/`0x83` → `0x71`/`0x72` |
| `test2.mmb` (not active) | `0x8D` `\system\setups\test2.mmb` → `0x8E` (2,192 B) |

So **`\system\setups\<name>.mmb`** is the setup directory, and the generic file
transfer writes there directly.

The `.MMB` file is *nearly* the compliance block: aligning the two gives
**1,968 / 2,086 bytes equal (94.3%) at a 4-byte shift**.  The file is 102 bytes
longer (2,192 vs 2,090) and puts the setup name at offset 38 where the `0x71`
block has it at 42.  So the on-disk setup and the wire compliance block are the
same structure with different framing — not two formats.

**This is the cleaner path for SFM.**  Writing a setup as a file needs two frames
and no `0xDA`/`0x68`/`0x82` ritual, and it does not disturb the active setup.

### File writes are chunked

The schedule's 1,304 bytes went as **two `0x8E` frames, 1,024 + 280**, each with
`offset` = that chunk's own length.  The 2,192-byte setup file went in **one**
frame, so 1,024 is not a hard ceiling — the chunking rule is unexplained and
recorded as observed.

### Six duty-cycle actions, not five

The full `_PSA()` dispatch — the scheduler's own action processor, reading records
via `_ReadRecord`:

```
DUTYCYCLE_START_MONITOR                            ← no setup
DUTYCYCLE_START_MONITOR_WITH_SETUP
DUTYCYCLE_START_MONITOR_WITH_SETUP_STOP_COMPLETE   ← a third start variant
DUTYCYCLE_STOP_MONITOR
DUTYCYCLE_CALLHOME
DUTYCYCLE_SELF_CHECK
```

**THOR exposes four.**  Two start variants it never offers, one of which needs no
setup file at all.

`_PSA() send ->> CMD_DUTYCYCLE_START_MONITOR` shows the setup-less action is live
code that the scheduler genuinely dispatches, not a dead case — it sends a real
message.  And since every one of these cases sits in the function that consumes
`_ReadRecord`'s `Action` field, **a schedule record can carry it.**

An alternative explanation was checked and ruled out: the Micromate does have a
separate *Timer Mode* (`MODE_TIMER`, `MODE_MONITOR_TIMER`, `Monitor Once Only`,
under Special Setup), so `START_MONITOR` could have belonged to that path
instead.  It does not — it is in `_PSA()`, the scheduler's dispatcher.

### The scheduler enable lives in SysPref, not in the setup

```
SysPref.bMonitorScheduler     = %s
_Task() CMD_SET_SCHEDULE -  MonitorScheduler ENABLED
_Task() CMD_SET_SCHEDULE -  ! MonitorScheduler NOT Enabled
```

This **confirms the failed prediction above from the other side**: the
`Scheduler On/Off` switch is a *system preference*, which is why the `0x71`
compliance block was byte-identical when the scheduler was turned on.  It was
never going to be in there.

It also suggests `0x47` — the two bare frames at the end of the capture — is a
SysPref get/set rather than anything scheduler-specific, which would explain its
`params[7]` selector and its 15-byte response shape being shared with `0x48`'s
page-0 descriptor.  ⚠ Still a hypothesis; the disable/enable capture settles it.

#### The schedule has a setup-less start action — Thor just never uses it

The coupling above looked like it might be a workaround for the unit crashing on
a missing setup.  The firmware says otherwise: **there are two distinct
start-monitoring actions**, and only one of them involves a setup file.

```
_PSA() case DUTYCYCLE_START_MONITOR
_PSA() case DUTYCYCLE_START_MONITOR_WITH_SETUP
```

The full duty-cycle action set, from the scheduler task's own dispatch strings:

| action | notes |
|---|---|
| `DUTYCYCLE_START_MONITOR` | **no setup involved** — uses whatever config is loaded |
| `DUTYCYCLE_START_MONITOR_WITH_SETUP` | the variant Thor always emits |
| `DUTYCYCLE_STOP_MONITOR` | |
| `DUTYCYCLE_CALLHOME` | `send ==> CMD_SCHEDULE_CALL_HOME` |
| `DUTYCYCLE_SELF_CHECK` | `call DailySelfCheck()` |
| `DUTYCYCLE_ON` / `OFF` / `NEXT` | scheduler state, not per-entry actions |

So the schedule↔config coupling is not a crash workaround — **Thor picks the more
demanding of two available actions, every time.**  SFM can emit
`START_MONITOR` and skip the config entirely.

On whether a missing setup *would* break the unit: the firmware suggests
graceful degradation rather than a crash — `Setup File Not Found` exists, and so
does `Invalid parameters reset to factory default - please review setup`, which
is a deliberate fallback.  ⚠ **Untested**, and not worth relying on until it is.

⚠ **Hypothesis, not a finding:** the schedule entry's leading byte may be the
action code — the one captured entry reads `03 00 00 00 0f 03 02 [namelen]
[name]`, and `03` would fit `START_MONITOR_WITH_SETUP` in an enum of the above.
Unverified; there is one entry and nothing to diff against.

**The capture that would settle it** (more valuable than the `0x47`
disable/enable test): build a schedule whose entry is **stop monitoring** or
**call home** — an action with no setup attached.  That would (a) confirm or kill
the action-code hypothesis, since a setup-less action should change the leading
byte and drop the name, and (b) prove from the other direction that a schedule
can be sent with no config push at all.

Note also `DUTYCYCLE_CALLHOME` → `CMD_SCHEDULE_CALL_HOME`: **a scheduled
call-home is a way to make a unit dial out on demand**, which is the one
remaining lever on the unsolved call-home direction.

### What this implies for the build order

The read path is fully known and needs no Thor-shaped decisions, so a read-only
client can be written now with confidence.  **The write path should not be built
by transcribing Thor's sequence** — the `0x68`/`0x82` question above decides
whether our setup write is 3 frames or 7, and it is answerable with one capture
plus one careful experiment.

Two reliability problems worth designing *against*, both observed rather than
assumed:

1. **A zero ack does not mean a write applied.**  Every ack seen is 11 zero
   bytes, across creates and overwrites alike, and no failing write has ever been
   observed.  Whatever SFM does, it should read back and compare rather than
   trust the ack.  `0x41` + `0x1A` makes that cheap.
2. **Nothing warns before clobbering a monitoring unit's active setup.**  The
   device will not stop it and the ack will not distinguish it.  That guard has
   to live in SFM.

## ⚠ Untested and unsafe-until-agreed

**Nothing below has been sent to a unit by us, and nothing should be without an
explicit decision.**

Note the distinction introduced on 2026-09-24: the setup-write sequence
(`0xDA`, `0x68`/`0x73`, `0x82`/`0x83`, `0x71`/`0x72`) has now been **observed**,
because Thor performed it while we recorded.  Observed is not the same as
exercised — **we have still never originated a write frame.**  The wire format
is known; our encoder is unwritten and unproven.

- **Writes** (`0x68`–`0x83`) — format known, never sent by us
- **Call-home write** (`0x7E` / `0x7F`) — ✅ observed 2026-09-25, acks `0x81` /
  `0x80`; enable flag isolated
- **Per-event delete** (`0xA8` / `0xAA`) — ✅ observed 2026-09-25, acks `0x57` /
  `0x55`.  **Destructive.**  Series III's `0xA3`/`0xA2` erase-all has *not* been
  seen on Series IV and may not exist
- **Start / stop monitoring** (`0x96` / `0x97`) — observed via Thor 2026-09-24,
  acks `0x69` / `0x68`
- **Scheduler enable** (`0x47`) — observed; `params[7]` semantics still unclear

Every one of these is now *observed*.  **None has ever been originated by us**,
and that is the line that still matters: the wire format is known, our encoder is
unwritten and unproven.
- `0x1F` (advance event pointer) — non-destructive on Series III but it does
  move device state, so it is parked with the rest

Also unknown:

- Whether `0x10` bytes inside **request params** need stuffing.  (Write-frame
  **data** stuffing *is* now settled — `10 XX` → `XX`; see *The write path*.)
- The **call-home session's bytes** — the device-initiated direction has still
  not been observed at all, and cannot be over USB.

  ✅ **The second half of this item is resolved** (2026-09-24, from the THOR
  manual): "how it learns an event was accepted so it stops re-sending it" was
  the wrong question — **it does not learn.**  Collection state is server-side
  bookkeeping and erase is a separate, optional, server-issued action.  See
  *What the THOR manual settles about the ACH session*.  What remains unknown is
  the framing of the inbound session and the **erase opcodes**, not any
  acknowledgement mechanism.

---

## Session provenance

Unit **UM12947**, firmware **`11.0CB`** (Blastware line), on the bench via
USB.  Every response in this document was checksum-validated.

Three sittings, all on the same unit:

| date | state | what was exercised |
|---|---|---|
| 2026-09-23 | zero events stored | read commands, direct from Python |
| 2026-09-23 | 5 events (4 waveform, 1 histogram) | the event chain + `0x5A` |
| 2026-09-24 | 5 events | **Thor pushing a setup**, via a recording relay |

The 2026-09-24/25 sittings used a topology worth keeping, because it is
reusable: `socat` on mint-mac shares `/dev/ttyACM0` on TCP, seismo_lab's TCP
bridge relays Thor to it and records both directions.  Thor is configured with
the unit at `127.0.0.1:<bridge port>` exactly as if it were a field modem.  No
modem, no SIM, and **nothing on the production Thor box is touched.**

**Run the relay with `-x` and keep its log.**  It hex-dumps every forwarded byte
in both directions, which makes it a complete second copy of every capture,
independent of whether seismo_lab was recording:

```
socat -d -d -x TCP-LISTEN:12345,reuseaddr,fork /dev/ttyACM0,raw,echo=0,b115200 \
    > ~/mm-captures/socat_<ts>.log 2>&1
```

That paid off on 2026-09-25, when a capture's `.bin` files never left the Windows
machine — the session was rebuilt from the relay log with
`scratch/socat_log_split.py`.  When the real bins arrived later the
reconstruction was **byte-for-byte identical in both directions** (3,595 and
4,004 bytes), so the log is a validated fallback rather than an approximation.
Sessions concatenate in one log; split them on the `accepting connection`
markers.

⚠ Firmware is the single biggest caveat on this document.  Every finding here
is from one unit on the Blastware build.  A `11.0BD` unit has not been
touched.

✅ The original "empty unit" limitation is gone — `0x08`, `0x1E`, `0x0A` and
`0x06` were all exercised against 5 stored events on 2026-09-23.