Tulip Stage A Initial Measurements

Headline numbers

• Picture size: 2592×1944. 5 MP JPEG; every Stage A and later run saved at this size.
• Sharpest edge width: 4.59 px. Flat target at 40 cm with autofocus, 10–90 % edge width. A truly sharp 5 MP image measures 2–3 px.
• Stage A 100-cycle run: 100 of 100. One early run, saved at full size with no failure, unexpected reset or SD error. Camera and SD stayed powered (power was not switched); later switched-power runs lost cycles, see Reliability.
• Sharp-sized frames, best recipes: 97 %. 25 cm, warm camera, autofocus right after the warm-up and 2 s in all (two recipes, about 35 cycles each).
• Powered time per photo: 10.8 s. tl_cycle v12b on the board, no retake; Stage A took 25.9 s.
• Mean JPEG size: 2.84 MB. Stage A indoor scene; 43,200 photos (one per minute for 30 days) would be about 123 GB.

Contents

• Sharpness and focus
• Exposure and colour
• Timing and powered time
• File size and storage
• Reliability
• Evidence and caveats

Sharpness and focus

How sharp the photos are and how well autofocus holds. Edge width and the line-pair, checkerboard and small-text scores are direct measurements; JPEG size and the sharp-sized share are the proxy used on long runs.

Edge width by distance

Width of a measured edge in pixels, so smaller is sharper. A missing autofocus-off bar means the edge was too blurred to measure.

Table view

Distance	Autofocus on, px	Autofocus off, px	Note
15 cm	4.95	—	No-autofocus frame too blurred to measure
40 cm	4.59	7.08	—
1.5 m	6.16	6.14	—
5 m	5.88	6.34	—
Far, outdoors	5.16	5.17	—

JPEG size against sharpness

Each dot is a photo scored with the line-pair test. The dashed lines mark the largest size seen for a blurred photo and the smallest seen for a sharp one, which is why file size can stand in for sharpness on long runs.

Table view

Photo	Run	Recipe	JPEG, MB	Line pairs	Checker	Small text	Class
c373	Overnight run 1	No AF	2.095	82	151	69	blurred
c4	Overnight run 1	2 s	2.102	97	182	81	blurred
c3	Overnight run 1	3 s	2.115	102	188	80	blurred
c298	Overnight run 1	4 s	2.12	76	97	82	blurred
c1	Overnight run 1	12 s	2.132	113	211	83	blurred
c37	Overnight run 1	No AF	2.139	86	158	74	blurred
c111	Overnight run 1	5 s	2.147	144	244	71	blurred
c356	Overnight run 1	12 s	2.168	104	218	79	blurred
c374	Overnight run 1	12 s	2.173	111	239	72	blurred
c11	Overnight run 1	5 s	2.193	375	541	140	partial
c130	Overnight run 1	5 s	2.21	108	192	73	blurred
c358	Overnight run 1	3 s	2.226	281	459	114	partial
c240	Overnight run 1	3 s	2.256	352	515	134	partial
c5	Overnight run 1	2 s	2.258	741	695	271	partial
c378	Overnight run 1	2 s	2.307	1127	755	436	sharp
c29	Overnight run 1	12 s	2.329	1187	784	496	sharp
N6 c2	Night 6	1+1 s / 1+2 s	2.12	99	183	—	blurred
N6 c100	Night 6	1+1 s / 1+2 s	2.12	84	153	—	blurred
New card c20	Card check	1+1 s	2.357	949	739	—	sharp
T04_001	v12b part b	1+1 s	2.356	516	664	155	not classed
T05_001	v12b part c	1+1 s	2.347	490	663	146	not classed

Close-range autofocus recipes

Sharpness scores at close range for each autofocus recipe, from three test patterns. Higher is sharper. The table view has the verdict, JPEG size and powered time.

Table view

Recipe	Photo	Wait, s	Powered, s	JPEG, MB	Line pairs	Checker	Small text	Verdict	Sharp cycles	Cycles scored
AF at once 1+4 s	T20_028	5	17.4	2.31	1272	930	494	sharp	3	4
AF at once 1+2 s	T20_022	3	15.2	2.29	1071	905	393	sharp	4	4
AF at once 0.5+1.5 s	T20_023	2	14.1	2.26	925	878	328	sharp	3	4
AF at once 0.5+0.5 s	T20_024	1	12.1	2.02	51	65	56	blurred	0	4
7 s, then AF 1+4 s (today)	T20_026	12	23.3	2.07	63	83	64	blurred	2	3
2 s, then AF 1+4 s	T20_027	7	18.6	2.11	85	169	71	blurred	0	4
No AF, 2 s	T20_025	2	13.1	2.06	79	155	70	blurred	0	4

Frame sharpness class by recipe

How each autofocus recipe’s frames split into blurred-sized, partial and sharp-sized JPEGs during the first night run.

Table view

Id	Recipe	Wait, s	Blurred	Partial	Sharp	Failed	Sharp, %	Powered, s
0	Today: 7 s, then AF 1+4 s	12	2	5	28	1	80	23.2
1	AF at once 1+4 s	5	1	6	29	0	81	16.3
2	AF at once 1+2 s	3	0	10	20	5	67	14.2
3	AF at once 0.5+1.5 s	2	0	1	32	2	97	13.3
4	AF at once 1+1 s	2	0	1	33	1	97	13.4
5	1+2 s, twice	6	0	18	15	2	45	17.1
6	0.5+1.5 s, twice	4	2	11	22	0	63	15.3
7	1 s, then AF 1+2 s	4	0	10	24	1	71	15.3
8	2 s, then AF 1+2 s	5	0	12	23	0	66	16.2
9	No autofocus (control)	2	35	0	0	0	0	12

Sharpness against powered time

Each circle is a recipe, labelled by its number. Further left costs less battery; higher gives more sharp-sized frames. The hollow circle is the control.

Table view

Id	Recipe	Sharp, %	Powered, s	Wait, s
0	Today: 7 s, then AF 1+4 s	80	23.2	12
1	AF at once 1+4 s	81	16.3	5
2	AF at once 1+2 s	67	14.2	3
3	AF at once 0.5+1.5 s	97	13.3	2
4	AF at once 1+1 s	97	13.4	2
5	1+2 s, twice	45	17.1	6
6	0.5+1.5 s, twice	63	15.3	4
7	1 s, then AF 1+2 s	71	15.3	4
8	2 s, then AF 1+2 s	66	16.2	5
9	No autofocus (control)	0	12	2

Blurred phase after a start

How long the camera keeps producing blurred-sized frames after a start. Faded bars were still blurred when the run ended, so the true length is longer.

Table view

Run	Interval, s	Cycles	Minutes	Phase ended
Overnight run 1, cold start	25	28	12	yes
Night 6, cold start	18	120	36	yes
Night 7, after a 40 min rest	18	15	4.5	yes
v12b smoke test and part a	60	170	170	no

Exposure and colour

Brightness, colour and how long the camera waits for exposure to settle before it shoots.

Powered time by settle window

The settle window is the pause before the shot so exposure and colour can stabilise. A longer window keeps the camera powered longer.

Table view

Window	Window, s	Powered, s	JPEG, MB	Brightness, %	Red	Green	Blue	Sharpness, %
12 s	12	23.8	2.254	0	1	1	1	0
4 s	4	15.9	2.252	-0.9	0.987	0.996	0.996	0.7
2 s	2	13.7	2.237	-0.3	0.995	0.997	1.004	-2.3
1 s	1	12.7	2.232	-1.5	0.953	0.996	1.002	-3.8
0.5 s	0.5	12.4	2.271	-3.4	0.877	1.007	1.011	0.3

Colour channels by settle window

Each channel’s level relative to the longest window. Below the dashed line means weaker than at the longest window.

Table view

Window	Red	Green	Blue
12 s	1	1	1
4 s	0.987	0.996	0.996
2 s	0.995	0.997	1.004
1 s	0.953	0.996	1.002
0.5 s	0.877	1.007	1.011

Brightness and sharpness change

Percent change against the longest window, for each shorter window.

Table view

Window	Brightness, %	Sharpness, %
12 s	0	0
4 s	-0.9	0.7
2 s	-0.3	-2.3
1 s	-1.5	-3.8
0.5 s	-3.4	0.3

Far-scene autofocus timing

Sharpness of the face, torso and a background pumpkin as a percentage of today’s recipe, for each autofocus timing tried. The table view has powered time and the red shift on the stone wall.

Table view

Condition	Wait, s	Powered, s	Face, %	Torso, %	Pumpkin, %	Red, %
7 s, then AF 1+4 s (today)	12	23.8	100	100	100	0
2 s, then AF 1+4 s	7	18.9	100	102	100	-1.1
AF at once 1+4 s	5	16.7	100	100	102	-1.1
AF at once 1+2 s	3	14.7	100	100	98	-1.1
AF at once 0.5+1.5 s	2	13.8	99	100	98	-1.8
AF at once 0.5+0.5 s	1	12.8	101	100	93	-5.4
No AF, 2 s	2	13.7	99	100	93	-1.8

Measured image properties

Each property with what was measured, the evidence for it and the project note it came from.

Property	Measured	Evidence	Source
Output size	2592×1944 (5 MP) JPEG	verified	stage-a-findings
Sharpest edge (10–90 % width)	4.59 px at 40 cm with autofocus; a truly sharp 5 MP image is 2–3 px, so roughly half the linear detail	verified	stage-a-findings
Noise	Low in good light; grainy in evening room light	verified	stage-a-findings
Highlight clipping	15–16 % of the frame pure white (bright windows, sky); EV −2 only lowers that to about 14 %	verified	stage-a-findings
Exposure control	Automatic exposure beat every manual setting in low light; manual saturates between 2,000 and 8,000 lines; gain 240 stayed darker than auto at dusk	verified	stage-a-findings
EV compensation	About 10–15 brightness levels per step in steady light; settings reset on every photo	verified	stage-a-findings
Evening room light	Dark, green in the middle, purple/pink in the corners, grainy (subject at about 1 m)	verified; fails the charter bar for room scenes	stage-a-findings
Colour settling	Red channel about 5 % low after 1 s and 12 % low after 0.5 s; indistinguishable from 12 s at 2 s	verified (one frame per window)	settle-window-sweep
Autofocus range	Essential under about 1 m; little or no difference beyond about 1.5 m (the lens rests near far focus)	verified	stage-a-findings
Field of view on focus	Sharp frames are shifted 12 px (0.5 % of the width) sideways compared with blurred ones	inference (likely lens movement)	settle-window-sweep
Wrong picture size	After a power cycle the camera can return 1600×1200 (about 0.9 MB) instead of full size; a retry fixed it every time	verified	stage-a-findings
Truncated JPEG	Reported length normal but no end marker: 3.2 % of cycles in the first overnight run, 14 % in the A/B night, up to 34 % of first attempts in the late v12b session	verified; cause open	settle-window-sweep, v12b-hardware-results
Camera SPI clock	8 MHz maximum (manufacturer specification)	documented	settle-window-sweep
Night scenes	Out of scope; dark frames averaged about 2 of 255 in brightness and held no usable edges	verified	stage-a-findings

Timing and powered time

How long the camera stays powered for each photo, and where that time goes.

Powered time along the optimisation path

Powered seconds per photo at each step. Faded bars are estimates rather than hardware measurements.

Table view

Step	Powered, s	Range, s	Basis
Stage A: tl_cycle v2, AF on, 12 s window	25.9	22.7–27.0	measured (100 cycles)
Window sweep baseline: 12 s window, AF off	23.8	—	measured (5 cycles)
2 s window, AF off, library read	13.7	—	measured (5 cycles)
2 s window, burst read, 32 KB, SD 16 MHz, with read-back	12.5	—	measured (6 cycles); read-back costs about 3.2 s
Same path without read-back	9.3	—	estimate from log timestamps
tl_cycle v12b on the board, 60 s interval, no retake	10.8	10.4–11.1	measured (459 cycles)
tl_cycle v12b with one retake	—	19.3–21.4	measured; +8.7 s per retake

Where the powered seconds go

Start-up, settle window, transfer to the SD card and everything else, for each configuration.

Table view

Configuration	Start-up, s	Window, s	Transfer, s	Other, s	Total, s	Basis
tl_cycle v11 (today)	1.2	12	7.4	0.6	21.2	split estimated from code and logs; measured total 21.0 s
Best path tested (estimate)	0.8	2	5.9	0.6	9.3	estimate from log timestamps; SD 16 MHz, no read-back

Transfer path timing

Time to move one photo from the camera to the SD card for each data path, split into camera read, SD write and everything else.

Table view

Path	Camera, ms	SD, ms	Other, ms	Transfer, ms	Powered, s
Library read, 8 KB, SD 8 MHz (today)	4380	5100	566	10046	18
Library read, 8 KB, SD 16 MHz	4365	3816	565	8746	15.4
Burst read, 8 KB, SD 8 MHz	2279	5083	371	7733	15.7
Burst read, 32 KB, SD 8 MHz	2273	4553	368	7194	15.2
Burst read, 32 KB, SD 16 MHz	2273	3239	367	5879	12.5

Start-up times by SD card

Milliseconds after power-on until the SD card is ready, the camera is ready and the first full-size frame arrives.

Table view

Card	SD ready, ms	Camera ready, ms	First frame, ms
16 GB card	452	705	830
128 GB card	557	809	—

Measured stage timings

Timings for individual stages of a capture cycle, with the project note each came from.

Stage	Value	Source
SD card ready after power-on	450 ms (16 GB card), 557 ms (128 GB card)	stage-a-findings, settle-window-sweep
Camera ready after power-on	700 ms (16 GB card), 809 ms (128 GB card)	stage-a-findings, settle-window-sweep
First full-size warm-up frame	about 830 ms after power-on (16 GB card)	stage-a-findings
Settle window (warm-up to photo)	12 s in tl_cycle v11; 2 s in v12b	settle-window-sweep, tl-cycle-12-draft
Capture call	median 105 ms in good cycles, 89 ms in failed ones (night run 6)	settle-window-sweep
Camera to controller read, 2.15 MB	4.38 s with the library; 2.27 s with burst reads (0.94 MB/s)	settle-window-sweep
Write to SD, 2.15 MB	5.1 s (8 MHz, 8 KB writes) down to 3.2 s (16 MHz, 32 KB writes); 4.7 s median on the 128 GB card at 8 MHz	settle-window-sweep
Wake to power-off, whole photo	25.9 s in Stage A; 10.8 s in v12b (no retake)	stage-a-findings, v12b-hardware-results
Cost of one retake	+8.7 s powered per retaken photo in v12b (+1.3 s for the earlier wrong-size retry)	v12b-hardware-results, stage-a-findings

File size and storage

How big the photos are and what a long run needs from the card.

JPEG size by scene

Smallest to largest file seen in each scene. The tick marks the mean or typical size.

Table view

Scene	Smallest, MB	Largest, MB	Typical, MB	Note
Dark spot, 100-cycle run (session 14)	0.49	0.84	—	Mean brightness about 2 of 255
Dark spot, 15 cm (sessions 10–12)	0.7	0.84	—	AF and no-AF about 0.79 vs 0.80 MB
Wrong default size 1600×1200	0.9	0.9	0.9	About 0.9 MB
Bright room scene (v11 baseline)	1.54	1.67	1.62	Cycles 7 to 10
25 cm focus page, blurred frames	2.03	2.18	—	Night runs and v12b
Two studio lights, 15 cm (session 12)	2.2	2.2	2.2	About 2.2 MB
Window-sweep still life (window means)	2.232	2.271	—	25 cycles, good lighting
25 cm focus page, sharp frames	2.28	2.37	—	Night runs, night 6 and the new card
Stage A indoor scene (100 cycles)	2.18	3	2.84	Mean 2.84 MB

Storage for a long run

Gigabytes needed at several JPEG sizes, with dashed lines at the capacity of common cards. Solid bars use the measured Stage A mean; grey bars use sizes from the plan table.

Table view

JPEG, MB	Run total, GB	Basis
1	43	plan table
2	86	plan table
2.84	123	measured Stage A mean
3	130	plan table
4	173	plan table
5	216	plan table

Card	Capacity, GB
256 GB (planned, not bought)	256
128 GB (in use)	128
16 GB (original)	16

Reliability

Truncated files, retakes, wrong-size captures and SD read-back failures.

Photo outcomes in the latest firmware build

Each session’s photos split into good on the first try, good after a retake, and failed. The table view has the retake rate.

Table view

Session	Cycles	First try	Rescued	Failed	Retake, %	Battery low, mV	Battery high, mV
2 (smoke test)	20	20	0	0	0	4004	4024
3 (part a)	150	150	0	0	0	4016	4084
4 (part b)	150	133	14	3	11.3	4076	4140
5 (part c)	139	92	31	16	33.8	4132	4188

Retakes through the late session

Retaken photos per block of cycles, to show whether retakes cluster late in a session.

Table view

Cycles	Cycles in block	Retakes
1–30	30	4
31–60	30	12
61–90	30	11
91–120	30	11
121–139	19	9

Truncated JPEGs through the A/B night

Truncated files per block of cycles across the night of the A/B test.

Table view

Cycles	Minutes	Truncated
1–30	9	5
31–60	18	9
61–90	27	11
91–120	36	16
121–150	45	19
151–180	54	7
181–210	63	1
211–240	72	4
241–270	81	1
271–300	90	0
301–330	99	2
331–360	108	0
361–390	117	0
391–420	126	0
421–450	135	2
451–480	144	0
481–510	153	0
511–540	162	0

Truncated JPEGs by condition

Truncated files per condition, split by whether the file length was still growing when read or had stopped without an end marker.

Table view

Condition	Kind	Cycles	Truncated	Growing	Steady	Powered, s
0: AF 1+1 s, poll on	condition	135	19	1	18	11.7
1: AF 1+1 s, poll off	condition	135	13	0	13	11.5
2: AF 1+2 s, poll on	condition	135	26	18	8	12
3: AF 1+2 s, poll off	condition	135	19	9	10	12.2
Poll on, pooled	pooled	270	45	19	26	11.84
Poll off, pooled	pooled	270	32	9	23	11.85

Wrong-size photos

Share of cycles whose first capture came out at the default size instead of the one requested, by test and autofocus setting.

Table view

Test	Autofocus	Wrong size	Cycles	Rate, %
v4, evening room light	not stated	7	20	35
v5, autofocus on	on	8	8	100
v5, autofocus off	off	1	10	10
Midday, bright, AF	on	0	10	0
Midday, dimmed, AF	on	0	10	0
Midday, dimmed, no AF	off	0	10	0
Dark spot, AF (session 10)	on	5	10	50
Dark spot, no AF (session 11)	off	0	10	0
Two studio lights, AF (session 12)	on	0	10	0
Dark spot, AF (sessions 13 and 14)	on	110	110	100

Failed SD read-backs

Share of verified photos whose read-back differed from what was written, by clock setting.

Table view

Condition	Clocks	Verified	Failed	Rate, %
Night run 1, autofocus cycles 1–200	write 16 / read 16 MHz	171	5	2.9
Night run 1, autofocus cycles 201–380	write 16 / read 16 MHz	159	32	20.1
Night run 1, no-AF control	write 16 / read 16 MHz	38	14	36.8
Night run 2, write 8 / read 8 MHz	41-cycle run before the hang	12	0	0
Night run 2, write 16 / read 8 MHz	41-cycle run before the hang	9	0	0
Night run 2, write 16 / read 16 MHz	41-cycle run before the hang	10	2	20

Evidence and caveats

How to read these numbers

Sharpness on long runs is judged from JPEG file size. The size classes were calibrated against photos scored directly with the line-pair, checkerboard and small-text tests, so a size class is a proxy rather than a measurement.

Faded bars on the powered-time path are estimates, and so is the stage split in the powered-seconds chart; the basis column says how each was worked out. The other powered-time bars were measured on hardware.

One step on the path, the latest firmware with a retake, only has a range, so it appears in that chart’s table but not on the chart.

Storage bars drawn in grey use sizes from the plan table where no measurement exists.

Headline facts

Headline fact	Value	Note
Picture size	2592×1944	5 MP JPEG; every Stage A and later run saved at this size
Sharpest edge width	4.59 px	Flat target at 40 cm with autofocus, 10–90 % edge width. A truly sharp 5 MP image measures 2–3 px
Stage A 100-cycle run	100 of 100	One early run, saved at full size with no failure, unexpected reset or SD error. Camera and SD stayed powered (power was not switched); later switched-power runs lost cycles, see Reliability
Sharp-sized frames, best recipes	97 %	25 cm, warm camera, autofocus right after the warm-up and 2 s in all (two recipes, about 35 cycles each)
Powered time per photo	10.8 s	tl_cycle v12b on the board, no retake; Stage A took 25.9 s
Mean JPEG size	2.84 MB	Stage A indoor scene; 43,200 photos (one per minute for 30 days) would be about 123 GB
Blurred-sized below	2.18 MB	Photos under this size were blurred in 9 of 9 measured (night run 1)
Sharp-sized from	2.28 MB	Photos from this size were sharp in 3 of 3 measured; sizes between are mixed

Numbers come from the project’s test notes and brief: the Stage A findings, settle-window sweep, hardware results, A/B test plan, test runner, build notes and USB photo transfer notes.