Skip to content

The demo shoot, the evidence screenshots, and the Instawards final report - #74

Open
AncungAulia wants to merge 18 commits into
mainfrom
docs/video-script
Open

AncungAulia wants to merge 18 commits into
mainfrom
docs/video-script

Conversation

@AncungAulia

Copy link
Copy Markdown
Owner

Fifteen commits, all documentation and evidence. No application code changes.

What a reviewer can now open

  • docs/FINAL_REPORT.md — the Instawards submission report in the five sections the form asks for: project and team, scope of work, evidence of completion, the verification checklist, and twenty coded links. Drawn from the same facts as EVIDENCE.md, which stays organised by SOW §6.1 because that is the shape for closing gaps.
  • docs/shots/ — 18 screenshots of the deployed app, nine pages at a phone and a laptop width, taken from production with no wallet, plus shoot.cjs that takes them.

The evidence page had claimed the directory screenshot was done. What existed was docs/design/polish/exports, taken on 15 September from a local dev server before any of this was deployed, and its two "console" images are the connect-wallet screen because that tool has no wallet and never got in. The runner profile, which the SOW names in as many words, had never been photographed at all.

Two failures while taking them are recorded in docs/shots/README.md, because both produced a picture that looked fine and was not: waiting on a static heading photographed three cards still reading "Race 41", and the header's rolling placeholder was caught with a word cut in half.

What the shoot needs

  • docs/video-script.md — the spoken script for the founder introduction, the nine-beat shot list against a 3:00 budget, the race brief field by field, the terms, five runners ready to type, and step-by-step for the two fraud beats.
  • docs/rehearsal/demo/posters/prambanan-sunrise.jpg — a fifth poster in the same template as Axel's four, for the race created by hand rather than seeded.

Two findings shaped it, both cited where they live: claim_racepack checks only the operator and the record's state, with no status or date condition, so entries need not be closed first; and a close date of today or earlier makes reserve_slot refuse with RegistrationClosed(20) whatever the status, which would lock the organiser out of her own race.

The SOW names create event as the first step of the loop and the shot list had dropped it. It is now beat 0.

Known, not fixed here

The lead featured card on the directory lets the poster's own text bleed through the fade, visible in docs/shots/directory-1440.png. It is the first image a reviewer opens.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU

AncungAulia and others added 15 commits September 27, 2026 09:36
STE-28 has been the last open SOW item for days and had no script, so the
three-minute cap was being spent on guesswork at record time. This writes the
eight beats against a time budget and says which two to protect when it runs
long: the fraud attempts, because they are what a reviewer came for.

The founder introduction is written as a separate export rather than a longer
single file. The SOW already names the demo's contents, and seventy seconds of
introduction in front of a three-minute cap costs a beat.

Two things here exist to stop the video from claiming more than the product
does. The claim limits are copied from social/x-intro-post.md, where they were
researched, because a spoken line is not reviewed the way a commit is. And beat
5 records that the wait must stay in the cut: a forwarded QR is accepted within
its step, so a clip of a fresh screenshot being refused would show behaviour we
do not have.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
Filming the loop on the seeded races looked like the cheap path and is not one.
Their organiser, desk and runner secrets only exist in the .env.demo of the
machine that ran the seed, because the script refuses to write a secret into
anything it commits, and without the organiser key there is no console and no
way to add a desk. The race that already ran is spent besides: claims and
published results are both one-shot.

So this records the setup for a race of our own, dated today, which also stops
the camera showing finish times published for a race three weeks away.

Two findings turn this into a job one person can do alone, and both are cited
where they live: claim_racepack checks only operator and Entered, with no status
or date condition, so entries need not be closed first; and Freighter runs on a
phone, which this app already pairs with over WalletConnect on testnet, so the
pass can sit on a real phone in front of the laptop's webcam.

Five runner accounts rather than three, because a pack is claimed once and a
published result is terminal, so a spoiled take needs a spare runner rather than
a second attempt.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
The evidence page claimed the directory screenshot was done. What existed was
docs/design/polish/exports, taken on 15 September from a local dev server before
any of this was deployed, and its two console images are the connect-wallet
screen, because that tool has no wallet and never got in. Everything else under
docs/design is the design the code was built from, drawn before the code.

So the runner profile, which the SOW names in as many words, had never been
photographed at all. These nine pages are the ones a script can reach without a
key; the pass, the desk, the signed-in console and the results review need one
and come out of the video shoot instead.

Two failures are worth keeping, because both produced a picture that looked
fine and was not. Waiting on the heading "Race record" photographed three cards
still showing "Race 41" and a category of "Unknown": that heading is static, and
a profile resolves its races after its records. And the header's placeholder
rolls on a CSS wheel, so the shutter caught a word cut in half. The script now
waits for a pattern to disappear rather than appear, and strips animations
before it fires.

1x rather than 2x: at 2x the landing page alone was 6MB, and these live in git.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
Uploading results was being grouped with the race-day beats and is not like
them: one browser, one file, no camera, no phone. What makes it awkward is only
its precondition, that a result needs a claimed record, so it has to come last.
Both CSVs are written out, the held one carrying a different anomaly kind per
row so the review screen has all four to show.

Scanning, the stale screenshot and the duplicate are one setup and three takes.
The duplicate in particular needs no second machine and no second network: one
laptop offline makes both desks offline, and the queues cannot see each other
because each Chrome profile has its own IndexedDB.

The decision that carries the rest is entering on the phone rather than the
laptop, because a pass reaches the runner in the browser they entered from.

And a fallback that costs one shot instead of the shoot: the desk's manual entry
is built for a camera that fails, so if Freighter will not pair on the phone
every mechanism still proves itself on the laptop alone.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
Both beats were written as outcomes rather than as things a person does alone
with a phone and a laptop, and each turns on a detail that decides whether the
take works at all.

Beat 5: a QR carries its own step, and the desk refuses a step more than one
away before computing anything, so a screenshot is good for 30 to 60 seconds
depending where in a step it was taken. 90 seconds clears it. And `claimed` is
checked before the code, so this beat needs a runner nobody has scanned and has
to precede that runner's own claim, or the screen reads claimed instead of
expired.

Beat 6: one laptop offline is two desks offline, because each Chrome profile
keeps its own queue, and sending is a tap on the claims page rather than
something that happens on its own.

Both carry the sentence to narrate, since both are easy to overclaim: beat 5
does not refuse a fresh screenshot, and beat 6 does not prevent a duplicate
hand-over. What they prove is staleness in under a minute, and one claim kept
with the losing desk told whom to chase.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
…er out

The setup step said "dated today, price low" and left the rest to the shoot day,
which hides a trap. A close date of today or earlier makes reserve_slot refuse
with RegistrationClosed(20) whatever the event's status, so a race dated today
with entries closed today cannot be entered by anyone, its own organiser
included. Leaving the field empty skips the wizard step and one signature.

The poster line was wrong as advice: render.sh needs macOS sips and hardcodes
the four seeded races, so it cannot make a poster for a new one. Any 16:9 image
does, and the field is optional.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
The demo race is created by hand rather than seeded, so it would sit on the
directory beside four races wearing posters and have none itself, which is the
first frame of the video.

Drawn in the same template as the other four so the family holds: same grid,
same chips, same "Stellar testnet demo race" tag, a place that exists and a race
that does not. No veil layer over the copy column, unlike the Solo poster --
here it would have to cross the ground band and the seam is visible in the
render, so the glow is pulled clear of the copy instead.

render.sh gains the fifth line, and a note on how this one was rendered: it
needs macOS sips, and this was made on Windows with headless Chrome writing the
JPEG directly.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
The setup step named four values and left the rest to the shoot day, which is
where a wizard with a dozen fields turns into guessing. This is every field
/org/new asks for, in order, with the value to type.

Three choices in it are not arbitrary. All five runners enter the 10K, so every
bib sits in one list and the results file needs no category_id and can never
come back ambiguous_bib. One add-on exists to carry the sentence in beat 2 that
nothing else carries, that the entry and the tumbler are one payment in one
transaction. And the description ends with a paragraph saying this is a
demonstration and nobody is expected at the start line, because a race page
that reads as real on a public directory is the one thing here that could
mislead somebody.

The close-date trap moves into the brief, next to the field it ruins.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
Two paragraphs was longer than anything on the directory. The four seeded races
run to three sentences: the route, the distances, and TESTNET_LINE. This
matches, so the demo race reads like its neighbours rather than like the one
somebody wrote by hand.

Also says plainly not to put placeholder text here. The field is on screen in
beat 1, and it carries the sentence that keeps the page honest for anyone who
finds it without the narration attached.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
…pped it

The SOW names the loop as create event, enter and pay, scan QR, claim race pack,
record finish, verify on the public profile, plus both fraud attempts. The shot
list opened on the directory, so the first item was missing and nothing in the
three minutes would have shown a race being published at all.

It is now beat 0 at twelve seconds, the one beat allowed to be a montage, and
the other eight give up two to four seconds each. The table still totals 3:00.

Beat 1 also still pointed at Kota Tua, from when the plan was to film on a
seeded race. It opens on the race being created instead, which makes beats 0
through 8 one continuous race.

Creating the race happens exactly once, so the setup list now says to record the
screen while doing it rather than staging it again later. The entries too: a
spare take of beat 2 costs nothing to capture and cannot be recreated once the
runners have entered.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
…edited

The brief reused the seeded races' four sentences, which are written for races
that are never scanned and never disputed. This race is filmed being both, so
the terms say what the desk will actually do and what the results actually mean.

Every line is checkable against the product rather than borrowed from a
template: no transfer because the contract exports none, one collection because
claim_racepack refuses a record that is not Entered, no refunds because Sterun
never holds the money, and a finish with no official time recorded as that
rather than as zero.

The warning is new and belongs before the field, not after it: terms go into the
event document, the document's fingerprint is the metadata_hash on chain, and a
typo is therefore permanent.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
The setup list said "enter the race and pay" as though it were one action. It
is a purchase with two wallet prompts that do different things: the first signs
the personal details into the off-chain vault, the second is the atomic payment.
Somebody approving the first while expecting the second has no idea what they
just signed.

It also names the race and the distance, which had been left implicit: her own
race, the 10K, all five runners in one distance.

And a warning that had nowhere to live until this step existed. These forms take
a real identity document number, the entry is filmed, and the value reaches a
real vault. Invented runners cost nothing and the seeded races already invent
theirs.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
…t was entered on a laptop

Nine fields on a phone keyboard four times over is the slowest part of the
setup, and it is avoidable. A pass entered on a laptop has to be fetched onto
the phone through /pass/<tokenId> regardless, so entering the rest on the laptop
costs nothing extra and one import of the recovery phrase brings all five
accounts to the phone at once. With the condition attached: only for a wallet
that holds nothing but testnet demo accounts.

The four runners are written out so the details are not invented at the keyboard
with the recording running, and each one is given its job so the wrong runner is
not spent on the wrong beat. Runner 4 in particular must never claim, or the
results file loses its not_claimed row.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
…it asks for

EVIDENCE.md is organised by SOW section 6.1, which is the right shape for
closing gaps and the wrong shape for the form. The submission wants project and
team, scope of work, evidence of completion, a verification checklist and a
coded list of links, so this is that, drawn from the same facts.

The checklist says what is true: deliverables 1 and 2 present, 3 partial, with
the demo video named as the single outstanding item rather than left blank.

It lives in the repository rather than in a hosted document because the
harness account here is not the one the frontend owner signs in with, so a
document created there is one she cannot open.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
…ubmission

The script is a working document for the shoot, not evidence of anything, and
pointing a reviewer at it advertises the one gap rather than closing it. L20 now
carries the recording in Drive instead.

The status cells are untouched: deliverable 3 still reads partial with the
three-minute video outstanding, because what is in that folder has not been
confirmed as the final cut. A link in the supporting list is inventory; a status
is a claim, and that one flips only when it is true.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
@vercel

vercel Bot commented Sep 28, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
sterun-app Ready Ready Preview Sep 28, 2026 1:17pm UTC
sterun-fe Ready Ready Preview Sep 28, 2026 1:17pm UTC
sterun-landing-page Ready Ready Preview Sep 28, 2026 1:17pm UTC

The evidence page had carried "not yet" against the demo video since it was
written, and the §6.2 checklist a warning beside deliverable 3. Both now say
what is true, and the row links the recording rather than describing a plan.

The screenshots row read "they come with the video", which was a promise while
the video did not exist. The pass, the desk, the signed-in console and the
results review still cannot be reached by a script; they are in the video, and
the row says that instead.

The final report follows: status Complete, the checklist row Present, and the
recording added to deliverable 3's proof. Both documents state the same facts
and change together.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
L21 carries the post on @sterunxyz, and deliverable 3's proof lists it beside
the Drive copy: a reviewer can watch it without a Drive link and without asking
anyone for access, which the folder link alone does not offer.

In the Word build, a cell that holds several links is now a labelled list, one
per line. The URL is the visible text there, so three of them separated by a dot
said nothing about which was which.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU
It listed the pass, the desk, the signed-in console and the results review as
things a script could not reach, which read as a gap. The SOW asks for
screenshots of the directory being browsed and of runner profiles, and both have
been there since the shots landed. The row now says what it has instead of what
it lacks.

The rehearsal note was written while the video did not exist and ended in the
future tense. Those six manual steps are still exactly what a script cannot
drive; the difference is that a person has now done them on camera.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PkgGbz4eNSdzzzAPiqV1ZU

This branch was successfully deployed

3 active deployments
Preview – sterun-landing-page — d7e58f91 Deployed Sep 28, 2026 by vercel[bot]
Preview – sterun-fe — d7e58f91 Deployed Sep 28, 2026 by vercel[bot]
Preview – sterun-app — d7e58f91 Deployed Sep 28, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant