# Bot detection — implementation plan

How we catch players who automate **research infrastructure** and **excavation
digs**, with evidence strong enough to punish.

This is a plan, not a shipped feature. Game-side signals and traps live on two
pages; everything an admin sees lives in the modern panel (`#/anticheat`,
`#/turnlogs`, a new `#/botwatch`). Legacy `edmin/` is out of scope.

**This file must not be web-readable.** See §6.0. Today the webroot serves
every `.md` and `.sql` under `app/`, this one included.

---

## 1. The problem we are actually solving

Turns accrue passively from `user.tick`. On Real-Time they arrive about every
5 seconds and the bank caps after roughly 5–7 minutes (60 turns, 90 if
premium). A human who walks away hits the cap and wastes generation. A script
does not.

Overflow turns (2026-10-23) soften this but don't remove it. After 30 minutes
at the cap a human earns 1 overflow turn per 30s, up to 360 more, so they still
waste five-sixths of what an always-present script collects. What changes is
the shape: a single spend of up to 450 turns on return is now normal human
play. `Detailed_Turns.OverflowTurns` records how much of each spend came from
the overflow bank, so a scorer can discount those bursts.

Not every action is worth scripting. The two that are:

| Action | File | POST field | Why it is worth a bot |
|---|---|---|---|
| Research infrastructure | `f_com_research2.cfm` | `turns` | Long grind. Burn the bank, wait for cap, repeat. One of five types (`rtype=1..5`). |
| Excavation dig | `f_com_col_find.cfm` | `turn` | Time-of-day bands. Rare / UnCommon / Mixed batches are minutes long and repeat four times a day. A human sleeps; a script fires every band. |

The obvious combined bot researches outside the rich bands and digs inside
them. It looks continuously present and never wastes a turn. Plan for it.

Explore, attack, ship build, colony upgrade, and the rest are out of this
plan. Do not put traps on them.

The bar for a ban is **certainty**. Behavioural scores are for *finding*
suspects. Punishment needs a signal the player could not have produced by
playing the game: a trap hit, or a spend request no browser click could have
made.

---

## 2. What already exists

### Game side

- `s_endturn.cfm` writes `gcc_log.Stat_Turns` (daily aggregate) on every
  spend, and `gcc_log.Detailed_Turns` (one row per spend, `Time`
  timestamp(3), `Page` = query string).
- `user.DetailedLogging` was set to 1 for every account by a one-off UPDATE.
  The code still gates on it: a per-spend `SELECT DetailedLogging` at the top
  of `s_endturn.cfm` and a `<cfif>` around the insert. `p_signup.cfm` does not
  set the column, so new accounts depend on the production column default.
  `p_login.cfm` and `i_f_800.cfm` only copy it into session.
- `Detailed_Turns` schema problems (per the dev seed, which mirrors prod):
  - Primary key `(ID, UserID)` and nothing else. Every `WHERE Time >= …` and
    `WHERE UserID = …` in `TurnLogs.cfc` is a full scan.
  - `Time … ON UPDATE current_timestamp(3)`: any UPDATE to a row rewrites
    its timestamp. Unacceptable for evidence.
  - `utf32` charset. `Page` is an ASCII query string; this is most of the
    182 MB.
  - No server column (resolved through `user.server`, as `TurnLogs.cfc`
    already does). `Time` is the DB clock; the dig schedule runs on Lucee's
    `now()`. `TurnLogs.gameClockOffset()` reconciles them.
- The insert interpolates `"#cgi.query_string#"` unparameterised in a
  double-quoted literal. Lucee does not escape `"`. Tomcat probably rejects a
  raw `"` in the request line, but this must be `cfqueryparam`.
- Both target pages have the `"Wait!"` dance (randomised JS function name,
  `type=button` + `onclick` → `obj.form.submit()`) and `session.buttoncheck`
  (= `url.ranURL`). A script that scrapes the form and POSTs
  `turns`/`turn` + `buttoncheck` never touches either.
- Both pages emit `AntiScriptFlag`: random empty `<a href=.></a>` tags. It
  changes nothing a script keys on. Delete it in Phase 3.
- Dig has stable selectors: `id="dig"`, `id="gc-dig-turns"`,
  `.gc-dig-control__submit`, `.gc-dig-control__turns`. Research has stable
  visible text `"Start Research !"`. Research renders **two** spend forms
  (`.mobileHide` / `.mobileShow`) on every response.
- Things on the dig page that depend on today's ids and must follow any
  rotation: the label's `for="gc-dig-turns"`, the Luck form's
  `onsubmit` (`getElementById('gc-dig-turns')`), and the Luck select's
  `onchange` (`.gc-dig-control__form [name=keepluck]`).
- Both pages are in `NATIVE_PAGES` in `JS/gc-htmx.js`: their forms submit as
  real browser POSTs, never htmx XHR. Request headers on these POSTs are
  therefore real navigation headers (§6.2 depends on this).
- IP: production is behind Cloudflare, and the server layer (outside this
  repo) restores the visitor IP, so `cgi.REMOTE_ADDR` is the real client on
  every request. Login, `AntiCheat.cfc` and `Base.clientIp()` all rely on
  that; new code does the same.

### Admin panel

- `#/anticheat` — shared IP / cookie, top turn burners, inter-account attacks,
  IP geo. Multi-account, not botting. ACL `anticheat` = level 3.
- `#/turnlogs` — honest reporting from `Detailed_Turns`. Explicitly does
  **not** score behaviour. ACL `turnlogs` = level 6.
- Player page shows 24h/48h/7d/30d spend vs generated and an action timeline
  labelled with the dig band each row fell in.
- Actions: freeze (2), suspend (4), disable turns (4), comment (2). No watch
  flag, no trap log. `gcc_admin.admin_note` and `admin_setting` exist.

### The gap

No record of *how* a spend request arrived (headers, UA, IP). No trap. No
scoring that fits a band-sniping bot. No dossier.

---

## 3. Principles

1. **Passive before active.** Header capture (§6.2) changes no markup, has
   nothing to leak or rotate, and cannot affect a Classic player. Ship it
   before any trap.
2. **Traps never fire on a human playing the page.** Calibrate first: log
   hits, do not auto-punish.
3. **Do not auto-ban from scores.** Scores are leads.
4. **Classic theme is sacred.** No visible chrome, no layout shift, no extra
   click.
5. **Rotating selectors are a moving target.** Names and ids live in
   settings so a leak is a setting change, not a deploy.
6. **Evidence is exportable.** One dossier an admin can paste.
7. **Canaries that hurt the economy are last.**

---

## 4. What we will build (and what we will not)

### Build

| Piece | Where | Why |
|---|---|---|
| Unconditional `Detailed_Turns` + schema fix | `s_endturn.cfm`, migration | Indexed, immutable, kept forever. |
| Spend-request signal log | `f_com_research2.cfm`, `f_com_col_find.cfm` POST paths | Headers no human click lacks and no script can fake from in-page JS. |
| Trap fields + rotated real selectors | Same two pages | Catches in-page scripts and form-scraping HTTP bots. |
| `gcc_log.bot_trap_hit` | New table | One row per trap fire. |
| `#/botwatch` | New Admin view + component | Suspect queue, signal/trap inbox, dossier. |
| Watch flag | `gcc_admin` | "We are looking" without a player-facing tell. |
| Band attendance + max gap scoring | Admin component | Lead generation that fits both bot shapes. |

### Deliberately skip (this round)

- Sitewide per-request log of every `url.f`.
- Client JS beacon (`mousemove`, `navigator.webdriver`). Stealth headless
  beats it; HTTP bots never load it. §6.2 covers the same ground passively.
- CSRF tokens on every form. A rotating token is a rate limit, not a
  confession.
- Auto-suspend on any signal. Calibration first.
- Cloudflare Bot Fight Mode / challenges on `i.cfm`. They *block* rather
  than log (the bot author learns immediately), and challenges land on
  humans. Cloudflare is for §6.0, not detection.
- Income-zero canaries, fake "turns disabled", silent 2× cost.

---

## 5. Threat model on these two pages

Three piles. The decoy design only catches two of them; that is fine.

### A. HTTP bots

Python `requests`, curl, AutoHotkey + HTTP. They GET the page, regex out
`buttoncheck`, POST the stable field `turns` / `turn`. Dig bots add a clock.

**Caught by:** §6.2 (no `Sec-Fetch-*` headers at all unless the author copied
a browser's, and a copied `Sec-Fetch-User: ?1` on every request at machine
cadence is its own tell), and §6.3 (they POST the legacy field name).

### B. In-page scripts — the main decoy target

Tampermonkey / Greasemonkey userscripts, browser extensions, a loop pasted in
the console, iMacros. Typically
`setInterval(() => document.getElementById('dig').click(), …)` or
`document.querySelector("input[value='Start Research !']").click()`.

A programmatic `.click()` ignores visibility, opacity and `pointer-events`,
so it fires the hidden decoy. It also carries **no user activation**, so the
resulting navigation has no `Sec-Fetch-User: ?1` — and page JS cannot set
that header.

**Caught by:** §6.3 (decoy) and §6.2 (missing `Sec-Fetch-User`), independently.

### C. WebDriver (Selenium / Playwright / Puppeteer)

They locate an element and send a synthetic native click. Against our decoy
(off-screen, `opacity:0`) Selenium raises *element not interactable* and
Playwright waits for actionability and times out: the bot **breaks** rather
than trips, and its author re-targets the visible button. A WebDriver bot
that clicks the real button produces real user activation and real headers.

**Caught by:** behaviour only (§7). Accept this. WebDriver bots are
expensive, fragile, and rare in browser-game botting. If a pile-C bot
re-targets after the decoy breaks it, that re-target is visible as a run of
GETs with no POST followed by normal spending — interesting colour, not
proof.

---

## 6. Game-side design

### 6.0 Stop serving docs (do this first, independent of everything else)

Before this, `http://<host>/Admin/docs/BOT-DETECTION.md`, `/Admin/sql/*.sql`,
`/AGENTS.md`, `/api/docs/*.md` and source backups (`application.cfm-`,
`*.bak`) all returned 200 with no login: `Admin/Application.cfc` only guards
CFML requests; static files bypass it.

- Cloudflare WAF custom rule, action **Block** (`ends_with` works on every
  plan; regex `matches` does not):

  ```
  ends_with(lower(http.request.uri.path), ".md") or
  ends_with(lower(http.request.uri.path), ".sql") or
  ends_with(lower(http.request.uri.path), ".bak") or
  ends_with(lower(http.request.uri.path), ".cfm-")
  ```
- Origin: `app/.htaccess` denies the same extensions. It only works if the
  Apache vhost allows overrides (`AllowOverride` covering `AuthConfig` /
  `All`); otherwise put the same `FilesMatch` in the vhost.
- Verify against prod after both land: each of those URLs must return 403.

### 6.1 Detailed turns: unconditional, indexed, kept

`s_endturn.cfm`:

- Delete the `getLogging` query and `session.DetailedLogging` assignment
  (lines 16–19) and the `<cfif>` around the insert. Logging no longer depends
  on the flag, so new accounts are covered regardless of the column default.
- Parameterise the insert (`UserID`, `Username`, `Turns`, `Page`) as its own
  script `Query` after the legacy `updatetick` query, so a row is only
  written once the spend has been applied.
- `DetailedLogging` stays in the `p_login.cfm` / `i_f_800.cfm` selects
  (harmless). Nothing reads it any more, so its column default no longer
  matters.

Migration `Admin/sql/detailed_turns_indexes.sql`: two online index builds,
dropping `ON UPDATE` from `Time`, then the utf8mb4 conversion. The last step
copies the table and blocks spends while it runs, so run it off-peak.

**Retention: forever.** At ~182 MB for several months in utf32, utf8mb4 plus
two indexes lands well under 1 GB/year. Long-term turn statistics for every
player are a feature. If it ever matters, partition by year on `Time`; do
not prune. `Stat_Turns` stays as the cheap daily rollup for long-range
charts.

### 6.2 Spend-request signal log (passive, Phase 0)

On every POST to the two spend paths — **including** ones that fail
`buttoncheck` — write one row to `gcc_log.spend_signal`. Do it in a tiny
shared include (`s_spend_signal.cfm`) called near the top of each page's POST
branch, reading `getHttpRequestData().headers`.

| Column | Point |
|---|---|
| `id`, `userid`, `at` timestamp(3) | Who, when (no `ON UPDATE`) |
| `action` | `research2` / `col_find` |
| `outcome` | `spent` / `bad_buttoncheck` / `no_turns` / `trap` |
| `turns_asked` | As typed |
| `sfs`, `sfm`, `sfu`, `sfd` | `Sec-Fetch-Site/Mode/User/Dest`, raw, short varchar, NULL if absent |
| `has_origin`, `has_referer` | Booleans |
| `ua` | `User-Agent`, truncated to 255 |
| `ip` | `cgi.REMOTE_ADDR` (already the restored client IP) |
| `cf_country` | `CF-IPCountry` (free with Cloudflare) |

What a real click looks like on a current browser (Chrome/Edge 80+,
Firefox 90+, Safari 16.4+): `sfs=same-origin`, `sfm=navigate`, `sfu=?1`,
`sfd=document` (or `iframe` — confirm which the game frame produces during
calibration). Enter in the turns box is also user activation → `?1`.

Readings:

| Pattern | Meaning |
|---|---|
| All four `Sec-Fetch-*` absent, modern UA | HTTP client wearing a browser UA. Strong. |
| `sfm=navigate`, `sfu` absent, modern UA | Navigation without a user gesture: script `.click()` / `.submit()` from a timer. Strong. |
| `sfu=?1` but posts every ~5 min around the clock | Headers copied into an HTTP bot, or WebDriver. Behavioural. |
| Absent, old UA | Legacy browser. Ignore; baseline per account. |

Score per account, not per row: "92% of 400 spends in 7 days lacked
`Sec-Fetch-User`, on a Chrome 141 UA, while the same account's other
sessions send it" is the confession. One odd row is noise.

Volume equals the `Detailed_Turns` rows for these two pages plus failed
posts. Keep forever.

### 6.3 Trap helper

One include, `s_bot_trap.cfm`, with `action = "research2" | "col_find"`:

1. On GET, mint `session.botReal_<action>` (random alphanumeric, never
   `turn`/`turns`/`submit`) and `session.botDecoy_<action>`.
2. On POST: a legacy field name (`turn` / `turns`) or the decoy name, with
   the real rotating name absent → trap hit — **only if
   `session.botReal_<action>` exists**. A tab loaded before the deploy never
   minted one, so deploy day produces no false hits. HTTP bots GET the new
   page before posting, so they still trip.
3. A *stale* rotating name (second tab overwrote the session value) is not a
   trap. `buttoncheck` already rejects that post exactly as it does today.
4. Calibration (`bot_trap_calibrate`, default on): still process the spend.
   Map the posted value back to `form.turns` / `form.turn` so the rest of
   the template is unchanged.
5. Write `bot_trap_hit` and set `outcome='trap'` on the §6.2 row.

### 6.4 Markup rules

Traps go on spend forms only (both research forms; dig control form). Luck
is not botted.

Real control:

- Turns box: `name` and `id` rotating; class, label text, size, defaults
  (`9999` / `99`), `maxlength` unchanged.
- Submit: `type=button` + existing Wait! `onclick`; `name`/`id` rotating;
  visible value unchanged.
- Dig: update the label's `for`, the Luck form `onsubmit`, and keep the
  `gc-dig-control__form` class on the **real** form only, so the keepluck
  `onchange` selector still finds the right field.

Decoy — a **sibling form**, earlier in the DOM than the real one so
`getElementById`, first-match `querySelector` and first-XPath all land on it:

- Same `action` as the real form.
- `input type=text name=turns|turn` (the legacy name) and a button with
  `id="dig"` (dig) / value `Start Research !` (research), plus the same
  `buttoncheck` hidden field so a scraper takes it for the real thing.
- Button `onclick` submits the decoy form (so a script `.click()` posts).
- Hidden by CSS, not `display:none`: `position:absolute; left:-10000px;
  width:1px; height:1px; overflow:hidden; opacity:0; pointer-events:none`,
  `tabindex=-1` on every control, `aria-hidden=true` on the form,
  `autocomplete=off` on the text input.
- No `gc-dig-*` classes on the decoy (theme CSS and the Luck JS must not
  see it).
- Inline styles in the include; no `theme.css` change.

Research: one decoy before the `.mobileHide` block covers both real forms.

### 6.5 Rotation

Decoy ids, decoy button values and legacy names live in
`gcc_admin.admin_setting` (repo is private, so defaults may also sit in the
include; settings win). Start with today's public selectors as the honey:
dig `id=dig`, research value `Start Research !`, legacy `turn` / `turns`.
Rotate only after hits stop landing; keep retired decoys live for a week.

### 6.6 What a trap hit records

`gcc_log.bot_trap_hit`: `id`, `userid`, `server`, `nic`, `at` timestamp(3),
`action`, `kind` (`legacy_field` / `decoy_name` / `decoy_click`), `page`,
`spend_signal_id` (FK to §6.2 row, which already has headers/UA/IP),
`processed`, `turns_asked`, `fields` (posted field *names* only, for
calibration review).

Never auto-suspend from this insert.

### 6.7 Do not touch `i.cfm`

§6.2 and §6.3 cover the two spend paths. A future "loaded the dig page 4,000
times, posted 0" signal is a narrow GET counter in `f_com_col_find.cfm`.

---

## 7. Behavioural layer (leads, not verdicts)

Compute per player, per window (7d / 30d), RT and TB separately, on the
**game clock** (`gameClockOffset()`), from `Detailed_Turns` + `spend_signal`.

| Signal | How | Catches |
|---|---|---|
| **Band attendance** | Of the Rare / Mixed / UnCommon bands in the window (4 per 6h cycle, 16/day), share with ≥1 dig. Bands: `Base.digBands()`. | Band snipers. No human makes the 00:30 and 06:30 Rare bands every day for a week. |
| **Max gap** | Largest delta between consecutive spends (any action). | Cap farmers and the research+dig combo. |
| **Presence** | Share of hour-of-day buckets with a spend. | Same, coarser; drives the heatmap. |
| **Off-band dig share** | Dig turns in Default vs rich bands, alongside research share. | The combo bot: research off-band, dig on-band. |
| **Header anomaly rate** | §6.2 share of spends missing `Sec-Fetch-User` / all `Sec-Fetch-*`. | Piles A and B. Close to proof on its own (§10). |
| **Cadence** | Spend interval clustered at the cap refill time. | Cap farmers. Weak alone. |
| **Ritual** | Same turn count every time. | Weak alone; humans leave the default too. |

Note the shapes: a pure band sniper has a daily max gap of ~3h (21:20 →
00:30), so **max gap must not be the entry gate**. Queue entry is any of:

- band attendance ≥ 80% over 7d (tunable), or
- RT max gap < 45 min over 7d (TB: tune separately; turnmin is 150s), or
- header anomaly rate ≥ 50% over ≥ 50 spends,

plus a minimum spend floor. Publish the formula in the UI.

`Stat_Turns` remains the long-range daily series (and covers accounts from
before logging was universal).

---

## 8. Admin panel

### 8.1 `#/botwatch`

ACL: queue and inbox under `anticheat` (3). Moderation buttons keep their
existing per-action levels.

Tabs:

1. **Suspects** — §7 scores; sortable; formula shown.
2. **Signals** — accounts with §6.2 header anomalies, grouped by user, with
   the UA breakdown.
3. **Trap hits** — `bot_trap_hit` grouped by user.
4. **Watched**.

Rows link IP → shared-IP scan, player → turn-log timeline.

### 8.2 Dossier

`botwatch.dossier(uid)`:

- Identity, server, latest UA / IP / country from `spend_signal`
- Header summary (counts per `sfs/sfm/sfu` combination, per UA)
- Trap hits
- Band attendance grid (days × bands), presence heatmap, max gap, action mix,
  7d / 30d
- Shared IP / cookie summary (existing `sharedIp` guts)
- Freeze / disable turns / suspend via existing `UserActions`, canned reason
  "Automated play (research / excavation)", required note
- Copy box (tab-delimited / JSON)

### 8.3 Watch flag

`admin_note` with type `watch`, or `admin_watch (userid, admin_id, since,
reason)`. Staff-only badge on the player page.

### 8.4 Settings (level 9)

`bot_signal_enabled`, `bot_trap_enabled`, `bot_trap_calibrate` (default
true), decoy / legacy name lists, §7 thresholds.

### 8.5 Do not fork Turn Logs

`#/turnlogs` stays plain reporting. It benefits from the §6.1 indexes for
free. A long-term per-player turn history chart belongs there, not in
`#/botwatch`.

---

## 9. Canaries (later, optional)

Post-trap only, on the page the bot is already on: nonsense research
completion text; Default odds behind a Rare label for that one user; a staff
PM "reply YES". Per-user override in the two templates. Only if trap or
signal evidence is contested.

---

## 10. False positives we must design against

| Source | Mitigation |
|---|---|
| Deploy-day stale tabs posting `turn` | §6.3 step 2: no minted real name → not a trap. |
| Password manager / autofill into decoy | Decoy is a sibling form, `autocomplete=off`, `aria-hidden`, off-screen. |
| Screen readers / keyboard users | `aria-hidden` + `tabindex=-1`; real control keeps its label. |
| Old browsers without `Sec-Fetch-*` | Judge §6.2 per account against its own UA; ignore UAs older than the support table. |
| Privacy extensions stripping headers | Rare; shows as *all* headers missing on *every* page for that user, including normal clicks at human cadence. Ask before acting. |
| Mobile in-app browsers / WebViews | Calibration week: tabulate header patterns by UA family before thresholds go live. |
| Double-submit / retry | `buttoncheck` already blocks replay. |
| Two research forms | One decoy before both; both real forms share the session name. |
| Staff / test accounts | Exclude `adminflag` and known test ids. |

Calibration: two weeks of signals + traps logged with posted field names.
Review the header table by UA family and every trap hit before any
punishment.

---

## 11. Phased delivery

### Phase 0 — hygiene and passive data (no player-visible change)

Built 2026-10-20; deploy order below.

1. §6.0: `app/.htaccess` in the repo; the Cloudflare rule is manual.
2. §6.1: unconditional, parameterised `Detailed_Turns` insert;
   `detailed_turns_indexes.sql`.
3. §6.2: `spend_signal.sql` + `s_spend_signal.cfm`, called once per spend
   POST from both pages with the outcome (`spent` / `bad_buttoncheck` /
   `no_turns` / `other`).
4. `#/botwatch` (ACL `anticheat`): Suspects and Request signals tabs,
   `BotWatch.cfc`. Dig bands and the clock offset moved to `Base.cfc`
   (`digBands()`, `gameClockOffset()`), shared with Turn Logs.

Deploy: Cloudflare rule → `spend_signal.sql` → code → index part of
`detailed_turns_indexes.sql` any time → its charset step off-peak.

Exit: staff can sort by band attendance, max gap and header anomaly rate,
and open a timeline.

### Phase 1 — traps, log only

1. `bot_trap_hit` table, `s_bot_trap.cfm`, wire into both pages.
2. Trap hits tab. `bot_trap_calibrate=true`.

Exit: two weeks of hits and header data reviewed. No bans yet.

### Phase 2 — dossiers and watch

1. Dossier + copy box, watch flag, canned reason.
2. Admin "What's New" entry only — never the public changelog.

Exit: inbox → dossier → freeze without edmin or raw SQL.

### Phase 3 — tighten

1. Keep processing spends on trap hit (preferred) or not.
2. Rotate decoys via settings.
3. Canaries for contested cases.
4. Delete `AntiScriptFlag`.

---

## 12. Files we expect to touch

Game: `s_endturn.cfm`, `s_spend_signal.cfm` (new), `s_bot_trap.cfm` (new),
`f_com_research2.cfm` (both forms), `f_com_col_find.cfm` (dig form + Luck
JS), `app/.htaccess`.

Admin: `Admin/sql/detailed_turns_indexes.sql`, `Admin/sql/spend_signal.sql`,
`Admin/sql/bot_trap_hit.sql`, `Admin/api/components/BotWatch.cfc`,
`Base.cfc`, `TurnLogs.cfc`, `api/index.cfm`,
`Admin/assets/views/botwatch.js`, `Admin/assets/app.js`,
`Admin/assets/views/changelog.js`, `Admin/docs/API-REFERENCE.md`,
`Admin/docs/DEVELOPMENT.md`.

Infra: Cloudflare WAF rule (§6.0).

Out of scope: `edmin/*`, other game pages, `i.cfm` request logging, public
`p_changelog.cfm`.

---

## 13. Open questions

1. **Frame destination.** Does the game run inside a frame/iframe in any
   theme? It changes expected `Sec-Fetch-Dest` (`document` vs `iframe`).
   Settle it from calibration data rather than guessing.
2. **Thresholds.** Band attendance 80%, max gap 45 min, header anomaly 50% —
   all placeholders until calibration.
3. **Who may punish.** Inbox at anticheat (3); actions at their existing
   levels.

---

## 14. Why this is enough to be sure

A human playing research or dig:

- Sees one turns box, one submit, no extra click
- Sends `Sec-Fetch-User: ?1` because they clicked or pressed Enter
- Never posts a field from a form they cannot see
- Misses most night bands and sleeps

A script against today's pages:

- HTTP: posts `turn` / `turns`, usually without `Sec-Fetch-*`
- In-page: `.click()`s `#dig` / `"Start Research !"` — hits the decoy, and
  the navigation carries no user activation
- Either way, attends every rich band or never stops

**Trap hit or header confession, plus inhuman presence** is the case.
Presence alone is a lead. A confession during an otherwise human-looking day
is still a confession: punish on it, cite the presence in the note.
