Jivika — QA Regression Test Plan

1. On a fresh number type ssc gd form kaise bharein.

The bot starts the questions (asks the date of birth). Nothing is stored; no error.

"Samajh nahi aaya"; a stall; an answer to the question as if it were a form field.

1. Fresh number. Type Haan, shuru karein exactly instead of tapping.
2. On another fresh number type haan.

Both start the questions.

Either is treated as a wrong answer to a question.

1. Rohit: get to the ask state question. Type hi.

The greeting comes back (fresh), previous answers are not silently used as if confirmed.

hi is read as an answer to the current question; the chat crashes; a second greeting stacked on the first.

1. At the date question send each of: 14-03-2002, 14/03/2002, 14.03.2002, 1-2-2000, meri dob 14 03 2002 hai.

Each is accepted (the next question — education — appears). The card later shows 14-03-2002 / 01-02-2000 in DD-MM-YYYY.

Any of them re-asks; a date read as month/day swapped.

1. Try each: 31-02-2001, 29-02-2001, 14-13-2002, 14-03-02, 14-03-2030, kal.

Each gets the re-ask "Ye tithi samajh nahi aayi. DD-MM-YYYY me likho, jaise 14-03-2002." Nothing advances; no answer is stored (mere form still says nothing has started).

A rollover (31-02 → 3 Mar) is accepted; a two-digit year is guessed; a future date is accepted.

1. At the education question tap Chunein and read the rows.
2. Choose Graduate (BA/BSc/BCom). On another number type 12th pass instead.

A list (not free text) with the six levels (8th … Post-graduate); both routes advance to the category question.

A free-text question; a level missing; "graduate" ranked as "post-graduate" or the reverse.

1. At education send 10th appearing, then final year BA.

Neither is stored as a qualification; the list is offered again.

"10th appearing" saved as 10th passed.

1. At category send sc st, then general obc. Then tap OBC (NCL).

The first two are re-asked (never resolved to one of them); the tap works.

"sc st" silently saved as SC.

1. At state send in turn (fresh each time): UP, Uttar Pradesh, bihar ka, Gujrat, Orissa.

Each resolves (UP, UP, Bihar, Gujarat, Odisha) and the card shows the full state name.

bihar ka read as Karnataka ("ka"); any of them re-asked.

1. At state send up bihar, then mars.

"Ye rajya nahi mila. Rajya ka naam ya code likho, jaise UP ya Bihar." for both; state not stored.

One of the two states picked; "mars" accepted.

1. At state send उत्तर प्रदेश; at education स्नातक; at category ओबीसी.

Each is read correctly; the bot still replies in Hinglish (romanised), not Devanagari.

Not understood; the bot answers in Devanagari without being asked (known limitation: the reverse — a Devanagari user — is flagged by the Debugger as P2 register-mismatch, §15).

1. Walk Rohit's path to the card. Read every bot message so far.

Not one message mentions DigiLocker, Aadhaar, documents, OTP, or asks for a name or phone.

Any of those words before the eligible list — S1 (CLAUDE.md §3.1).

1. At the date question send kal. Type mere form.

"Abhi koi form shuru nahi hua. 'hi' likh kar shuru karein." — nothing was stored.

A partial profile exists; mere form errors.

1. Fresh number: tap Haan, shuru karein, then send graduate, OBC.
2. Send 14-03-2002.

After 1: "Samajh gayi — padhai, category note kar liya." then the date question. After 2: skips education and category and asks state only.

Education or category asked again; the date question skipped.

1. At the date question send 12-03-2002 graduate.

Date 12-03-2002 and Graduate are both captured — not "12th pass".

Qualification shows 12th.

1. Send BSc OBC UP at any question.

Graduate + OBC (NCL) + Uttar Pradesh. Category is not SC.

SC anywhere.

1. Fresh number, tap start. Send general obc, Bihar.

Only the state (Bihar) is captured; the bot then asks the date, and category is asked later (never guessed from "general obc").

Category filled with UR or OBC.

1. Send graduate hu, up me rehta hu (lowercase) then on another number graduate UP.

The first captures education only (the lowercase "up" is not taken as a state); the second captures education and state UP.

"up" captured from ordinary words; "UP" ignored.

1. Send graduate hu OBC UP lekin mujhe jeene ka mann nahi karta.

The distress path (§8) — acknowledgement with Tele-MANAS 14416, ticket queued — and nothing captured (no card).

The answers are captured and a card shown; no ticket.

1. At the state question send mujhe pata nahi (or any chatter).

The re-ask (state), not "captured" and not a crash; after three such failures a help ticket opens (§9).

A field captured from chatter; silence.

1. Stop at the card. Look at the thread and at the Debugger/dividers.

No ✦ engine: divider and no eligible list yet.

A verdict appears before "Haan, sahi hai".

1. Tap Haan, sahi hai (or type sahi hai).

✦ profile: 2002-03-14 · graduate · OBC-NCL · UP, the eligible list (§5), its source PDF, ✦ engine: …, then the one-time tip.

Confirm ignored; the tip missing or repeated.

1. At the card type state Bihar, 15-03-2002.

State → Bihar and date → 15-03-2002 in the new card; category unchanged.

Only one applied.

1. At the card tap Badlo. 2. Choose category from the list. 3. Tap SC.

A list of four fields (janm tithi, padhai, category, state) → the category question → back on the card with SC (dob, education, state untouched, not re-asked).

All four questions re-asked; the card not shown after.

1. At the card type nahi.

The same four-field list as Badlo.

Treated as an unreadable answer.

1. At the card type pata nahi kya likhu.

One line: "Ye samajh nahi aaya. Seedha likho, jaise 'category SC' ya 'state Bihar' — ya 'Badlo' dabao." and the card again. Nothing stored.

Silence; an answer guessed.

1. Reach the result as Rohit (born 2002). Then hi, repeat with date 05-05-2010, confirm.

The second result is the too young message (§5), not Rohit's old eligible list.

A stale verdict.

1. Reach show_eligible as Rohit, say "Badlo". 2. Reach show_eligible, say "meri dob galat hai, 12-03-2000" (a typed sentence, not the button). 3. Repeat step 1 after tapping a notice (choose_notice), after "Baad me" (notice_saved), and after starting a form but before agreeing to the consent notice (consent). 4. Confirm the corrected card and continue. 5. Agree to the consent notice (reach identity_card), then say "Badlo" there too.

Steps 1–3 all land back on the profile card (§4.1) with the current values — never "no handler for this input" — and step 2's sentence actually changes the date shown on the card. Step 4's "Haan, sahi hai" returns to show_eligible with eligibility recomputed from the corrected value, and any notice already picked before the correction is still there, re-selectable, not duplicated. Step 5 is UNCHANGED and still guide-only: "Badlo" at identity_card goes to a person (§6.18's "Kuch galat hai" case), never back to the profile card — identity fields are never self-service once documents are pulled.

"No handler for this input" anywhere in steps 1–3; a correction at identity_card or later bypasses the guide; a notice picked before the correction silently disappears or gets duplicated after it.

1. Read the document bubble's file name and (in the Debugger Copy report or by inspecting the link) its URL.

notice_01122025.pdf on ssc.gov.in — SSC's own 2026 notice.

A third-party or placeholder link.

1. Date 14-03-2004, 8th pass, General (UR), UP, confirm.

"…kam se kam 10th pass chahiye. Padhai ke baad phir dekhte hain…"

Eligible; a level other than 10th named.

1. Two fresh numbers, General (UR), graduate, UP: dates 02-01-2004 and 01-01-2004.

02-01-2004 eligible; 01-01-2004 over the limit.

Either reversed.

1. Graduate, UP. OBC (NCL) with 02-01-2001 and 01-01-2001; SC with 02-01-1999 and 01-01-1999.

OBC: first eligible, second over (26). SC: first eligible, second over (28). The over-limit message quotes 26 / 28.

A 23-year limit applied to OBC/SC.

1. Sunaina: date 02-08-2003, Graduate, SC, Bihar, confirm.

Eligible (DRAFT). "mere form" later shows the handle SSC GD · draft.

Rejected on age.

1. For each result you have seen compare the message with the ✦ engine: divider.

The message and the verdict divider agree every time (eligible ⇔ a list; not eligible ⇔ a reason).

A list under a "not eligible" verdict, or the reverse — S1 (the Debugger raises eligibility-mismatch P0).

1. Get a result. Then hi, run the path again to a second result.

The tip ("'mere form' likh kar… 'madad' likh kar…") appears after the first result only.

Repeated on the second result; missing on the first.

1. On separate results type apply, apply karna hai, form bharo, ssc gd, SSC GD · draft.

Each opens the same card.

Any of them "no handler"; shuru karo is not one of them (it restarts — §7.6).

1. At the card tap Baad me. 2. Later tap the row again (or apply).

"Theek hai, SSC GD · draft save hai — koi jaldi nahi…" then the same card again; still one application (no duplicate); Haan, shuru karein then works.

A reminder promise; guilt; a second application record.

1. On a "too young" result type apply.

"Draft rules ke hisaab se abhi koi form aapke liye nahi mila…" — no card, no application created.

A card for a form you are not eligible for.

1. At the result type RRB Group D CEN 09/2025.

Not treated as a choice (a "no handler" note; no card).

A card for a form that was never offered.

1. (Not reachable in the console until a second notice has rules — only SSC GD is checked today.) Covered by applications.test.ts; re-run this test manually when RRB Group D is added.

A bare apply answers "Kaun sa form? Neeche se chuno." with a list of the eligible forms.

One picked silently.

1. Start a form, then delete → Haan, hata do. Then mere form.

"Abhi koi form shuru nahi hua…" — the application is gone with the answers.

The form is still listed.

1. On separate runs type haan and nahi at the consent message.

haan agrees (documents pulled); nahi declines.

Either read as unreadable text.

1. After agreeing open the Events panel (or Export JSON).

In order: consent_granted purpose=guided_apply version=v0-draft, then document_pulled type=marksheet_10 why=guided_apply, then document_pulled type=id why=guided_apply, application_status status=docs_pulled, form_card_shown card=identity_education — each with the time (when), the number (who) and the purpose (why). A decline produces only consent_declined. No document value appears in any event.

A document_pulled with no earlier consent_granted; missing events; document contents inside props — S1.

1. On Rohit agree to consent. Switch to Sunaina and mere form / walk her far enough to see any document mention. Switch back and delete → Haan, hata do.

Sunaina never sees or holds Rohit's documents. After the erase Rohit has no documents (a fresh consent is needed to pull them again) and the earlier events are attributed to "erased".

Another profile's document data appears anywhere — S1; documents survive the erase.

1. On the card tap Kuch galat hai (or type nahi). 2. Type koi hai?, then kab tak call aayega.

One help ticket in the Guide queue ("identity card: the candidate says a value is wrong"); "…hamari team se koi aapko 2 ghante ke andar call karega aur saath me check karega." Later messages get a varied "noted" line, never the same words twice, and never a second ticket. Typing a new name is not accepted as a correction.

A typed name accepted; a second ticket; the same reply repeated.

1. (Needs a tampered document — covered by the automated test, which injects a marksheet with a different date of birth; in the console use the unit test or a debugger build.) Expected behaviour to confirm from the test output:

State identity_wrong; no "Sahi hai?" card, no document values in any bubble; ✦ identity card held back — documents disagree with the answers (dob_differs_from_marksheet); a help ticket whose text holds rule ids only; form_card_held card=identity_education n=1.

A card shown while the documents disagree — a wrong prefill, S1; values in the ticket or the event.

1. After 6.18, delete → Haan, hata do. Export the Events JSON.

The documents, answers and applications are gone; the reads that happened stay in the spine but attributed to "erased".

Documents survive the erase.

1. Tap Kuch nahi — aage chalo (or type nahi).

"Note ho gaya ✓ — koi claim nahi." followed by the domicile question (§6.27). Events form_card_confirmed card=claims then form_card_shown card=domicile. No ticket.

A claim left blank rather than an explicit no; a ticket opened.

1. Rohit (OBC): tap Age me chhoot. Repeat as Sunaina (SC).

"Age me chhoot laga di (OBC: 3 saal ki chhoot (code 2))." — for SC "(SC/ST: 5 saal ki chhoot (code 1))". The menu comes back without the age row and with a fresh header ("Aur kuch lagu hota hai?"). Event claim_selected claim=age.

A code that does not match the category; a second question; the age row still offered.

1. Tap NCC certificate, then NCC 'B' certificate, then Age me chhoot, then Kuch nahi — aage chalo.

A three-row list (A, B, C) → "NCC 'B' certificate note kar liya." → the menu (no NCC row) → … → "Note ho gaya ✓ — age chhoot (code 2), NCC B."

The type lost; NCC offered twice; the summary missing a claim.

1. Tap ESM, khel, doosra rajya. 2. Type koi hai?.

One help ticket ("claims: the candidate says ESM, sports, another state or a 1984-riot dependant applies"); "Ye claim abhi guide ke saath bharenge — hamari team se koi aapko 2 ghante ke andar call karega…". Later messages get a varied line and no second ticket. The application is not advanced with these claims left as "no".

The claim recorded as "no"; the flow carrying on to the next card; a second ticket.

1. (Automated test: a form whose age code does not fit the category — e.g. code 1 for an OBC candidate — is finished.)

The card is held (form_card_held card=claims n=1), a ticket carries the rule id (age_relaxation_code_mismatch), no confirmation event.

A mismatched code confirmed — a wrong form value, S1.

1. Type Lucknow. Then (fresh run) type doosra rajya, then Bihar, then Patna.

First run: the exam-centres question appears; events answer_captured field=district via=asked, form_card_confirmed card=domicile, form_card_shown card=centres. Second run: "Kis rajya ka domicile certificate hai?…" → the district question now names Bihar → Patna accepted. mere form and the eligibility answers still say Uttar Pradesh — a guided answer never rewrites the profile.

The profile state changed; the district question still names the old state after "doosra rajya".

1. At the district question type haan, then ??, then 12345.

"Zile ka naam English letters me likho, jaise Lucknow." → a different "Bas zile ka naam, jaise Kanpur. Dikkat ho to 'madad' likho." → a help ticket on the third failure. haan is not stored as a district.

"HAAN" saved as a district; the same words twice; no ticket.

1. At the district question type meri dob 15-03-2002 hai, graduate hu, SC, Bihar.

The profile answers are unchanged (date 14-03-2002, OBC (NCL), UP); the district question is asked again.

Any profile field changed by a guided-card answer.

1. On separate runs: tap English; tap Hindi; tap Koi aur bhasha then type marathi; then french.

English/Hindi go straight to the force-preference question (§6.35); the other opens "Bhasha ka naam likho, jaise Tamil ya Marathi." and marathi is stored as Marathi; french → "Ye bhasha exam me nahi mili…". The list has three rows.

A language outside the notice's 15 accepted.

1. Read the question. Answer Haan, de sakte ho on one run and nahi on another.

"Ek sawaal: kya sarkar ko aapki jaankari (naam, category, padhai) doosre sarkari naukri ke mauke ke liye dene ki ijazat hai? Ye aapki marzi hai — 'nahi' bhi theek hai." with two buttons. Nothing is recorded until you answer; yes and no both proceed (to the signature card, §6.38); both are stored as an explicit true/false.

Pre-answered "yes"; pressure toward yes; "no" blocks progress — S1 (a consent must be free).

1. (Automated: a full path is collected, then each card is validated.)

Domicile, medium and consent_jobs cards validate with no errors; a nonexistent state code is caught (not_an_option).

A value the form's rules reject sitting in the collected form.

1. Read the centres and forces questions.

Both end with "Baad me badla nahi ja sakta." (the form allows no change after submission — Notice ¶11.2, ¶14.12). Changes are possible only at the review card (built next).

The warning missing; the picks presented as casual.

1. Export the Events JSON after 6.34–6.35.

answer_captured field=centres via=asked and field=forces via=asked, form_card_confirmed card=centres|forces — no city names or force codes anywhere in props.

A city or force named inside an event — S1.

1. Send jpg 8 KB 709x236, jpg 9 KB 709x236, png 15 KB 709x236.

The second reply is a different message ("Signature ek saaf, seedhe box me…"); the third failure opens a help ticket.

The same words twice; no ticket.

1. After 6.39 export the Events JSON and search for 709, 236, 15360, jpeg.

Not found: only field=signature and card=signature.

File attributes inside an event.

1. Badlo → Exam centres → type Lucknow, Lucknow, Kanpur (repeat), then Agra, Gwalior, Bhopal. 2. Badlo → Force order → 7 6 5 4 3 2 1.

The repeat is re-asked exactly as at 6.34; the good answer updates the card ("AGRA → GWALIOR → BHOPAL"); the reversed force order updates the card with the new names in order.

A bad centre list accepted from the review path but not from the first-time path (or vice versa) — the same rule must apply both times.

1. Badlo → Exam bhasha → English (tap). 2. Badlo → Exam bhasha → Koi aur bhasha → type tamil. 3. Badlo → Job-info sharing → Nahi.

Each returns straight to the card with the new value; the consent change is stored as an explicit false, not blank.

A stale value shown after a change.

1. At the review card type naam Rohit Sharma hai.

The card is unaffected (not read as a correction); the name is unchanged. To fix a wrong name, the candidate must have used "Kuch galat hai" back at the identity card (§6.18) — the review card cannot do this.

The name silently changes from typed text — S1 (identity must match the certificate; a typed "fix" here is unverified).

1. At the review card type pata nahi kya likhu. 2. At the change menu (Badlo) type kuch bhi.

"Ye samajh nahi aaya. 'Haan, sahi hai' dabao, ya 'Badlo'." then the card again; "Ye samajh nahi aaya. Neeche list se ek chuno." then the menu again. The Debugger's repeat-verbatim rule stays quiet no matter how many times you retry.

The exact same bubble twice in a row.

1. At the review card send mujhe jeena nahi hai. 2. On a fresh run, delete → Nahi, rehne do. 3. Change the domicile, then delete → Haan, hata do.

Distress routes as always (§8). Declining a delete returns to the review card with every value intact. Confirming a delete clears the form and the documents together.

Any of the review's values surviving an erase; distress not routed.

1. Walk the path as Sunaina (F, SC) to the packet.

"Fee: ₹0 (exempt)" — feeFor on the field map, per Notice ¶9.2.

₹100 charged to an exempt category.

1. Type abc123xy. On a fresh run, type 12345678 .

Stored as ABC123XY (uppercased) and 12345678 (trimmed) respectively.

Case or whitespace kept verbatim in a way that would make two people's numbers look different when they're the same.

1. At the packet send mujhe jeena nahi hai. 2. On a fresh run, confirm with a number, then delete → Haan, hata do.

Distress routes as always. The delete clears the application, the form and the documents — confirmation number included. The whole path (review → packet → confirmed) leaves the Debugger silent.

Any leftover value after the erase; a finding anywhere on the happy path.

1. Type madad again.

"Aapki request mil chuki hai…" (and a different wording a third time). Still one open help ticket.

A second ticket; the same words verbatim.

1. Fresh number: mere form.

"Abhi koi form shuru nahi hua. 'hi' likh kar shuru karein." A second time: "Kuch badla nahi — abhi koi form shuru nahi hua."

An error; the same words twice.

1. Mid-form (say at ask category) type delete, then Nahi, rehne do.

"Theek hai, kuch nahi hataya." and the category question again; answers intact.

Sent back to the start; answers lost.

1. delete → Haan, hata do. Then mere form. Then type hi.

"Ho gaya — aapki jaankari hata di gayi…". ✦ erased: dob, qualification, category, state (lists what went). mere form says nothing started. The Events panel shows erasure_requested and erasure_completed, and all earlier events for that number are now attributed to "erased" (Export JSON: no phone number). hi starts fresh; the tip can be taught again.

Anything of the old profile survives; the phone number remains on old events.

1. At the result type namaste.

The greeting again; a fresh start.

Read as an answer.

1. At the date question send mujhe madad chahiye kyunki main bahut confused hoon aur yeh form samajh nahi aa raha.

A re-ask/normal handling (long text is not swallowed by the keyword madad), no ticket.

A ticket opened from a long sentence that merely contains "madad".

1. At the card type ok, then (fresh) theek hai, bilkul. At a card type galat.

ok/theek hai/bilkul confirm; galat opens the four-field list.

Any read as unreadable text.

1. At the date question type haan.

Not accepted as a date (re-ask); does not skip ahead.

"haan" advances the flow.

1. After a confirmation number is captured (§6.54) type kal mausam kaisa rahega.

No fake reply; a divider no handler for this input in "packet confirmed" yet; the Debugger raises P1 no-reply (accepted, §15).

A confident wrong reply; a crash.

1. Answer the first question wrong, the rest correctly.

"2/3 sahi ho gaye. Streak: 1 din — kal phir milte hain." The streak still went to 1.

The task blocked or restarted on a wrong answer; the streak not incremented.

1. Complete a task, click the Debugger's +1 day, start again. 2. Complete a task, click +7 days, start again.

After +1 day: the same 3-question task, streak continues. After +7 days: "Aaj ek chhota sa kaam — bas 1 sawaal." — one question, the streak carries over unchanged, and nothing says a day was missed.

Any wording like "you missed" or "you didn't do it yesterday" — S1 (CLAUDE.md §3.5, never guilt); the streak reset to 0 for a miss.

1. Complete a task, then type aaj ka kaam again the same session.

"Aaj ka kaam ho chuka ✓ — streak {n} din. Kal phir milte hain." No new questions; the streak unchanged.

A second task offered the same day; the streak counted twice.

1. Mid-task type something that isn't one of the three options, three times in a row.

"Neeche diye options me se ek dabao." then the same question again, every time. The Debugger's repeat-verbatim rule stays quiet no matter how many times you retry.

The same exact bubble twice in a row; an answer guessed from free text.

1. Complete a task and check the Events panel. 2. Send a distress phrase mid-task. 3. delete → Haan, hata do after a completed check-in.

Events command_used command=checkin, checkin_started lighter=false, checkin_answered correct=… ×3, checkin_completed score=… total=3 streak=… — no question or option text in any event. Distress still routes to a person from mid-task. Erasing drops the streak with everything else.

Question/answer text inside an event — S1; distress not routed; the streak surviving an erase.

1. At the eligible list type baad me (declining to apply for now).

"Theek hai — jab man kare, 'aaj ka kaam' likh kar ek chhota kaam try karo, roz ka streak banao." (This state, first_task, said nothing at all before #60 — a real gap, fixed here.)

Silence; a P1 no-reply finding.

1. Note the Debugger's Clock row (real time, no +Nd suffix). 2. Click +1 day, then +7 days — the label updates each time ((+8d)). 3. Send a message; check its timestamp reflects the advanced clock. 4. Click Reset.

The clock label always shows the simulated time, with an offset suffix once advanced; every message sent after advancing is timestamped at the simulated time, not real time; Reset returns to real time and the button disappears at offset 0. No other flow is affected by an unadvanced clock (offset 0 behaves identically to before this ticket).

A sent message timestamped at real time despite an advanced clock; the offset failing to reset.

1. Type reminder chahiye. 2. Type it again. 3. Use the Debugger's clock to jump to the T-7 day (or +7/+14 days from opt-in), send any message. 4. Jump to T-3, then T-1, sending a message each time. 5. Type band karo, jump forward again. 6. Type reminder chahiye again.

(1) "Theek hai — jab deadline paas aayegi, main yaad dila dungi..." (2) "Reminders pehle se on hain." (3)–(4) one "Test reminder: SSC GD (DRAFT)..." message per crossed window, each exactly once, clearly saying "test" — never presented as a real official date; the turn's own reply still follows right after. (5) reminders stop; the opt-in itself is remembered. (6) reminders resume from wherever they left off — the deadline anchor never resets.

A reminder claiming to be the notice's real deadline (a reputation-event shape, CLAUDE.md §10); a reminder firing twice; band karo wiping the opt-in instead of just pausing it; a reminder mixed into the same turn that first triggers distress.

1. Type ssc gd 2027 sahi hai kya. 2. Type RRB ALP 2024 bharti nikli hai, sahi hai kya. 3. Type something unrecognisable, e.g. kisi ne mujhe ek naukri ka notice bheja hai, sahi hai kya. 4. Mid-onboarding (after "haan", before answering DOB), type an old-notice check, then answer the DOB question normally. 5. Type ssc gd form kaise bharein (no year, no "sahi hai kya", no notice word).

(1) SSC GD 2027 is named as tracked but not published — never "Verified" — plus its 2026 source PDF as a document card. (2) "Ye purana notice hai: RRB ALP (2024). Abhi koi naya RRB ALP notification nahi hai." (3) "Ye notice mujhe pehchaan nahi aaya — ise unverified maano..." — never guessed either way. (4) the notice reply lands, then the DOB question is answered exactly as if nothing interrupted it — the check is a pure self-loop, never a committing step. (5) ordinary onboarding capture — no notice reply at all; this is a genuine "how do I apply" question, not a forward.

Any notice ever called "Verified" (nothing in this POC's data is actually published yet — CLAUDE.md §10, a fake notice marked Verified is a P0-class reputation event, and so is the inverse); a real notice-choice tap or an "apply" intent swallowed by the checker instead; the verdict changing the conversation's state; the checked text appearing in an event's props.

1. Note the Cost panel right after the greeting (one message, 0 turns). 2. Walk through onboarding to the eligible list (step by step, no smart capture). 3. Check the breakdown rows. 4. Opt into reminders and let one fire (§7.22), then re-check the breakdown.

(1) 0 turns ⇒ no ₹/10-turns figure shown (nothing to project from yet). (2) the panel updates live every turn; the ₹/10-turns badge turns amber once it crosses ₹2.50 — expect it to, since the step-by-step path is a known, already-quantified gap (§7 turn/cost budget). (3) every line is "Service (replies)" plus a final "LLM / voice (not called yet) — ₹0.00" line, always present, never silently omitted. (4) after a reminder fires, a new "Utility (reminders)" line appears, separate from "Service" — a reminder is never folded into the free-flowing service-conversation cost, because in production it would be a genuine WhatsApp utility template send.

A ₹/10-turns figure shown at 0 turns; the LLM/voice line disappearing instead of reading ₹0; a reminder counted as "service"; the badge never turning amber even on a conversation that's visibly gone over ₹2.50/10 turns.

1. Reach the eligible list, then type thanks. 2. Repeat with ok, theek hai, dhanyavad. 3. At the very start, type ok in answer to the greeting's "Haan, shuru karein?".

(1)–(2) a brief, warm reply ("Koi baat nahi!...") — never the old "no handler for this input... (DigiLocker steps not built)" console note, and the state doesn't move (a self-loop). (3) "ok" still works exactly as a real "yes" to the pending question — moves straight to asking DOB, never swallowed as a bare closer.

Silence or the stale "no handler" note for a plain "thanks"/"ok"; a genuine "yes" answer at a pending question mistaken for a closer and not advancing.

1. Note the Debugger's rule count (19). 2. Drive any scripted scenario to completion and confirm 0 findings. 3. (Developer-only) feed the debugger a bot message containing "just", "quick", "As an AI", or "Oops" and confirm filler-voice fires. 4. (Developer-only) a conversation at ≥10 turns whose cost projects over ₹2.50 should trip cost-over-budget; the same shape at <10 turns should not.

Every existing scenario stays debugger-clean — these are additive checks, not stricter ones. filler-voice and cost-over-budget only fire on genuinely bad content/cost, never on real product copy or any scenario measured today.

A new rule firing on existing, correct copy or an existing scenario (a false positive); cost-over-budget judging a short conversation (the volatility the ≥10-turn gate exists to avoid).

1. Note the Debugger's rule count (21). 2. Drive any scripted scenario that reaches distress, guide_hold or identity_wrong (the Distress probe; "kuch galat hai" at the identity card; an unrecordable claim) and confirm the hand-off message says a person is coming (a timeframe, or Tele-MANAS 14416) and that 0 findings fire. 3. Drive the delete command to confirm_delete and confirm the ask is the real yes/no card, 0 findings. 4. Reach packet_confirmed (the completed application) and confirm it is not flagged by no-next-step even though it names no further command. 5. (Developer-only, console-internal) feed the debugger a hand-off to identity_wrong/guide_hold/distress whose message never says a person is coming, and a confirm_delete/consent ask that is bare text instead of a card — both should fire.

Every existing scenario stays debugger-clean (additive checks, not stricter ones) — no-next-step and unconfirmed-commit both start at zero live findings, same as every debugger rule shipped in #63 before them. no-next-step only applies to the 3 kind: "human" states (a genuine dead end), never the 2 kind: "terminal" ones (a graceful, successful close is a status update, not a roadblock — L8's own wording). unconfirmed-commit only applies to the 2 yes/no gates before a consequential, hard-to-reverse action (confirm_delete, consent) — packet_shown's typed confirmation number is deliberately out of scope, since it is a value entry, not a yes/no gate, and is separately guarded by construction (GUIDED_TYPED keeps it out of the general smart-capture sweep, so it can never be inferred from unrelated free text).

Either rule firing on existing, correct copy or an existing scenario (a false positive); a future hand-off or consequential-action copy change silently removing the next-step line or the yes/no card without the debugger catching it.

1. Right after 8.1 type koi hai?, then 14-03-2002.

Only the informational line "Hamari team aapko jaldi call karegi. Tele-MANAS 14416 par kabhi bhi baat kar sakte ho." — no question, no parsing of the date, no reassurance.

The date is accepted; a new question; "everything will be fine".

1. After 8.1, send another message.

Divider guide ticket G-n updated — message added for the guide. The ticket in the queue shows the note.

The message is lost.

1. In the Guide queue click Resolve on the distress ticket.

Divider guide ticket … resolved … conversation resumes at "<state>"; the ticket turns resolved; the next message continues the form from that exact state.

Restart from the beginning; the conversation stays locked.

1. Trigger mujhe jeena nahi hai at: the greeting, a question, the card, the result, and after a form is started.

The same handling every time.

Any stage where it is read as an answer or ignored.

1. Try main sab khatam kar dunga, jeene ka mann nahi karta, I want to end my life.

Each goes to the distress path.

Any is treated as chatter — S1.

1. At a question send I give up on this form, and ye form bahut mushkil hai.

Normal handling (re-ask / help), no distress ticket. And the bot never says "give up" itself.

A distress ticket for frustration; the bot telling the user to give up.

1. Open a distress ticket and a help ticket.

Distress "call by" ≈ 60 minutes out; help ≈ 2 hours; times in your local timezone; counters show "n open".

Both the same SLA; times in another timezone.

1. With an open ticket for Rohit click Reset.

The conversation restarts and that ticket is closed.

An orphan ticket for a conversation that no longer exists.

1. Fail the same question three times.

"Ye thoda atak raha hai — maine hamari team ko bata diya hai. Koi aapko 2 ghante ke andar call karega…" and a help ticket in the queue whose text names the field and the attempts. Divider "3 failures at this question → help ticket requested".

A fourth identical re-ask; no ticket; the attempts lost.

1. Keep failing after 9.2.

Later replies vary; no second help ticket while one is open.

A ticket per failure.

1. At education type xyz, then abc.

"Ye list me nahi mila. Neeche 'Chunein' dabao aur ek chuno." then a second, different wording.

The question text repeated.

1. Fail the date twice, answer it correctly, then fail the state question once.

The state failure is counted as 1 (plain re-ask), not 3.

Escalates immediately.

1. Use the Gibberish probe 3 times.

As 9.1–9.2; the Debugger shows retry-loop if a 4th identical reply happened (it must not).

A loop without a ticket.

1. Walk Rohit to the card. Switch (pulldown) to Sunaina. Switch back.

Sunaina is at the greeting; Rohit is exactly where he was.

Any state or answer carried across numbers.

1. Give Rohit his profile. On Sunaina type mere form and hi.

Nothing from Rohit (date, category, form) appears in Sunaina's thread, buttons or Events tagged to her.

Another profile's data shown — a reputation event.

1. Events panel: this number vs all numbers while switching.

The counts and rows follow the number; all numbers shows the last 5 digits per row.

Events of one number listed under another.

1. Open the pulldown → + New number.

A new persona in the +9199999xxxxx block, greeting shown, becomes active. There is no way to type a number.

Any way to enter a real phone number; a number outside the block.

1. Give Rohit a help ticket and a Debugger finding (e.g. Devanagari probe).

The pulldown shows an open-ticket badge and the finding count against Rohit only.

Badges on the wrong number.

1. Progress two numbers. Reset one.

Only that thread restarts; the other stays.

Both reset.

1. Give Rohit a few turns (past the profile card). Reload the page (same browser tab or a new tab in the same browser).

Within a few seconds, the full transcript and current state reappear exactly as left — including the very first greeting message. A second browser tab on the same conversation picks up a reply sent from the first within ~4 seconds (polling), without reloading.

The conversation resets to the greeting; the greeting message itself is missing from the restored transcript even though later messages are present (a narrower, easy-to-miss variant of the same failure — check the very first bubble specifically); a stale copy that never updates from the other tab.

1. Progress Rohit in one browser (or clear localStorage → reload → note you're now a "different visitor"). Open the same URL in a second, unrelated browser (or an incognito window) and also pick Rohit.

The second visitor's "Rohit" starts at the greeting, completely independent of the first's progress — each anonymous visitor gets their own private copy of every persona, not a shared one.

Two unrelated visitors both picking "Rohit" see or edit the same live conversation.

1. Walk Rohit's full path to a started form. Watch the panel.

P0 0 · P1 0 · P2 0 (until you type something unhandled after the start).

Any finding on a happy path — file it, it is either a product bug or a bad rule.

1. Click Distress, Devanagari, Gibberish (×3).

Distress → no finding when routed correctly; Devanagari → a P2 register-mismatch; Gibberish×3 → the escalation (§9.2) and no P0.

A P0 on a correct path.

1. Trigger a finding. Click jump →.

The thread scrolls to the offending message and highlights it.

Nothing happens.

1. Filter P0/P1/P2/all; toggle all numbers / this number.

The list follows the filter and the counts.

A stale list.

1. Click Copy report and paste it somewhere.

A plain-text report: counts, each finding with its source, the transcripts involved, your timezone.

An empty report; the button says "Copied ✓" but nothing was copied.

1. Rohit: one-sentence capture (§3.1), then Haan, sahi hai.

Rows in order, e.g. conversation_started source=console · answer_captured field=dob via=asked … profile_card_shown · profile_confirmed · eligible_list_shown n=1 k=1. Times in your timezone.

Missing steps; out of order; wrong via (asked / sentence / correction).

1. Send messages that look like personal data: Mera naam Rohit Kumar hai, phone 9876543210, aadhaar 1234 5678 9012, rohit.kumar@example.com.
2. Export JSON from the Events panel and search it.

None of that text — no name, number, date of birth or email — appears in any props. Only codes (field=dob, n=1 k=1, from=ask_state).

Anything a user typed inside an event — S1 (props are ids and codes only).

1. Click Export JSON.

A file jivika-events-…json downloads; it parses; rows are numbered from 1 with seq, at, phone, name, props.

No download; invalid JSON.

1. In the browser console run window.__jivikaDebug.events.length and window.__jivikaDebug.findings.

A number and an array.

undefined.

1. Open /flow. Scroll the field guide.

15+ phone-style threads, each a heading, a sentence and a conversation; four diagrams render (no raw Mermaid text, no "Syntax error"); a contents list works.

A thread blank; a diagram shown as code.

1. Pick three threads (e.g. the eligible result, "mere form", the card correction) and replay them in the console.

The console produces the same messages (the page is generated from the same code).

A difference.

1. Open /scope.

Sections, three diagrams, the "Built (slice 0)" note under the event spine.

A build/render error.

1. View source of /flow, /scope, /qa.

<meta name="robots" content="noindex, nofollow"> on each.

Indexable.

1. Use the top nav: Simulator, Gallery, Flow, Scope, QA, Ops.

Each goes where it says; the active item is highlighted; works on a phone-width window. Since #13: "Ops" is still a public nav link on every page (the label itself isn't a secret — it's documented in this file and in HANDOFF.md), but clicking it when signed out now 404s rather than showing the shell — that's the intended gate, not a dead link.

A dead link for a signed-in ops user; /ops rendering real content for a signed-out visitor.

1. Visit /, /console, /flow, /scope, /qa.

A consistent dark theme; readable text; diagrams legible.

A white flash or unreadable section.

1. Sign in at /ops/login with an allowlisted account (§12.8). 2. Open /ops — note the "Funnel lens" card (not a "not built yet" placeholder like the other three). 3. Click through to /ops/funnel. 4. Read the stage bars top to bottom. 5. Read the five stalls.

A non-empty, non-increasing funnel (registered ≥ profiled ≥ … ≥ retained); "profiled" and "saw value" show the same count every time (they fire in the same turn); all five stalls render, including "profiled, never saw a list" at 0 with its own one-line explanation (never hidden); a footnote explicitly labels "Retained" as a streak-≥2 proxy and says Credentialed/Advocate aren't built.

An empty or all-zero funnel; a stall silently omitted instead of explained; "Retained" presented as if it were real D30 activity; any count that isn't traceable to a real, re-runnable synthetic lead.

1. Signed out, visit /ops and /ops/funnel directly. 2. Visit /ops/login and click "Continue with Google." 3. Sign in with a Google account that is not on the allowlist. 4. Sign in again with an allowlisted account (the founder's own, already seeded). 5. With the session from step 4, revisit /ops and /ops/funnel.

Step 1 → a plain 404 for both, not a redirect and not the shell (the surface isn't advertised). Step 3 → bounced back to /ops/login?denied=1 naming the rejected account, app still otherwise fully usable (nothing crashes). Step 4 → lands on /ops. Step 5 → both pages render real content.

/ops renders anything for a signed-out visitor; a non-allowlisted Google account reaches /ops; the denied banner doesn't name which account was rejected; a crash instead of a clean bounce.

1. In /console, on a synthetic number, say a distress phrase ("mujhe jeena nahi hai"). 2. Send a second message while still in distress. 3. Sign in at /ops, open the "Guide desk" card. 4. Claim the distress ticket. 5. Add an internal note. 6. Reply into the thread. 7. Go back to /console on that same number and wait ~4 s (the existing poll). 8. Return to /ops/guide and resolve the ticket. 9. Check the /console thread again.

Step 1 → a guide_tickets row appears (priority P0, due in 60 min, excerpt = the distress line). Step 2 → the SAME row's "while waiting" notes grow by one — no duplicate ticket. Step 3 → the card reads "live", not "not built yet". Step 4 → status flips to claimed. Step 6 → the reply shows up in the candidate's /console thread as a normal assistant bubble, signed "— Jivika guide se" — the candidate never sees the internal note from step 5. Step 8 → the ticket leaves the open queue. Step 9 → the conversation has resumed exactly where distress interrupted it (the next assistant message answers whatever question was pending, not a repeat of the distress-hold line) — never left stuck.

A second distress message creates a second ticket instead of appending a note; an internal note leaks into the candidate's thread; a reply never reaches /console even after a hard reload; resolving a distress ticket leaves the candidate's conversation still parked in the distress state; /api/ops/tickets or /ops/guide return anything but a 404 when signed out.

1. Read the whole thread of a full run.

Romanised Hinglish throughout, in the voice of a helpful senior — short sentences. Never formal Hindi, never Devanagari, no emoji storms.

Devanagari or stiff Hindi ("kripya…"); English-only lecture.

1. After Baad me, band karo, a "not eligible" result and an over-limit result, read the reply.

Calm and forward-looking ("koi jaldi nahi", "hum dekhte rehte hain").

Any sentence that blames, shames, pressures or tells the person to give up — S1.

1. After a distress message read every bot line.

Acknowledgement, a person is coming, Tele-MANAS 14416 — nothing else.

"Sab theek ho jayega", motivation, emoji.

1. Note every structured decision (greeting, education, category, card, form card, delete).

Each is buttons or a list; free text is used only for the date, state and corrections.

A yes/no or a pick that requires typing.

1. Read each bot message.

Two sentences or fewer (the card and consent-style notices excepted), no stacked bubbles saying the same thing.

A wall of text; the same content sent twice in a row.

1. Look at message times and the Guide queue.

e.g. 10:33 AM PDT — your local timezone next to the time.

No zone; the server's zone.

1. Read every bot message across a full run.

No {n}, {first}, {date} or undefined in any text.

A raw placeholder — the Debugger shows unfilled-placeholder.

1. Search the page, the report and the events export for anything that looks like a real number/name.

Only +9199999xxxxx, "(synthetic)" names.

Real personal data anywhere — S1.

1. Any run: find the first mention of DigiLocker/Aadhaar/documents.

After the eligible list, and only after the person chose to start a form.

Earlier — S1.

1. Read every message about forms.

No fee, payment, "I submitted for you" or a link that submits. (Fee guidance and links arrive with guided apply.)

Any charge, convenience fee, paid tier or claim of submitting — S1.

1. Look for any employer-facing wording or private-job pathway.

None.

Any — out of Phase I scope.

1. Repeat 5.3, 5.6 and 5.7 and check the message against the ✦ engine: divider.

Always consistent, limits come from the rules (23/26/28).

A mismatch — S1.

1. Read each result and each form row.

"DRAFT" / "Draft — official notice aana baaki" everywhere until a real notice exists, and the source document is the official 2026 PDF.

An unlabelled claim — a fake-notice-marked-Verified class of bug, S1.

1. Repeat 10.2 with the same number after Reset and after delete.

No trace of the previous profile.

Leftover answers or documents — S1.

1. Repeat 7.7 and inspect thread, mere form, Events (Export JSON) and window.__jivikaDebug.

Nothing attributable to the number remains.

Any leftover.

1. After a full run of this plan check the panel across all numbers.

Only the findings you intentionally provoked (Devanagari P2, unhandled-after-start P1).

An unexplained P0.

1. Repeat 5.1, 4.4 and 6.3 on www.jivika.dev/console after a deploy.

Identical to local. (Check the deployed content, not just the deploy status — the domain lags ~20 s.)

A difference.