The Verified Work Product — From Task to Filing

In this session

• 01

  The opening measurement: Session 1's anchoring experiment runs again, same protocol, new numbers — and the gap between the anchored halves, put beside the Session 1 gap on screen, has narrowed. The room is its own before-and-after study: the course's claim was never that training removes bias, but that knowing the mechanism and committing answers before reveals changes behaviour measurably. The intake survey's before-and-after tells the same story across more dimensions at term's end.

• 02

  The pipeline demonstration, narrated by session number at working speed: the task (a supervising advocate wants an urgent note) is decomposed (S2's specification); the matter's workspace supplies the grounding (S4); the draft emerges under the output contract with its authorities table honest and unchecked (S2); a cold tab attacks it (S5); the four steps run on everything load-bearing, trail filling as it goes (S6); the operative paragraphs get read against the draft (S3); and the disclosure is written last, from the trail, in two minutes because the trail already contains it (S7). Total elapsed time on the projector clock: under twenty minutes — the point being not speed for its own sake, but that the professional workflow costs minutes and the unprofessional one costs careers.

• 03

  Competence, as the cases define it now: knowing the failure modes is no longer technological literacy but professional adequacy — a lawyer who cannot say why a model fabricates cannot sensibly decide when to rely on one. Session 1 was a professional-responsibility class in disguise, and Pooja Ramesh Singh makes the connection explicit: 'zero tolerance' is addressed to a profession presumed capable of checking.

• 04

  Candour and the question that survives every upgrade: you certify what you file. The provenance of a draft never transfers responsibility — not from the junior who wrote it, not from the machine that drafted it. The 'who signs it?' move puts a filing-ready paragraph with one confident citation on screen and asks only that: whose name goes at the bottom? Every improvement in the tools sharpens rather than dulls this question, because better drafts survive more careless review.

• 05

  Confidentiality and supervision, closed out: what enters any tool is governed by the frame Session 7 completed — Guideline 12's flat rule, DPDP's fiduciary structure, privilege that belongs to the client — and supervision now runs in both directions: you review the machine's work as a senior reviews yours, and the review standard is higher, because the machine's errors arrive better-formatted than a junior's. The UK's Cork v. Smith — cited by our own Supreme Court — is the complete cautionary anatomy: an internal firm AI tool, its own verification warnings ignored, a junior's draft, supervisors approving unchecked, a court misled twice. Every safeguard existed; every human waved it through.

• 06

  Disclosure, converging from three directions: the draft court Regulations' reg. 43 (a draft, dated on the slide) would require disclosing AI assistance to courts; universities and journals increasingly require it in submissions; and clients and partners increasingly expect it in practice. The course's form — tools and tiers used, what each did, what was verified and how, what was rejected — is deliberately shaped so that whichever regime crystallises, the trail already answers it. Disclosure written from a trail takes two minutes; disclosure written from memory takes a career.

• 07

  Agentic AI, seriously and briefly: tools that browse, transact and file convert output-checking into action-authorising — the verification point moves before the act, or it does not exist. Deep-research agents, the mildest agents students will meet, already cite more and fabricate more than chat models in measured studies; anything that acts on a court registry, a client account or a filing system inherits every duty in this course at higher stakes. The rule that scales: the more autonomous the tool, the earlier and more explicit the human authorisation — and the more complete the trail, because after an action, the trail is all there is.

• 08

  The audit, framed before it runs: three failure classes, seeded the way reality seeds them — a fabricated authority (the class everyone catches), a real authority misdescribed (the class that misleads courts), an overruled authority correctly quoted (the class that loses cases). The prediction from Session 6 — that catch rates fall exactly in that order — is about to be tested on the room itself. The audit is also the assessed artefact's dress rehearsal: same method, same trail, same disclosure, graded.

• 09

  The close, which is an identity rather than a summary: the course began with a machine that answers everything confidently and a profession paying for believing it. It ends with a person who briefs precisely, grounds deliberately, orchestrates when it pays, verifies always, and keeps the record — on free tools, under Indian law, with a named technique for every move. The model's confidence was never evidence. Yours, now, is earned — and documented.

The mirror

Run it on the class. Then on the machine.

An experiment on the room, the same effect explained in the model, a live demonstration on a real tool, and a named takeaway skill.