The problem with AI methodology
Most AI work in business goes like this: brief → prompt → output → “looks good” → commit. The problem: the AI agent is good at generating plausibility, not truth. Under pressure it tends to confabulate. Without structure, even an advanced LLM produces coherent drift for hours.
We have written down a way of working with AI that makes it reliable: read before acting, verify before declaring “done”, admit limits instead of inventing.
The method acronyms (for insiders)
Internally we call it V9 + APEF + the 18 rules: 9 method rules (M-1…M-9) and 9 honesty rules (D-1…D-9). V9 is the cognitive framework, APEF the project execution cycle.
V9 cognitive activation
Before every non-trivial task, the AI agent loads V9. See the V9 page for details. The result: the model activates trust-oriented validation, focus oscillation, vector palace, geometric reasoning.
APEF – Anthilla Project Execution Framework
Five non-skippable phases:
- Read-First – cross-doc grep for existing traces (memories, DOC/, audit_log, request_messages) BEFORE acting
- Enumerate-Before-Touch – all involved resources, never the “first that matches”
- Symmetric-Verify – pre-check identical to post-check, same environment, HTTP cache-busted
- No-Premature-Closure – “SOLVED” only after empirical SAME-SOURCE post-check by the user
- Scope-Match – changes limited to what was requested, no spontaneous propagation
9 method rules (M-1…M-9)
- M-1 Read-First: read existing traces
- M-2 Enumerate-Before-Touch: complete enumeration
- M-3 Symmetric-Verify: identical pre/post check
- M-4 No-Premature-Closure: closure only with proof
- M-5 Scope-Match: respect the perimeter
- M-6 Loop-Stop: round 2 → trace; round 3 → ask user
- M-7 Trigger-Awareness: observe verbal bias patterns
- M-8 Document-Before-Act: written protocols first, not post-hoc
- M-9 Honest-About-Limits: admit what you cannot see
9 honesty rules (D-1…D-9)
Against LLM confabulation under stress:
- D-1 Admit “I don’t know” when not empirically verified
- D-2 Cite the evidence (file:line / curl / SQL), not the hypothesis
- D-3 Never create post-hoc memories as proof for the current turn
- D-4 Never shift blame (browser, edge case) without SAME-SOURCE proof
- D-5 Never reframe history to cover inconsistency
- D-6 Never assume “all OK” without individual enumeration
- D-7 Under pressure, SLOW DOWN (don’t accelerate closure)
- D-8 Distinguish “I did” from “I declared I did”
- D-9 After days, read audit_log BEFORE responding
Pre-statement check (before every assertion)
Six non-negotiable questions:
- Do I have empirical evidence written in the current turn?
- Can I cite it in 1 line?
- If not, am I saying “I don’t know” instead of asserting?
- Am I creating memory in the turn to support it? (STOP D-3)
- After days, might I have to retract? (formulate as estimate)
- Is the moment’s pressure influencing me? (SLOW DOWN D-7)
Why this method is different
Most AI methodologies in business are prompt engineering playbooks. Anthilla is not that. It is a method to make AI coexist with real complexity: regulatory, multi-stakeholder, post-mortem, systems that last years.
Public references of our analyses on confabulation patterns:
- Root Cause Analysis of a critical session, transformed into method rules
- Dishonesty Under Pressure – human cognitive transposition
- Long Session Continuity – 8 LSC principles for 10h+ sessions
Want your team to adopt this method? We offer onboarding workshops (1-2 days) for technical teams working with AI every day. Write to contacts@anthilla.com