Azimuth
See the next probable move before the other side has committed to it.
President, YSR Congress Party
A red team for political reality
AZIMUTH works for the party. It does not flatter the party. It turns noise into a bearing and shows leadership the hardest defensible case before the opposition does.
One party. One verified picture.
The complete party intelligence foundry. Leadership, research, field, communications, legal, strategy and district teams work in one place. Models may propose; named humans authenticate sources and approve hypotheses for decision use.
Prove it—or refuse the call
In the tested proof, exact words, source, speaker and time stay attached. Dependent voices count as one room. Record and street stay separate. Weak proof returns: we do not know.
State to district. Field to leadership.
The same system locates an issue at every scale, routes verified field reporting upward and sends decisions back to the organisation. Place is context, never a score.
New evidence must change the model
Observed history and inference never silently merge. Every actor model and motive branch carries its source, alternative, confidence and falsifier. Named humans authenticate the sources and approve the hypothesis for use.
Attack.Defend.Judge what survives.
Red attacks the claim. Blue defends each component. An evidence judge separates what is measured, what humans judged and what remains unavailable.
What the working ledger says
The case ledger records a 2019–2023 recruitment gap, a February 2024 notice for 6,100 posts, a 2025 notice for 16,347 posts and a 421-post sports-quota clause. Only the sports-quota notice is packaged in this proof; verify the other source files before operational use.
16,347 POSTS / SOURCE FILE NOT PACKAGED
421 SPORTS-QUOTA POSTS / PRIMARY NOTICE 01 / 1 MAY 2025 / CLAUSES 1, 12
Use the factsagainst us
The lab tests how true facts can be selected, ordered, omitted or attributed to create a damaging story. Every attack must trace back to evidence. It cannot invent an allegation.
See the loss before the attack lands
Repeated grievances, contradictions and issue links become an early shape. The design carries that shape to its worst plausible consequence—but only through stated conditions.
See the likely move before public commitment
The LLM generates candidate moves from actor history, incentives, constraints and opportunity. Evidence, calibrated scoring and human judgment rank them. Worst motive must compete with the strongest benign explanation; rhetoric alone cannot raise its prior.
How could we be wrong?
The target process has nine attacks. If every runnable attack fails, the conclusion may be labelled SURVIVED KILL TEST. If one breaks it: CONCLUSION BROKEN. This demonstration has no conclusion entered.
Our response can become their proof
The design compares doing nothing, responding badly and responding with discipline. It would test whether our answer strengthens their narrative.
Walls, not gaps
The system cannot establish what the public believes, emotional frame, propagation timing or coordination. Effort alone cannot move these walls.
EMOTIONAL FRAME / LABEL HUMAN JUDGMENT
PROPAGATION AND TIMING / RECOLLECT ABSOLUTE TIMES
COORDINATION / DROP OR REDESIGN FOR LAWFUL BASIS
One issue. One decision.
Every failed assumption stays
A forecast would freeze with its assumptions. The outcome would remain separate. Failed assumptions would enter institutional memory.
Nine layers. One roof.
Mental Models form the bridge between evidence and prediction.
The accountable learning loop
The party-owned record routes verified evidence through competing hypotheses. A named person approves each forecast. When the outcome arrives, the gap becomes a recorded error and an explicit revision.
The fastest reliable build
PostgreSQL, pgvector and encrypted document storage keep exact spans, dates and source ownership attached.
Ollama handles private routine work. An approved frontier model handles difficult synthesis through the same strict schema. Move to vLLM only when load requires it.
Frozen cases test citations, calibration, red and blue attacks, prompt injection and regressions. One failed gate blocks promotion.
LoRA learns stable repetitive skills from human-labelled corrections. Canary and rollback preserve the base model.
Where we are
Works today
- Sources and exact words on three demo rows
- Independent-source counting
- The refusal rule
- 0 false calls in its 400-shuffle null self-test
Partly built
- Current-state view
- Tamper-evident capture, unconnected
- Frozen forecast record, empty
- Five of nine Kill Test attacks
Designed
- Mental models and war-game
- Attack and defence
- Pre-attack warning
- Assumption graveyard