extropy estimate assumes early convergence, ignores timeline events #112

Activity

1. DeveshParagiri commented on Feb 18, 2026

   

   DeveshParagiri

   on Feb 18, 2026

   CollaboratorAuthor

   More actions

   Deeper analysis: estimator is 4-5x off

   The estimator undercounts for three compounding reasons:

   1. Ignores re_reasoning_intensity

   The ASI scenario has re_reasoning_intensity: extreme on timesteps 1, 2, 3, and 6. When extreme, the engine forces ALL aware agents to re-reason (engine.py:656-660). The estimator doesn't model this at all — it uses a flat 2% re-reasoning rate (estimator.py:214).

   2. Ignores early convergence auto-disable

   As noted in the original issue, the estimator assumes early stop at ~3 timesteps. The engine runs all 6 because allow_early_convergence: null + future timeline events = disabled.

   3. Doesn't model conversations

   Medium/high fidelity adds 4-8 LLM calls per conversation (all fast model). The estimator counts zero conversation calls.

   Actual vs estimated for ASI (5,000 agents × 6 timesteps)

   	Estimator	Actual (projected)
   Effective timesteps	3	6
   Reasoning calls	~5,100	~21,000-22,000
   Total LLM calls	~10,200	~42,000+
   Cost	$5.93	$25-35

   Per-timestep breakdown (projected)

   Timestep	Intensity	Reasoning calls	Why
   1	extreme	~4,000	Seed exposure (broadcast)
   2	extreme	~5,500	Remaining exposed + ALL re-reason
   3	extreme	~5,000	ALL aware agents forced
   4	high	~300	Multi-touch only
   5	high	~1,500	Some re-reasoning
   6	extreme	~5,000	ALL aware agents forced

   What needs fixing

   The estimator needs to:

   1. Read re_reasoning_intensity from each timeline event and model forced re-reasoning (extreme = all aware, high = fraction)
   2. Apply the same early convergence auto-disable logic as the engine
   3. Account for conversation calls based on fidelity setting
2. DeveshParagiri commented on Feb 18, 2026

   

   DeveshParagiri

   on Feb 18, 2026

   CollaboratorAuthor

   More actions

   Example: estimator vs reality for ASI scenario (5,000 agents × 6 timesteps)

   What the estimator says

   Effective timesteps: ~3 (early stop at ~100% exposure)
   Total calls: ~10,200
   Cost: $5.93
   

   What actually runs (projected)

   The estimator misses three things: (1) re_reasoning_intensity: extreme forces ALL aware agents to re-reason on timesteps 1/2/3/6, (2) early convergence is auto-disabled because every timestep has a timeline event, (3) conversations are not counted.

   Reasoning calls per timestep:

   Timestep	Intensity	Reasoning calls	Why
   1	extreme	~4,000	Seed exposure (broadcast)
   2	extreme	~5,500	Remaining exposed + ALL re-reason
   3	extreme	~5,000	ALL aware forced
   4	high	~300	Multi-touch only
   5	high	~1,500	Some re-reasoning
   6	extreme	~5,000	ALL aware forced
   Total		~21,000

   LLM calls breakdown by fidelity (gpt-5-mini @ $0.25/$2.00 per MTok):

                           MEDIUM FIDELITY          HIGH FIDELITY
   ─────────────────────────────────────────────────────────────
   Reasoning events:       ~21,000                  ~21,000
   Pass 1 calls:           21,000                   21,000
   Pass 2 calls:           21,000                   42,000 (2× for public stmt)
   Conversation calls:     ~21,000 × 15% × 1 × 4   ~21,000 × 15% × 2 × 6
                           = ~12,600                = ~37,800
   
   Total LLM calls:        ~54,600                  ~100,800
   

   Cost:

   	Medium	High
   Pass 1 (21K × ~2.2K in / 200 out)	~$20	~$20
   Pass 2 (21-42K × ~300 in / 70 out)	~$4.5	~$9
   Conversations (12-38K × ~800 in / 150 out)	~$6	~$18
   Total	~$30	~$47

   Time at 1000 RPM (Azure gpt-5-mini):

   	Medium	High
   Total calls	~55K	~101K
   At 1000 RPM	~55 min	~101 min
   With burst headroom	~45-55 min	~80-100 min

   Estimator vs actual:

   	Estimator	Actual (medium)	Actual (high)
   Calls	~10K	~55K	~101K
   Cost	$5.93	~$30	~$47
   Time	N/A	~50 min	~90 min

   Conversation % is a guess (15% of agents request talk_to). For a scenario like ASI where everyone has strong opinions, could be 30%+ — which would double the conversation line items.
3. RandomOscillations commented on Feb 24, 2026

   

   RandomOscillations

   on Feb 24, 2026

   Collaborator

   More actions

   Verification update (2026-02-24): issue still reproduces in current code.

   Current estimator logic remains timeline-unaware:

   • extropy/simulation/estimator.py:235-238 applies a generic exposure early-stop.
   • It does not mirror simulation runtime gating that uses future timeline events / allow_early_convergence policy.

   Given this mismatch, the CLI estimate surface has been temporarily stubbed to prevent misleading pre-run numbers while preserving estimator internals for later parity work.

   Temporary CLI behavior:

   • estimate is hidden from top-level help.
   • Direct invocation returns a temporary-disabled message referencing this issue.

   Keeping this issue open until estimator parity is implemented and tested against runtime stop behavior.