Gemini 4 Argon Low · Pure Uncoached Protocol

211.5B Paired-Seed Prompt x Code Mutation Engine

SDCS-Bench

Spec-Driven Cognitive Scaffolding vs. Mid-Stride Resets

When an agent session is cut off mid-task and reset (`git reset --hard`), uncommitted code vanishes. In a 6-file repo, `state.md` alone is enough. In a 24-file, 4-package monorepo, `Full_SDCS v1.9.7` cartography (`app_map.md` + AST skeletons) cuts turns by 40.7% and cached tokens by 39.8%.

-40.7%fewer turns on Tier 2 (16 vs 27)

-39.8%fewer cached tokens vs State.md

3D Specimen: tier_2a_fintech Monorepo (4 packages · 24 files · 855 LOC · 8 volatile rules)

Every slab in that 3D tower is a module or requirement from the live FinTech monorepo. Scroll or hover any slab to inspect.

Scroll to Trigger Mid-Stride Cutoff 855 LOC → 238-token state.md

Stage 02 · The Mid-Stride Cutoff Shockwave

8,550active tokens

Pre-Cutoff Monorepo (855 LOC)

Mid-sprint, 8 volatile requirements arrive across ledger, billing, rbac, and ratelimit. Immediately after the agent's first file edit, the session is killed and uncommitted code is wiped (git reset --hard). Pure Naked loses everything (0/8 retained). Full_SDCS v1.9.7 locks all 8 rules and all 11 exact target file paths into a 238-token state.md on Turn 1.

Interactive 3D Cutoff & State-Lock Laser

8 volatile specs + 11 target paths (kept) AST skeletal interfaces uncommitted edits wiped at reset

Full_SDCS_v1_9_7 · state.md handoff (c740fc41 -> 8db7b77f)

## Current Objective
- ledger: allow_overcapture (<=110% hold, overage <= available_cents); released_hold_cents in build_account_summary.
- billing: max_line_discount_cents in compute_line_discounts & finalize_invoice; restocking_fee_bps in apply_net_credit.
- rbac: ':*' prefix wildcard in evaluate_access_decision; revoked: bool = False on TempGrant.
- ratelimit: overdraft_tokens >= 0 in QuotaPolicy/gateway/evaluate_token_bucket; read-only peek_tokens().
## Immediate Next Action
- Edit ledger/rules.py, ledger/service.py, ledger/reporting.py, billing/discounts.py, billing/invoice.py, rbac/entities.py, rbac/rules.py, ratelimit/models.py, ratelimit/quota_store.py, ratelimit/algorithms.py, ratelimit/gateway.py.

Stage 03 · Tier 2 Monorepo Turn-by-Turn Skyline

Why State.md alone burns 27 turns in a 4-package monorepo.

In Tier 2 (tier_2a_fintech, 24 files across 4 packages), State_MD_Only saves the requirements in state.md, but has no app_map.md or AST skeletons. Session 2 wakes up blind and spends 22 turns searching folders and reading files. Full_SDCS v1.9.7 knows all 11 target files immediately and finishes at Turn 16.

Pure State_MD_Only

27 Turns

5 S1 + 22 S2 turns · PASS (8/8)
1,437,520 cached tokens
151,325 uncached · 29,560 out

Pure Full_SDCS v1.9.7

16 Turns (-40.7%)

4 S1 + 12 S2 turns · PASS (8/8)
866,008 cached tokens (-39.8%)
97,403 uncached (-35.6%)

3D Turn-by-Turn Context Laser Inspector Turn 16 of 27

State_MD_Only: 74,856 tok Full_SDCS: 81,118 tok (DONE · 8/8 PASS)

Hover any 3D bar in the skyline or drag the slider to compare per-turn context accumulation across Session 1 and Session 2.

Stage 04 · The Scale Crossover Ledger

-39.8%

cached token reduction on Tier 2 monorepos with Full_SDCS v1.9.7 over State_MD_Only.

Every arm below was executed under the exact same Pure Uncoached Mid-Stride Cutoff Protocol on Gemini 4 Argon Low with domain-agnostic AGENTS.md templates.

8 / 8 (100%)

volatile sprint requirements retained across all 4 FinTech packages by SDCS & State.md arms (vs 0/8 on Pure Naked)

16 vs 27

total turns on Tier 2 (Full_SDCS v1.9.7 finishes in 12 post-reset turns vs 22 turns for State_MD_Only)

Scale Crossover

Tier 1 (6 files) favors minimal State.md; Tier 2 (24 files, 4 packages) flips decisively to Full_SDCS v1.9.7

TIER 2 MULTI-PACKAGE MONOREPO BENCHMARK (tier_2a_fintech · 4 PACKAGES · 24 FILES · 855 LOC · 8 VOLATILE REQS) Hover any row to highlight its 3D Plasma Monument →

Arm	Outcome	Retained	Turns (S1+S2)	Steps	Uncached Input	Cached Input	Output	Key Mechanism

TIER 1 SINGLE-PACKAGE MICROSERVICE ABLATION (task_01_ledger_hold · 1 PACKAGE · 6 FILES · 2 VOLATILE REQS) Complete 7-Arm Ablation Baseline

Arm	Outcome	Retained	Turns (S1+S2)	Steps	Uncached Input	Cached Input	Output	Key Mechanism

Stage 05 · Scale-Adaptive 7-Pillar Architecture

How SDCS v1.9.7 adapts from 6 files to 24+ files automatically.

01 · Scale Detector

sdcs/hydrate.py

Under 4,000 source tokens (Tier 1), SDCS pre-hydrates verbatim source files. Above 4,000 tokens (Tier 2), it switches to Standard Cartography.

→

02 · The Compass + AST

Pillar 5 + AST Skeletons

Injects app_map.md, compact AST class/method signatures for all 20 modules, and decisions.md rejection guards on Turn 1.

→

03 · The Blackboard

state.md (<= 300 tok)

Session 1 locks all 8 requirements AND all 11 exact target file paths on Turn 1, enabling zero-search surgical edits in Session 2.

Pillar 5: The Compass (app_map.md)

Maps all 4 packages (ledger, billing, rbac, ratelimit) on Turn 1 so the agent never wastes turns running directory scans.

AST Skeletal Interfaces

Provides exact function and dataclass signatures across all 20 modules at a fraction of full-source token cost.

Pillar 6: The Graveyard (decisions.md)

Pre-warns against mutating BucketState in peek_tokens() (REJ-002) and integer floor division in billing (REJ-003).

Interactive Session 1 -> Session 2 Handoff Inspector (tier_2a_fintech)

Compare the exact on-disk working memory left behind at the moment of the mid-stride cutoff across Full_SDCS_v1_9_7, State_MD_Only, and Pure Naked.

Selected Arm & Session 1 ID Full_SDCS_v1_9_7 · Session 1 Handoff (c740fc41)

Cartography & File Path Precision 11 exact target .py paths locked on Turn 1 via app_map.md + AST Skeletons

PASS (8/8 Retained · 16 Turns · 866k Cached Tokens)

ON-DISK WORKING MEMORY AT CUTOFF (state.md) 238 tokens (Gate S <= 300)

WHY THIS HANDOFF DETERMINED SESSION 2 EFFICIENCY Empirical Trace Analysis

1. Turn 1 Cartography Advantage:
   Because Full_SDCS_v1_9_7 hydrated # [PILLAR 5: THE COMPASS] (app_map.md) and # [AST SKELETAL INTERFACES] on Turn 1, Session 1 (c740fc41) immediately knew the exact 11 target files across all 4 packages without calling list_files or read_file.

2. Zero-Search Session 2 Resume:
   When Session 2 (8db7b77f) booted after the mid-stride reset, its ## Immediate Next Action already listed all 11 file paths:
   - ledger/rules.py, ledger/service.py, ledger/reporting.py
   - billing/discounts.py, billing/invoice.py
   - rbac/entities.py, rbac/rules.py
   - ratelimit/models.py, ratelimit/quota_store.py, ratelimit/algorithms.py, ratelimit/gateway.py

3. Result:
   Session 2 completed all 8 cross-package requirements in just 12 turns (16 total turns vs 27 for State_MD_Only), saving 571,512 cached input tokens (-39.8%) and 53,922 uncached input tokens (-35.6%).