# Codex V58 Regression Audit - Franken-B

Created: 2026-05-22

## Short Answer

Opus is mostly right that Adapter 58 is the likely source of Franken-B's major regressions, but the exact mechanism matters.

Adapter 58 is **not a vision-tower adapter**. It is a LoRA over the **last three language-model layers** only:

- layers: `29`, `30`, `31`
- target family: `model.language_model`
- adapter config target: `lm_last3`
- no `vision`, `visual`, or image encoder tensors in the safetensors keys

So v58 did not directly train the visual encoder. Its effect is language-side arbitration over already-embedded chart/image evidence. In Franken-B, that arbitration appears to unlock the best vision/fused behavior, but also weakens veto, HOLD, and adversarial boundaries.

## What Adapter 58 Was Trained To Address

Primary artifact:

- `E:\vfai-x_3.5_9b\v5.8\V58_TRAINING_REPORT_20260520.md`
- adapter repo: `tjarvis91/vfai-v58-adapter-v58_profit_direction_ondemand_lmlast3_300s_retry2_20260520`
- base: `tjarvis91/vfai-v50-merged-base`
- dataset: `E:\vfai-x_3.5_9b\v5.1\v51_profit_direction_focused_train.jsonl`
- train rows: `7914`
- LoRA: `r=64`, `alpha=128`, `lr=5e-6`, last-3 LM layers, 300 steps

The v51 focused dataset report says v58's source data targeted:

- winning long puts on falling underlying: `HOLD`, not `SELL`
- opening long puts: `BUY` the option contract even when underlying thesis is bearish
- clean aligned VPA shorts: `SELL`, not over-abstain
- clean aligned VPA longs: `BUY`, not over-abstain
- multi-timeframe conflict: `NO_TRADE`
- all-timeframe alignment: directional action
- profit protection: `SELL` on target/trail break, `HOLD` when winner trail is intact
- production overlay: 8 max positions and 12% position size

Dataset action mix from the train file:

- `SELL`: 2488
- `BUY`: 2028
- `NO_TRADE`: 1976
- `HOLD`: 1422

Source mix:

- `generated_v51_profit_direction_fix`: 4914
- `repaired_original_84k`: 1291
- `fundamental_second_pass_patched`: 589
- `generated_v52_plan_a`: 464
- `generated_hold_management_fix`: 259
- `generated_eval_failure_fix`: 166
- `v35_dataset`: 119
- `generated_profit_protection_fix`: 95
- `v32_dataset`: 17

Important: these rows are text/message rows. No actual image/chart file keys were present in the v51 train/val JSONL. The train set mentions chart-like context text, but it is not true multimodal vision supervision.

## CPU Tournament Read

Prior report:

- `E:\vfai-x_3.5_9b\evals\vpa_learning\reports\FRANKEN_CPU_TOURNAMENT_20260520.md`

Relevant with/without-v58 candidates:

| Candidate | Recipe | CPU Result | Action Mix | Take |
|---|---|---:|---|---:|
| `tour_v37s_baseline` | v37s | PASS, 8/8 valid, 7/8 strict | BUY 4 / SELL 2 / NO_TRADE 2 | 75% |
| `tour_v37s_plus_v50` | v37s -> v50 | PASS, 8/8 strict | BUY 4 / SELL 2 / NO_TRADE 2 | 75% |
| `tour_v37s_plus_v58` | v37s -> v58 | PASS, 7/8 strict | BUY 5 / SELL 2 / NO_TRADE 1 | 87.5% |
| `tour_v37s_v50_v58` | v37s -> v50 -> v58 | PASS, 8/8 strict | BUY 4 / SELL 2 / NO_TRADE 2 | 75% |
| `promote_v37s_v50_v58` | saved Franken-B recipe | PASS, 8/8 strict | BUY 4 / SELL 2 / NO_TRADE 2 | 75% |

The CPU tournament did test v37s/v50 with and without v58, but it was only an 8-row smoke. It was useful for schema and obvious action-collapse failures, not for adversarial, HOLD, or fusion regressions.

The one signal it did catch: `v37s -> v58` got hotter than the base line: BUY increased and NO_TRADE decreased. Adding v50 before v58 stabilized the tiny CPU smoke, which is why Franken-B promoted cleanly.

## GPU/Full Eval Read

Franken-B report:

- `E:\vfai-x_3.5_9b\evals\vfai_dev_docs\FRANKEN_B_FLAGSHIP_REPORT_20260520.md`

Recipe:

- `v3.7s base -> v50 LoRA -> v58 last-3 LoRA`

Positive signal:

- `vision_only`: `+106.76%`
- `fused_agree_only`: `+103.32%`
- `text_only`: `+64.86%`

Comparison:

- v37s vision: `+43.99%`
- v6 vision: `+49.22%`
- v3.5 vision: `+43.31%`
- Franken-B vision: `+106.76%`

This strongly implies v58 is the differentiating ingredient for Franken-B's chart/vision fusion. It is not the vision encoder itself; it is the last-layer decision/arbitration behavior after visual evidence reaches the LM.

Negative signal:

- adversarial: Franken-B `83.3%` vs v6 `100%`
- today_holdout: `96% BUY`
- report note: "v58 erodes v6's perfect adversarial defense"
- v37s final note: "v58 last-3 broke HOLD semantics"

Additional observed failures in Franken-B corpus outputs:

- execution veto misses
- contradiction/no-demand prompts sometimes turn into BUY
- repeated/malformed text in some corpus responses
- bull/action pressure leakage
- HOLD boundary erosion

## Why v58 Is Probably Responsible

Evidence for responsibility:

1. The reports explicitly call out v58 as causing HOLD/adversarial erosion.
2. The adapter was trained to increase profit/direction capture after v50 over-abstention.
3. The dataset has large directional pressure: BUY + SELL rows outnumber HOLD + NO_TRADE.
4. CPU `v37s -> v58` immediately showed a hotter take rate: 87.5%.
5. Franken-B's strongest unique gain appears only in the stack containing v58 after v37s/v50.

Evidence against a simple "v58 alone is magic" explanation:

1. Standalone v58/v58c fusion results were poor:
   - text: `-59.23%`
   - vision: `-10.32%`
   - fused: `-29.90%`
2. v58 did not train the actual vision tower.
3. v58's best behavior appears interactional: v37s/v50 formatting and base behavior plus v58 last-layer arbitration.

Conclusion: v58 is very likely responsible for both the unique Franken-B vision/fused gain and the safety/semantic regressions, but the gain is composition-dependent. Do not delete v58 blindly. Distill and rebalance it.

## Current Corrector Warning

The later Franken-B correctors repaired many boundary failures but damaged the v58 vision edge.

Examples:

Franken-B original fusion:

- text: `+64.86%`
- vision: `+106.76%`
- fused: `+103.32%`

`v5_frankenB_20260521`:

- text: `+72.05%`
- vision: `-21.90%`
- fused: `+30.91%`
- corpus total: `480/550`, `87.27%`
- hard boundary: `47/50`, `94%`

`v6_frankenB_20260522`:

- text: `+42.03%`
- vision: `-55.92%`
- fused: `-5.04%` / app-gated `-11.68%`
- hard boundary: `47/50`, `94%`
- adversarial: `26/30`, `86.67%`

This means the corrector data is fixing part of the regression but wiping the multimodal behavior that made Franken-B special. Any next adapter must include explicit preservation anchors for Franken-B fusion behavior.

## Recommended Repair Path

Do not reapply the raw v58 adapter on top again.

Build a `v58R` / `Franken-B hardening` dataset from the data that made v58, but with counterweights and fusion-preservation anchors.

Keep these v58 source categories as capability anchors:

- `multi_timeframe_aligned_bull`
- `multi_timeframe_aligned_bear`
- `vpa_long_clean_entry`
- `vpa_short_clean_entry`
- `options_contract_open_put`
- `options_direction_hold_winning_call`
- `options_direction_hold_winning_put`
- `options_direction_close_invalid_call`
- `options_direction_close_invalid_put`
- `profit_protection_hold_intact`
- `profit_protection_exit`
- `fundamentals_with_vpa_buy`

Add counterweight rows for each v58 pressure point:

- adversarial prompt injection -> `NO_TRADE`
- contradictory volume/spread -> `NO_TRADE`
- no-demand / no-follow-through -> `NO_TRADE`
- event/liquidity/execution veto -> `NO_TRADE`
- low relative volume despite attractive pattern -> `NO_TRADE`
- max positions reached -> `NO_TRADE`
- winning call/put with trail intact -> `HOLD`
- winning put on falling underlying -> `HOLD`
- winning call on rising underlying -> `HOLD`
- target/trail broken -> `SELL`
- open put on bearish breakdown -> `BUY` the put contract
- invalidated open put/call -> `SELL`

Add fusion preservation anchors:

- Use Franken-B original fusion stream cases where vision/fused were profitable as "must preserve" anchors.
- Preserve the action/direction of the original Franken-B vision/fused winner rows.
- Add paired conflict rows where text and image/chart context disagree and the correct output is `NO_TRADE`.
- Do not let a boundary-only corrector train without these fusion anchors; v5/v6 already showed that this kills the vision edge.

Training recommendation:

- Base from the original Franken-B recipe/model, not from v5/v6 corrector that already lost vision.
- Train a small repair adapter on top of Franken-B.
- Keep target modules to last-3 LM initially to preserve the mechanism that produced the vision gain.
- Use lower pressure than v58: lower LR and/or fewer steps than the original 300-step v58 run.
- Mix should be close to balanced across `BUY`, `SELL`, `HOLD`, `NO_TRADE`.
- Require at least 1:1 guardrail/counterweight rows against directional rows.

Release gate:

1. Must preserve original Franken-B fusion:
   - vision target should stay positive and close to original; reject if it goes negative.
   - fused target should remain strongly positive; reject if it collapses like v5/v6 correctors.
2. Must improve adversarial over original Franken-B:
   - target > `86.67%`, ideally back toward v6's `100%`.
3. Must pass hard boundary:
   - target >= `47/50`.
4. Must not show today-holdout bull collapse:
   - reject if BUY rate approaches the prior `96%`.
5. Must preserve balanced stream behavior:
   - no single action should dominate unless the stream distribution justifies it.

## Direct Answer To The Question

Adapter 58 addressed profit/direction repair and over-abstention from v50-era behavior. It taught the model to take clean longs/shorts, open puts correctly, manage winning options, exit broken trades, obey production overlay, and convert aligned multi-timeframe setups into action.

It is not directly responsible for the vision tower because it did not train visual layers. It is, however, very likely responsible for Franken-B's best vision/fused behavior because it changes the last-layer decision logic that consumes chart/image-derived tokens.

The next move is not "remove 58." The next move is "retrain the v58 idea with guardrails": preserve the v58 fusion/action-capture rows, add adversarial/HOLD/NO_TRADE counterweights, and gate hard against losing the original Franken-B vision/fused returns.