Correcting Is Not Reusing: What a Failed Attempt Contributes Beyond the Verdict

A rejected answer leaves a retry two things, the verdict and the transcript that produced it, and self-correction scores them together. We separate them on exact-answer mathematics with a retry protocol in which both sides enforce the verdict. Both retries use the same output-token ceiling and a controller that discards every draw in the rejected answer class, with every solver draw billed. They differ in whether the failed transcript is retained or replaced by a withholding notice, with the instruction worded to match, and the difference in success is the retention contrast. In four instruction-tuned models not trained for retry it is negative: keeping the transcript lowers success by .106 to .235 (.235 on Llama-3.1-8B, 161 problems). The penalty persists after final-answer redaction: under one instruction at matched prompt length, the redacted transcript stays .074 to .260 below a content-free placeholder. With the transcript in context, the share of first calls that return to the rejected answer class rises by .17 to .50, and the controller's redraws recover only part of the loss. Among nine complete retry policies priced against a fresh draw, our full-history continuation baseline is the worst in every family and ceiling. LoRA supervised fine-tuning on retry prompts lifts the kept transcript above the withheld one on a fixed failure set in each of three seeds, yet end-to-end success stays below the untrained model's (.616 against .739, first seed). Under a one-step revision protocol, S2R's self-correction RL raises the correction rate while the revised retry still succeeds less often than a fresh draw from the same checkpoint. Correction rate, retention contrast and end-to-end success give different verdicts and must be reported separately.