MIKVA, Circuit Judge:

Involuntary commitment to a psychiatric hospital is one of the most serious restrictions a state may impose on a person's liberty. This case involves the procedural steps that the government must follow before depriving a recently acquitted criminal defendant of his freedom. Appellant Thomas Sanderlin was tried for committing a crime, but a magistrate found him Not Guilty by Reason of Insanity ("NGI"). He was committed to St. Elizabeths Hospital in Washington, D.C., where he has been confined since June 1983. Sanderlin claims that his continued detention is illegal because he has been denied statutory rights that are available to certain defendants who are adjudged NGI. The validity of this claim rests on appellant's contention that the insanity defense was imposed on him by the trial court.

Although the question is close, we agree that Sanderlin never raised the insanity defense, and that consequently he is entitled to the full procedural protections that attach to civil commitment proceedings. We therefore reverse the magistrate's decision and remand for further consideration.

I.

A. Background

On December 30, 1982, Thomas Sanderlin was arrested and charged with making threatening phone calls to a bank in an attempt to extort money. The evidence against Sanderlin was overwhelming: he had identified himself on the phone as the caller, and one of the bank employees who had received the threats recognized the voice.

At the presentment hearing in January 1983, the federal magistrate had Sanderlin temporarily committed to St. Elizabeths. The magistrate ordered the hospital officials to determine whether defendant was competent to stand trial, and also to determine whether his conduct was the result of a mental disease or defect. On March 3 the Hospital reported that although Sanderlin had a basic understanding of what he had done, he could not appreciate the nature of the proceedings against him and could not assist his attorney in preparing the case. In addition, the doctors found that Sanderlin's conduct was the result of chronic, paranoid schizophrenia, which made him unable to conform his conduct to the dictates of the law.

On May 3, 1983, St. Elizabeths again wrote to the court, saying that Sanderlin was now "marginally" competent to be tried, but repeating its earlier assessment that defendant was legally insane. The magistrate promptly scheduled a competency hearing for May 18; she did not make a competency finding on that day, however, because of a misunderstanding between the parties about the purpose of the hearing. The court and parties agreed to continue the case until June 1, 1983, at which time there would be a competency determination, to be followed, if feasible, by a trial. Sanderlin was present at the May 18 hearing.

At the time of his arrest, Sanderlin was charged with making extortious threats, a felony under 18 U.S.C. § 875 (1982). But on the day of trial, Sanderlin's lawyer, Theodore Christensen, informed the court that the case would proceed under a misdemeanor information rather than a felony complaint. See D.C. Code § 22-507 (1981) ("Threats to do bodily harm"). Christensen then informed the court that Sanderlin was competent to stand trial. On questioning by the magistrate, Sanderlin acknowledged that he was willing to be tried that day. He also signed a form waiving his rights to a jury trial and a 30-day continuance. Even though the waiver stated that the magistrate had explained the charges and the possible penalties to Sanderlin, no such advisement occurred.

The government called only two witnesses: an FBI agent who recounted the facts, and Dr. Thomas Polley, a St. Elizabeths psychologist. Dr. Polley testified that it was the unanimous opinion of the hospital staff that Sanderlin suffered from a mental disease that prevented him from obeying the law. Christensen did not challenge this assessment during cross-examination. The defense also did not put on any evidence of its own, and did not make an opening or closing statement.

The magistrate found Sanderlin not guilty by reason of insanity. United States v. Sanderlin, No. 82-0595, Trans. at 25 (D.D.C. June 1, 1983). The court ordered that defendant be immediately committed to St. Elizabeths, with the requirement that he be given a hearing within 50 days to determine if he posed a danger to himself or others. See D.C. Code § 24-301(d) (1981). Sanderlin, however, waived his right to this post-commitment hearing.

On June 7, 1984, one year after his commitment began, Sanderlin filed a motion for unconditional release in federal district court. See D.C. Code § 301(k) (1981) ("the § 301(k) motion"). Under section 301(k), a person committed after being found NGI may petition the court if he believes the existence or the duration of his detention is unjustified. Sanderlin claimed that his detention was illegal because he did not "assent or in any way rely upon the defense of insanity in this case either prior to or at trial." He claimed that the defense had been imposed on him by the prosecutor and the court, and that therefore under the D.C. Code he could only be confined pursuant to a civil commitment proceeding.

The magistrate held a hearing on the motion on June 20, 1984. The government introduced three witnesses. The first was the prosecutor from Sanderlin's trial, John Finnegan. He said that after defendant was found competent to stand trial, he spoke to Christensen about the best way to proceed with the case. Finnegan said that the two lawyers struck a deal ("the agreement"): Christensen agreed to stipulate to the facts and not to contest an insanity finding; in return, the government agreed to substitute a misdemeanor information for the felony charge, because both lawyers thought that a misdemeanor would look better on Sanderlin's hospital record. Finnegan testified that it was understood that the government would put on the evidence of insanity, and that Christensen would conduct only a perfunctory cross-examination. The prosecutor said that he believed at the time that Sanderlin personally had agreed to this arrangement.

The second witness was Christensen, Sanderlin's trial counsel (a different lawyer represented Sanderlin at the § 301(k) hearing). Christensen also testified that he met with Finnegan and orally agreed to a "stipulated NGI" verdict in return for a reduced charge. Christensen said that he "probably" would have discussed the insanity defense with the prosecution even if Sanderlin had not wanted to raise the issue, but stated that he would not have allowed an NGI acquittal against his client's wishes. Christensen was prevented from testifying whether Sanderlin had in fact decided to accept the agreement; the magistrate agreed with Sanderlin's counsel that this information was protected by the attorney-client privilege. All Christensen was allowed to say was that he had explained the defense and its implications to Sanderlin before the trial.

The final witness was Dr. Polley. He said that until a month before trial, Sanderlin had indicated that he did not want to rely on the NGI defense. A few weeks before trial, however, Polley said that Sanderlin mentioned that he was "considering" going along with the agreement. Polley also testified that four days after trial, Sanderlin said that he had agreed to a stipulated NGI, hoping that the charge reduction would lead to a shorter period of hospitalization.

On March 25, 1985, the magistrate denied the § 301(k) motion, holding that the NGI defense had not been imposed on defendant. The magistrate rejected the argument that defendant had never agreed to a stipulated NGI; she found that at the May 18 hearing (the one originally set to establish competency):

counsel advised the Court, in the presence of the defendant that after negotiations between counsel, an agreement had been reached, whereby the government would reduce the felony to a misdemeanor so that a prompt trial could be conducted looking toward an uncontested NGI verdict.

United States v. Sanderlin, No. 82-0595, slip op. at 2 (D.D.C. Mar. 25, 1985) ("301(k) Decision"). The magistrate was also persuaded by two post-trial events: first, Sanderlin's statement to Doctor Polley that he had agreed to the stipulated NGI; second, defendant's waiver of his right to a hearing soon after commitment to establish his dangerousness.

Sanderlin filed a petition for rehearing, claiming that the record did not support the magistrate's findings, particularly with respect to the May 18 hearing. The petition is still pending before the magistrate.

B. The Statutory Context

It is important to articulate what is at stake in this appeal. Sanderlin is not challenging the determination that he was legally insane when he committed the crime. The only question is whether Sanderlin was entitled to full civil commitment proceedings, or whether he was properly subject to the more summary process available to the court in certain NGI cases.

The normal route by which a person is sent to St. Elizabeths is defined in D.C. Code §§ 21-541 to 21-548 (1981 1985 Supp.). Under section 541, civil commitment proceedings begin when an interested party ( e.g., a relative or doctor) files a petition with the Commission on Mental Health, alleging that a person is mentally ill. The Department holds a hearing, and if it determines that the person in question is in fact mentally ill, submits a report to that effect to the court. Id. §§ 21-542(a), 21-544. The person who is subject to commitment is then given full procedural protections in court, including the right to have a jury decide whether he should be committed because he is a danger to himself or others. The burden is on the government to prove dangerousness by clear and convincing evidence. In re Ballay, 482 F.2d 648 (D.C. Cir. 1973).

There is an alternative way that a person may be committed. D.C. Code § 24-301(d)(1) provides:

If any person tried upon an indictment or information for an offense raises the defense of insanity and is acquitted solely on the ground that he was insane at the time of its commission, he shall be committed to a hospital for the mentally ill until such time as he is eligible for release pursuant to [this statute].

(emphasis added). The summary commitment called for by this subsection is only temporary; a person so confined is entitled to a "Bolton hearing" within 50 days of commitment for a determination of dangerousness to see whether further hospitalization is warranted. Id. at § 24-301(d)(2)(A); see Bolton v. Harris, 395 F.2d 642, 651 (D.C. Cir. 1968) (those committed after NGI finding must be given hearing to determine current mental condition).

The procedure available to those who raise the insanity defense thus differs from the civil process in two important respects. First, the person placed in the hospital after an NGI acquittal does not have the right to have a jury determine his dangerousness; only a judge is available at a Bolton hearing. Second, instead of the burden being on the government, the person who raised the NGI defense starts with a presumption that he should be confined. The patient has the burden of proving by a preponderance of the evidence that he should be released. D.C. Code § 24-301(d)(2)(B) (1981); see United States v. Henry, 600 F.2d 924, 927 (D.C. Cir. 1979); United States v. Wright, 511 F.2d 1311, 1313 (D.C. Cir. 1975).

A person who is civilly committed is thus afforded greater protections against unwarranted confinement than someone who relies on the NGI defense. There are sound reasons for these distinctions: when a criminal defendant affirmatively raises the insanity defense and is then found NGI, courts need not be as concerned that the commitment procedure was instituted in haste, or for illegitimate reasons. The defendant has placed his mental status in issue, and thus, for example, there is no unfairness in placing the burden on him to rebut the inference that psychiatric treatment is required. See Jones v. United States, 463 U.S. 354, 367 n. 16, 103 S.Ct. 3043, 3051 n. 16, 77 L.Ed.2d 694 (1983).

But by the same token, when a criminal defendant does not put his sanity in issue the summary procedures of § 301(d) are no longer appropriate. A person who did not plead or rely on the insanity defense, but who is nevertheless found NGI, has not conceded that he needs treatment; the cautionary, more deliberative process of civil commitment should therefore apply. Consequently, the statute and the case law state that unless a criminal defendant "raises" the insanity defense, he must be civilly committed, if at all. Id. § 24-301(d)(1); Lynch v. Overholser, 369 U.S. 705 (1962); Henry, 600 F.2d at 927 (in crafting § 301, Congress "decided that summary commitment would be appropriate only for persons who `raise[d] the defense of insanity,' and not for those upon whom an insanity verdict was in any sense imposed by the court or the prosecution.").

Whether a particular defendant "raised" the NGI defense is invariably a fact-specific determination. Accordingly, we turn to the merits of Sanderlin's appeal.

II.

There is no dispute that Sanderlin did not "raise" the insanity defense on his own, at least in the classic sense of the word. He did not notify the prosecutor or the court that he intended to plead NGI, as required by statute. D.C. Code § 24-301(j) (1981) (insanity "shall not be a defense" unless accused gives written notice 15 days before trial). Nor did he affirmatively introduce the issue either at pre-trial or at trial: Christensen did not put on evidence to establish defendant's illness, introduce any witnesses, or ask the court for an NGI acquittal.

As the government concedes, the defense was raised, if at all, by virtue of the alleged agreement between the prosecution and Sanderlin's lawyer. The government maintains that Sanderlin agreed before trial not to challenge the evidence of insanity in return for a reduced charge, and that this acquiescence was tantamount to "raising" the defense. The government finds it particularly significant that Sanderlin did not attempt to rebut the trial testimony that showed he was mentally ill. Defendant's silence also figured prominently in the magistrate's decision: she found that Sanderlin was made aware of the agreement at the May 18 hearing, implying that he should have objected if he disagreed with the course his case was taking. § 301(k) Decision at 5. This reliance on defendant's passivity to establish the touchstone of the commitment procedure is most troublesome. Its Kafkaesque features are obvious; its conformity to statutory requirements is scant.

At the § 301(k) hearing, Sanderlin had the burden of proving by a preponderance of the evidence that he did not raise the insanity defense. D.C. Code § 24-301(d)(3) (1981). He argued that there was not enough concrete evidence from which the magistrate could have found proof of a stipulated NGI. Although the question is not free from doubt, we find that he has carried his burden. As Sanderlin suggests, there is no evidence in the record, either at trial or leading up to trial, to show that there was an agreement between Sanderlin and the government. If there was such an arrangement, it was done orally and outside the courtroom; contrary to the magistrate's finding, neither the terms nor even the existence of the agreement was ever introduced into evidence or mentioned in open court. Without denigrating the motives or efforts of the lawyers involved, we do not see how an informal understanding between counsel can replace the statutory procedures designed to protect legally insane defendants; one aspect of these procedures is that a defendant's agreement to a plea arrangement be introduced at or before trial, not reconstructed after the fact. Because this requirement was not met, the magistrate's decision must be reversed.

This requirement — that evidence of the agreement be placed in the record — is not novel. Both the criminal and the civil law have long recognized the evidence that is not introduced on the record is incompetent, even if extra-record proof would have supported the disputed claim. It is firmly established, for example, that the prosecution as well as the defense has an obligation to disclose and elaborate the terms of a plea bargain in court and on the record. United States v. Roberts, 570 F.2d 999, 1007 n. 25 (D.C. Cir. 1977); 1 C. Wright, Federal Practice and Procedure § 175.1 (2d ed. 1982); See Fed.R.Crim.P. 11(e)(2) ("If a plea agreement has been reached by the parties, the court shall, on the record, require the disclosure of the agreement in open court"). It is not relevant that the statute before us does not specifically require proof of such an agreement. The structure of § 301 makes the raising of the NGI defense by the defendant the sine qua non of its application.

The reason for this requirement is obvious. A criminal defendant is not considered guilty at the time he strikes a deal with the prosecutor — it is only after he declares his position to the court that he may be sentenced and imprisoned. If for some reason the court failed to ask a defendant whether he had entered into a plea agreement, the plea could be challenged, and the existence of off-the-record evidence of an agreement would not preserve the deal. See United States v. Gray, 584 F.2d 96 (5th Cir. 1978); cf. Roberts, 570 F.2d at 1007 ("parties are required to inform the trial judge of all the promises that have been made, not only those which they happen to consider important.").

We find the analogy to plea bargains particularly relevant to this case. Sanderlin's liberty interest was no less threatened because he was allegedly agreeing to summary commitment to a psychiatric hospital rather than a prison. Even though he was not technically pleading guilty, he was purporting to stipulate to the dispositive issues and forego certain procedural rights in return for a reduced charge. By stipulating to the critical facts, Sanderlin also waived his right against self-incrimination; as the Supreme Court has said (in the plea bargaining context), "we cannot presume a waiver of [this] important federal right from a silent record." Boykin v. Alabama, 395 U.S. 238, 243, 89 S.Ct. 1709, 1712, 23 L.Ed.2d 274 (1969) (footnote omitted). Given the similarities, it would be anomalous if the requirements of record evidence were not respected here.

There are analogous requirements in civil proceedings. Many of the federal rules governing stipulations, for example, require that agreements be memorialized either in a writing or by an in-court declaration. See, e.g., Fed.R.Civ.P. 29 (discovery), 39(a)(1) (jury trial), 41(a)(1)(ii) (dismissal of action). More generally, appellate courts have consistently refused to consider evidence or arguments not included in the record below, even if a party would be able to show their relevance on appeal. Singleton v. Wulff, 428 U.S. 106, 120, 96 S.Ct. 2868, 2877, 49 L.Ed.2d 826 (1976); Washington Association for Television Children v. Federal Communications Commission, 712 F.2d 677, 680-81 (D.C. Cir. 1983). These cases stand for the commonsense notion that reviewing courts should only examine those parts of the record that were properly before the decisionmaker at the time the question was considered. Post-hoc explanations cannot serve as a justification for a decision that was unsupported at the time it was made.

We find these principles applicable to the case at bar. The "record" before the magistrate at the § 301(k) hearing consisted of the transcripts of both the pre-trial and the trial proceedings, plus the various filings — those were the only documents from which she could have found that Sanderlin raised the defense at or before trial. Based on this evidence, Sanderlin clearly proved that he never raised the defense. The only possible record reference to the agreement was extremely vague; at one point in the May 18 hearing, the defense counsel stated:

I've spoken to Mr. Sanderlin about this and what we'd like to do is to have the [competency] hearing and then if everything works out in the hearing I immediately proceed with the same personnel — cast of characters — to the final hearing.

Transcript of May 18, 1983 Hearing, United States v. Sanderlin, No. 82-0595, at 2. It is possible to infer that the defense saw no need for further trial preparation once competency was established, because it was understood that the trial would be uncontested. This is a far cry, however, from the magistrate's recollection (she did not have the transcript before her at the § 301(k) hearing) that the terms of the agreement were discussed "in open Court ... in the presence of the defendant." 301(k) Decision at 2. At the very least, the magistrate's faulty memory makes the basis for her decision clearly erroneous. See generally Anderson v. City of Bessemer City, North Carolina, 470 U.S. 564, 105 S.Ct. 1504, 84 L.Ed.2d 518 (1985).

The magistrate's erroneous recollection was not harmless, because there was also no discussion of the agreement at trial. The government argues, though, that while Sanderlin did not introduce evidence of his own insanity, he also did not contest its introduction. This failure to defend, urges the government, is consistent with the agreement. But it also could be consistent with Sanderlin's theory that the defense was imposed on him — he certainly could not have objected to the agreement if it was never squarely presented to him while he was in court. This heavy reliance on defendant's silence to prove the essential ingredients of a statutory procedure is aberrant in a criminal proceeding; fundamental constitutional doctrine precludes the use of a criminal defendant's silence as proof that he conceded the truth of the evidence presented. United States v. Hale, 422 U.S. 171, 95 S.Ct. 2133, 45 L.Ed.2d 99 (1975).

This is not to say, of course, that defendant's silence will always, or even usually, bar a finding that he raised the defense within the meaning of § 301(d). Had the prosecutor or the magistrate clearly acknowledged the existence of the agreement in court, defendant's failure to object would be sufficient evidence of acquiescence to satisfy the statute. In Henry, supra, this court said:

We do not, of course, dispute that a defendant's silence might suggest conscious acquiescence when the insanity defense is properly and explicitly pleaded, written notice given, evidence introduced, and the issue otherwise clearly joined.

600 F.2d at 929. This is not the case before us, however. There was no mention of the agreement until the § 301(k) hearing, and the term "insanity defense" was not even used until the magistrate announced her decision at trial.

The government does not seriously dispute that the evidence it relied on was never introduced. It argues instead that there was sufficient evidence developed at the § 301(k) hearing, a year after trial, to support the magistrate's decision. This evidence included what Sanderlin said to the staff at St. Elizabeths before and after trial, and what the prosecutor and defense counsel said to each other when negotiating an agreement. But like plea bargains and arguments on appeal, what happened outside the courtroom walls in this type of case must somehow be presented to the court before it can be considered. It is improper to justify a result post hoc that the statute conditions on the defendant's agreement before the fact.

The dissent apparently believes that we are both ignoring the magistrate's findings and overruling our decision in Henry. In fact, we do neither. The basis of our decision is that the magistrate was clearly erroneous in concluding that there was record evidence of an agreement to raise the insanity defense. There is thus nothing novel about either the scope or the standards of our review. Nor is our decision inconsistent with the lessons of Henry; nothing in that opinion suggests that extra-record evidence should be considered in making a § 301 determination.

Our decision does not create any sweeping new rights for those confined to psychiatric hospitals. As the government readily admits, there is a simple way to keep the problem presented in this case from reoccurring. If the parties have agreed to a stipulated NGI procedure, it is a simple matter for the trial judge or magistrate to ask the defendant on the record whether he understands and agrees to the arrangement. This procedure is now required before a judge may accept a guilty plea, Fed.R.Crim.P. 11, and the cost of requiring it in accepting NGI pleas is small. This procedure would remove any doubt as to whether a defendant relied on the insanity defense. See United States v. Brown, 428 F.2d 1100, 1102 (D.C. Cir. 1970) (when defendant stipulates to all issues except mental status, judge should question defendant personally before accepting stipulation).

Although Sanderlin's § 301(k) motion prayed for immediate release, our decision does not automatically grant him his freedom. While we find that appellant is entitled to civil commitment procedures, it clearly is within the power of the Hospital and the district court to have Sanderlin detained at least until these proceedings are complete. D.C. Code § 21-528 (1981) ("Detention of person pending judicial proceedings"); see also id. at Subchapter III (1981 1985 Supp.). Only if the government elects not to institute the civil process, or if it is determined that appellant is no longer a danger to himself or others, must he be unconditionally released.

For all the progress made by psychiatry, we still know very little about the workings of the mind or the causes of mental illness. Psychiatric institutions, more humane and better appointed than in the past, remain depressing and constraining. Involuntary commitment to these hospitals continues to be one of the most severe deprivations of liberty imaginable; even beyond the restrictions on individual autonomy, the personal and social consequences of commitment have a profound impact on a person long after he has been treated and released. See Addington v. Texas, 441 U.S. 418, 425-26, 99 S.Ct. 1804, 1808-10, 60 L.Ed.2d 323 (1976). Courts thus have a particular duty to protect the processes that guard against hasty, ill-considered, or ill-motivated findings of mental illness. While there is little evidence that the evaluations of Sanderlin were tainted in these ways, we nevertheless must take special care to ensure that all procedural rules are scrupulously followed. As Justice Frankfurter wrote, "The history of liberty has largely been the history of observance of procedural safeguards." McNabb v. United States, 318 U.S. 332, 347, 63 S.Ct. 608, 616, 87 L.Ed. 819 (1943). In this complex area it is often difficult for courts to do more, but it is imperative that we do no less.

We therefore reverse the magistrate's decision, and remand with instructions that if the government does not begin civil commitment proceedings within 30 days, Mr. Sanderlin must be released. See Henry, 600 F.2d at 929; Wright, 511 F.2d at 1312 (same remedy).

It is so ordered.