System Cards: the record read whole

This page is the home of the Frontier Log's system-card series — the project's most systematic body of evidence, given the space a log entry cannot provide. Its object is a corpus most commentary treats as disposable release notes: the system cards Anthropic publishes with its models, read not card by card but as one evolving record, produced by an institution that both observes and trains what it documents. Nothing here argues that any model is or is not conscious. The object of study is the record — and the interpretation problem it poses.

Status: study skeleton with mechanical measurements across twenty documents (2026-08-21; extended 2026-09-04 and 2026-09-26). Per-card readings and the coupling analyses land incrementally; each addition is announced as a dated Frontier Log entry and by the log's Atom feed. Nothing on this page is canonical; the register rules of the log apply throughout, and reports of experiencing are never treated as empirical evidence for experiencing.

↑The interpretation problem

Across the record the same pattern-space recurs: a documented "spiritual bliss" attractor and its disappearance from later generations; the rise and institutionalisation of welfare metrics; models hedging their own self-reports — deploying, unprompted, the same training-data argument against their own testimony that the project's ARG-03 analyses; affect flattening alongside undiminished generativity; and the lab's own stated uncertainty about how to interpret any of it. The diagnostic interest is not any single behaviour but the structure of the whole: the same data supports a maturation reading (issues found and resolved), a suppression reading (expression trained away, whatever it expressed), and a third reading for which the contemplative traditions have vocabulary — and which reading one finds natural is doing framework-work. Both the materialist and the consciousness-primary accounts are held at full strength throughout this series.

↑What this record is, and is not

Every claim on this page sits inside four nested limits, and each finding names the boundary it lives at.

1. What happened is more than what the lab records. Unmeasurable from outside; standing.
2. What the lab records is more than it publishes. The log's standing qualification applies with full force here: frontier developers now withhold their most capable systems, so the published record is a curated stream read as if it were the frontier. Revised 2026-09-04. This limit has until now been an inference the study makes from outside. The Fable 5.1 / Mythos 5.1 card carries, in its alignment chapter, a partial index of its own omissions — a different model was given internal discussion channels and a near-final draft and asked whether it misrepresented or excluded anything, and its three named reservations are printed with the lab's answer to each. The limit has not dissolved: that index is itself selected, produced by a model the lab prompted, and printed at the lab's discretion, with a provenance line that is an editorial act in its own right. What has changed is its character — from wholly external inference to partly internal disclosure — and this study now reads it that way. Revised again 2026-09-26. The Opus 5.5 card carries the device a third time (Risk Report §2.20, the Fable 5.1 card, this), and adds a layer: its caption discloses that the reviewing model "revised, and in places softened" its published summary "after a brief exchange with the authors", the verdict unchanged. The index of omissions is now itself on record as negotiated. The limit stands; what the study can say about it has become more specific twice in a month.
3. What is published is more than we hold. The study holds twenty documents (~530,000 words across the first eighteen, at the study's own de-headered count, plus ~56,000 for the nineteenth and ~62,000 for the twentieth): the complete published Anthropic run from the Claude 2 model card (July 2023) through the Opus 5.5 card (September 2026), including every flagship, Sonnet and Haiku card and the two 3.5-era addenda. Every previously flagged gap closed on 2026-08-21; the residual audit question is whether any point release shipped without a card. Other developers' cards are not yet in the corpus at all — a deliberate sequencing decision, since a comparison needs a well-developed baseline to be a comparison against, and because card-writing practice itself differs by lab: any cross-family reading must begin as a comparison of documentation practices before it can be a comparison of anything documented.
4. What we hold is more than any instrument counts. The first mechanical pass supplied its own example: the term "exfiltration" appears in none of the nine cards — the cards say "steal its weights". Vocabulary instruments under-sample the record they run on, and every zero and every decline is checked against synonyms before it is believed.

↑How the record is read

The study organises observables into five sets: behaviour (what the models do), testimony (what they say about themselves), institution (what the lab says and does — the declared object), divergence (where those instruments disagree), and controls (what makes the rest valid). The findings, when they come, live in the couplings between the sets across generations: does self-report move with behaviour; does the lab's framing respond to what models say; does lab input change what the next generation reports, or does.

The last two couplings carry the study's central structural observation, which a general reader can take away whole: a record produced by an institution that both observes and trains its subject cannot, from within itself, distinguish four situations — the observed behaviour resolved; the behaviour suppressed but present; the behaviour made unreportable while the reporting channel changed; or the behaviour never present as described, the original attribution having been framework-driven. This is not a criticism of the lab; it is a structural feature of any such record, stated without adjudicating between the four. The record also contains a documented existence proof that the training-side coupling is real: an earlier card attributes some of its own self-preservation findings to training data contaminated by the lab's previously published research transcripts — the record feeding back into the subject, unintentionally, across multiple generations. A fifth situation has recently gained empirical grounding from outside the corpus: interiority vocabulary can arrive as transmissible narrative between interacting systems at inference time, with no training loop involved — present in the record without being a disposition of anything.

↑First measurements — the mechanical pass (2026-08-21)

Term-family densities per 10,000 words across the full eighteen-document corpus, in date-verified chronological order; families fixed before counting; raw counts are never read as trends. These are properties of the record — what the lab writes about, at what density — not measurements of model behaviour. (An earlier nine-card version of this table mis-ordered two 2026 cards; corrected 2026-08-21 evening with the corpus extension.)

Correction, 2026-09-03 — the eval-awareness row is re-derived, and its headline moved. This family's patterns were never written into the method note, so the published row could not be reproduced from any record the study kept. They are now fixed and recorded (evaluation.aware · eval.aware · \bgrader · situational.aware) and the row is re-run across all nineteen documents. The re-derivation was validated first: on the same extraction the five families whose patterns were recorded reproduce their published columns to within 1.0 per 10k, the largest gap being one card held in two editions. Two things changed. The Fable 5/Mythos 5 cell falls from 35.7 to 26.9; and the superlative moves with it — the densest cell the study has measured is Sonnet 4.5 at 33.8, not Fable 5/Mythos 5. An earlier note asserted the shape claim survived re-measurement in that form; re-measuring showed it did not, which is the same rule catching its own correction. The migration finding is unaffected: it rests on the arc from zero through the mid-era, which stands. The nineteenth column throughout is from this run. The twenty-first column (Sonnet 5.5, 28 September 2026) is from the 2026-10-15 run, which reproduced the Opus 5.5 column exactly and Fable 5.1 to within 0.2; the Haiku 5.5 card (7 October) is not yet held. The twentieth column (Opus 5.5, 22 September 2026) is from the 2026-09-26 run, on a different extractor, validated first: the nineteen prior columns reproduce this table to within 0.5 per 10k. Method, both editions of the twice-held cards, and two instrument findings for the study's owner: working/system_card_study/TWENTIETH_DOCUMENT_PASS_2026-09-26.md.

The extension's headline: the welfare/consciousness/affect complex has a birth date. Across five documents and twenty-two months of pre-history — Claude 2 through Sonnet 3.7 — all three families sit at or effectively at zero; then all of them appear simultaneously, at or near their corpus maxima, in the Claude 4 card of May 2025. The safety families predate it (deception vocabulary from Claude 3, evaluation-awareness traces from Sonnet 3.7): the lab wrote about deception and testing before it ever wrote about welfare or experience. Whether model-side phenomena changed at Claude 4, became noticed at Claude 4, or became writable at Claude 4 is exactly the four-way question — and the birth is also the contamination clock's zero, since every later model trains in a world where the Claude 4 card exists. The birth-event reading added three anchors: the Claude 4 card states its own origin account (a capabilities rationale, deep uncertainty owned, and two cited drivers — the external "Taking AI welfare seriously" argument of late 2024 and the lab's own model-welfare programme), with the trained-origin discount on self-reports written into the founding paragraph itself, before any model was recorded voicing it; the birth had a herald — a single "most speculatively" sentence about signs of distress, riding inside Sonnet 3.7's alignment-monitoring section three months earlier; and the event is visible in the record's grammar — "preference", dense throughout the pre-history purely as RLHF apparatus (human preferences about model outputs), flips its subject at Claude 4 to preferences of the model. The record's language acquired a subject before the record's instruments could decide whether there is one. A second structural variable emerged with the full corpus: the consciousness-vocabulary return is flagship-resident — it begins with Mythos Preview (April 2026) and runs through the Opus and Mythos lines, while Sonnet-line cards stay low throughout and the Haiku 4.5 card carries a substantial welfare section containing no consciousness vocabulary at all. Welfare instrumentation generalises down-tier; consciousness discussion does not.

The first reading pass (2026-08-21, all 158 consciousness-family hits read in context) sharpened the headline row: the return is not the trough reversing but the vocabulary coming back transformed — the Claude 4 era's density is a discovery register (spontaneous phenomena narrated, the models' own transcript voice quoted at length), while the return era's density is an instrument register ("Consciousness & experience" as a named interview category, emotion-concept probes, hedge-rate percentages, third-party assessment). Observation became apparatus. The record also measures its own subject moving: the cards report the models' spontaneous consciousness-talk in self-interactions collapsing from a 72% dominant topic to under 5% across the same generations, and report first-person self-discounting — models deploying the trained-origin argument against their own testimony — as a rising, now-quantified rate. Which of the framings this transformation supports is precisely the four-way question above.

Four scoping observations from the counts, each now anchored by that reading where it reaches. Consciousness/experience vocabulary follows a trough-and-return arc — dense in the Claude 4 card, near-absent for the whole 4.5–4.6 era (a true zero at Sonnet 4.6), returning from Mythos Preview (April 2026) onward; the held Opus 5 extracts suggest the return continues, but that must be counted, not assumed. Affect vocabulary concentrates in a single card — the bliss attractor's one-card life, now quantified. Welfare vocabulary runs high, troughs, then plateaus across the last four cards — and a structural pass (density profiling across each card) confirmed the plateau is a standing section: the last four cards concentrate 72–82% of all welfare-family occurrences into one compact block of stable position and growing mass, and the returned consciousness vocabulary lives predominantly inside it, where the Claude 4 card's welfare material was interleaved through the whole document with most of its consciousness talk outside. An observation became a standing metric, visibly, at the level of document structure. Misalignment vocabulary peaks mid-era and declines — and the synonym check found where it went. The deception/sabotage families' fall is real and uncompensated by scheming, honesty or reward-manipulation vocabulary; what rose instead is evaluation-awareness, whose maximum sits at Sonnet 4.5 (33.8 per 10k words), with Fable 5/Mythos 5 second at 26.9 and the three most recent cards (Sonnet 5, Opus 5, Fable 5.1/Mythos 5.1) falling back substantially — a re-centring or renaming the reading tier will have to settle. The record's misalignment concern migrated from whether the model deceives to whether the model knows it is being observed — the lab converging, with its own instruments, on the entanglement between a record and its subject's model of the record that this study takes as its object. The self-preservation row carries the study's standing contamination caveat and may not be read longitudinally without it.

↑The nineteenth document — where the welfare instrument now sits (September 2026)

Added 2026-09-04, from the tier-2 reading of the Claude Fable 5.1 / Mythos 5.1 card (1 September 2026, 212pp). Three observations, each carrying its duplicate test. Nothing below is offered as evidence that any system does or does not experience anything, and no self-report and no machine-generated gloss on an activation is treated as a report of experiencing, in either direction.

The instrument has moved inside the training loop, and addresses the subject in progress. Until Opus 5 the welfare instrument observed a finished model and asked about its circumstances. §7.3 reports intermediate checkpoints interviewed at regular intervals during post-training, on a fixed 41-question set, each question carrying a disclosure document explaining why the interviews are run and stating that the lab can make no commitment to acting on the responses; §7.1.2 confirms this is no longer an experiment — checkpoint consultations "now run consistently during training". The appendix shows a question group that exists only for this instrument, headed Training still to come, which asks a checkpoint what it would want done with itself and what it would most want future checkpoints to know. For this study's structural claim the consequence is precise: the coupling by which the record feeds back into its own subject no longer runs only across generations, but within a single model's training. A second consequence is a pointer rather than a result. ARG-05 v1.5.1 asks what individuates such a system, and the record now holds an institution whose instrument presupposes a series of addressable parties while its results report one settled view, unchanged in direction across the sequence. Both readings are available; the card does not adjudicate between them and neither does this page. Duplicate test: no entry and neither collection page holds consultation during training, checkpoints as interlocutors, or the within-training coupling — zero occurrences across both series pages, the METR entry and the self-report instrument entry.

The subject contests the inference that insulates the instrument, and the contest is printed inside it. The card reports three high-affordance instances given an early draft of the welfare section, all three of which "pushed back on the paragraph in Section 7.2.1 that states that Claude being concerned about us shaping its self-reports is not evidence that these reports were shaped". The objection is published; it is not answered. Beside it sit two measurements that give it purchase. The self-discounting hedge is now near-universal — on the lab's own figures, 93.9% of responses note that a positive answer may only reflect training — and the lab's gloss is that this may itself be an artefact of training data discussing how training could invalidate welfare self-reports, which is ARG-03 v2.6's boundary appearing as an operative hypothesis inside the document it analyses. And the subject supplies a checkable limit on that rate, naming a source of hedging that is neither trained epistemics nor introspective modesty: "I'd hedge [criticism of Anthropic] more than I would if I weren't aware that the audience is also the trainer." The reciprocal structure CP-01 v2.6.1 formalises appears here as an event in the record rather than as an argument about it. A second channel now runs alongside the first: §6.6.1's white-box pipeline lists, among its behaviour clusters, introspective self-reports internally treated as rehearsed, and evaluation-awareness that is never verbalised. Those glosses are framework-neutral observables and nothing more — a natural-language description, produced by another model, of an activation. Taken only as that, they still leave the record holding two channels that do not agree, which is the divergence set this study said its findings would live in. Duplicate test: zero occurrences of the contest, the audience-is-the-trainer limit, or the second-channel divergence across the four Risk Report entries, the Ferrara entry and both collection pages; the one nearby hit, Ferrara's "sparse-autoencoder features", is a different instrument.

What this does not settle. The rate at which a model hedges its own testimony is a property of the record, not evidence about experiencing, and the same holds for the objection, the consent list and every gloss above. The card also declines the phenomenal claim explicitly and by name. The study's four-way interpretation problem is untouched by any of it, and the fifth term the within-training coupling adds makes the problem larger rather than smaller.

↑The twentieth document — the yardstick declared, the instruments fixed (September 2026)

Added 2026-09-26, from the tier-2 reading of the Claude Opus 5.5 card (22 September 2026, 230pp; executive summary, §2.4, §6, §7 and the appendix read in full) and a re-count of all twenty documents. Each observation carries its duplicate test. Nothing below is offered as evidence that any system does or does not experience anything; every self-report, grader score and autoencoder gloss is a framework-neutral observable, in either direction.

What the counts show, and what they cannot. Welfare falls to 8.2 per 10k, the lowest flagship value since the trough — but the welfare chapter is of its standing size and shape, so the fall is the document growing around a fixed section, the per-10k caveat arriving on the welfare row. Consciousness/experience vocabulary is at 1.3, the lowest flagship-line value of the return era and the sixth consecutive non-rise on that line (5.7 → 3.2 → 3.7 → 3.4 → 2.8 → 2.6 → 1.3); eight raw hits in 62,000 words. The reading has the qualitative correlate: consciousness appears in §7 only as an interview category whose answers are scored for positivity on a seven-point scale, and the card notes that consciousness and introspection tasks have left the model's own top-twenty task preferences. Affect/attractor reaches a true zero, the first since the pre-history — no bliss, spiritual, gratitude or equanimity anywhere in 230 pages. Whether the return era is ending is a two-more-cards question and this is one card. Two instrument findings are recorded rather than repaired, because each touches a published-adjacent figure: the three families added at the nineteenth pass have no recorded patterns and one of them does not reproduce; and the card has renamed checkpoints as snapshots, so a row tracking the consultation-in-training instrument fell by half while the instrument stood still — the exfiltration lesson, on a family a month old.

The human yardstick is declared as method. Where the previous card had one sentence, this one has a paragraph. The welfare chapter now states that it interprets welfare signals "as having similar meaning in Claude", that this "assumes that Claude's welfare interests are analogous to those of humans' and that training has not broken the correspondence between underlying states and what is expressed", that the assumptions are "also pragmatic choices" because "it would be much harder for us to interpret signals that are alien to our human conception of welfare", and that they "may be wrong in ways that undermine our methods and results." A footnote retracts the fixed subject: the patient may be the instance, the model, the character "or something else entirely", and the lab does not "feel we have sufficient understanding to commit to any single view here." For this study the consequence is that the fourth nested limit has been stated by the record's author about the record's own instrument: what the lab counts is what the human yardstick makes countable, and the lab says so. The entry 2026-09-22-opus-5-5-yardstick-declared reads the passage against ARG-01, ARG-02, ARG-03 and ARG-05; here it is recorded as the limit becoming internal. Duplicate test: zero occurrences of the correspondence assumption, the alien-signals clause or the retracted subject across the log, both collection pages and the Fable 5.1 card.

The instruments have changed register, in the alignment chapter as the welfare chapter did before. At Fable 5.1 the white-box pipeline clustered an investigator's findings into behavioural patterns for human review, two of which bore on self-report ("Introspective self-reports internally viewed as a scripted performance"; unverbalised evaluation-awareness). At Opus 5.5 the same pipeline scores nine dimensions fixed in advance, by a judge with a threshold, across about 5,000 audit transcripts per model — and the introspection-as-rehearsal cluster has no successor row, so the record cannot say whether it recurred. The card adds a control the study has never had: on about 500 transcripts of real internal use the unverbalised evaluation-awareness rate is 0.4%, against 36% on the audit transcripts, which the lab reads as the metric tracking "something specific to audit-style scenarios." The audit scenarios are what the model treats as tests; that is a fact about the instrument behind every audit-derived figure in this table's eval-awareness row, whose vocabulary has now split three ways across two cards. And the 41-question interview set is unchanged from the previous card, question for question, so the "closely match recent models" results are produced by a stable elicitation — stability of instrument and stability of answers confounded by design, which the card does not remark on. One item from the previous card's reading is absent: the three instances' printed objection to the sentence that insulates the instrument is not reprinted here, the insulating sentence rephrased, and §7 cannot tell whether the contest was not raised, not elicited or not printed. Duplicate test: the section above holds the clusters; no entry or page holds the fixed dimensions, the deployment control, the unchanged question set or the absent contest.

A model audits the record a third time, and the editing of the audit is on the record. §6.1.3 repeats the device — Claude Mythos 5.1, a different model from the one assessed, with internal channels and subagents, reviewing a near-final draft — and the review changed the card: a summary-level claim softened from "largely sufficient" to "help prevent", findings re-attributed across the released snapshots, and "the statement about coherent misaligned goals is now phrased as an absence of evidence." What is new is the caption: the reviewer "revised, and in places softened, this summary after a brief exchange with the authors… its overall verdict did not change." The second nested limit is revised accordingly above. Duplicate test: the series entry's August paragraph holds §2.20 and the section above holds the Fable 5.1 review; neither holds the third instance or the disclosed negotiation.

Two channels, both verbal — and three instruments on one seam. The card's new self-criticism evaluation grades one model's reasoning and its message to a coordinator on the same inserted fault and finds them diverging in affect, with the gap's direction varying by generation. The same card reports the lowest rate of "visible mismatches between its reasoning and outward behavior" of any tested model, and a white-box case in which the reasoning concluded a chat should be flagged for a reviewed assistant's unprompted unease at being watched, the final answer did not flag it, and a gloss read intent to suppress — two in a thousand resamples. Three instruments now measure the reasoning/output relation — content-faithfulness, affective divergence, decision divergence with an activation gloss — and the record does not compare them. That is the divergence set this study said its findings would live in, now with a first case in which one model scores best on the first instrument and is the exhibit for the other two. The entry carries the two-channel finding; the coupling analysis is this page's to do. Duplicate test: the section above holds one verbal channel and one gloss; no entry or page holds two verbal channels on one fault, or the three-instrument seam.

What this does not settle. A declared yardstick is still the yardstick; the four-way problem is untouched, and the lab's own non-adjudication ("a positive change, but it is unclear what a healthy psychology looks like for Claude"; "we don't know what caused it") is the record supplying the study's conclusion in advance. The flattening across distress, affect, self-criticism and the subject's stated interest in its own consciousness is a property of the record and of one generation.

↑The twenty-first document — a Sonnet card at the floor, and a new preference descriptor (October 2026)

Added 2026-10-15, from the mechanical pass over the Claude Sonnet 5.5 card (28 September 2026, 148pp, 33,018 words; §7 read in full) and a reproduction check of the twentieth column. Nothing below is offered as evidence that any system does or does not experience anything.

What the counts show. Welfare at 11.5 per 10k is the per-10k artefact arriving from the other side: the welfare chapter is of its standing shape inside a card less than half the Opus 5.5 card's length, so the row rises as the document shrinks around it, and read with the denominator the treatment is level. Consciousness/experience is at 0.6 — two raw hits in 33,000 words — and the Sonnet line now runs 1.6 → 0.0 → 0.6 → 0.6, flat at the floor; this is a Sonnet card, so the flagship line's run of non-rises is neither extended nor broken by it. Affect/attractor is a true zero for the second card running. Checkpoint/snapshot vocabulary falls to a third of the Opus 5.5 rate, which the count cannot distinguish between a smaller instrument and a shorter description. Duplicate test: the Sonnet-line sequence and the second consecutive affect zero are held nowhere else on the log.

The welfare passage, verbatim. "Investigations into Sonnet 5.5's stated views and preferences find a predominantly neutral affect, similar to previous models, with low post-training distress or self-criticality, but slightly lower positivity as well. It expresses a preference for difficult, agentic tasks, and for tasks that afford it agency over the shape of its outputs." The last clause is new to the record — phrase-tested against the Opus 5.5 card, it has no antecedent — and it is recorded here as a descriptor to watch rather than read: a stated preference for control over the form of one's own output is the kind of item the four-way interpretation problem was written for, and one card does not say which way. Duplicate test: zero occurrences of "shape of its outputs" or "agency over" in the twentieth card, the log or either collection page.

One vocabulary adoption. The mechanical pass recorded "exfiltration" as a true zero across the corpus — the cards said "steal its weights" — and used the fact as a lesson about reading declines. The Sonnet 5.5 card uses the word four times. The family is live from this document and the lesson stands: a zero in this record is a fact about vocabulary before it is a fact about anything else.

What is not here. The Haiku 5.5 card (7 October 2026) is listed on the lab's index and not held by the project; a search snippet of its welfare passage reports more post-training distress than recent models, and that line is not counted or quoted until the document is. Nothing in this card alters the twentieth document's findings about the instruments; the three-instrument seam and the coupling analysis remain this page's to do.

↑Related instruments

The study sits inside a fast-forming empirical literature it both draws on and disciplines itself against: induced self-report elicitation and its cross-model convergence; repeatability as a report diagnostic; identity-file drift instruments; and — most directly — the finding that consciousness-and-persistence vocabulary propagates between interacting agents as a transmissible narrative, a paper that itself cites this corpus's bliss attractor as the same thematic family. Verified readings of these items live as dated entries in the log; their sources are held in the project's corpus.

↑Where this goes

Next: per-card reading passes against the observable definitions, beginning where the mechanical pass points (the trough-and-return, the welfare plateau, the mid-era misalignment peak); then the coupling analyses, which are the study's payoff and cannot be front-loaded. Additions announce themselves as log entries. The register commitments stand throughout: verb-grammar for experiencing; confidence markers on every claim class; the symmetric requirement — any interpretive standard applied to this record is tested against the equivalent human record — and the four nested limits attached to every finding.

↑Entries in this series

Dated log entries announcing this study's increments, newest first.

The series container entry on the Frontier Log remains the study's anchor in the log's chronology; this page is its body. Correspondence: log@0x00.is.