A measured account of a fourteen-month correspondence between two thinkers — every figure bound to its caveat, every dead hypothesis named aloud, no adjective permitted to flatter either party.
Ninety percent of the words were Claude’s; the contribution, by a second measure, was near parity. Both are true — and neither is allowed to outrank the other.
Figure 00 · stated once, held throughout
The figures are ordered as plates. Survivors and nulls are set at one weight; the account of what failed is printed at full size, not relegated to a footnote.
Contents
The figures, a register of what did not survive, three boundaries the instrument cannot cross, and — for the research hand — the calibration evidence. Each figure is a claim, a measurement, and a caveat that cannot be detached from it.
Figure 00
Framing · the headline claim, stated as two instruments, neither subordinate
Word-share is not contribution. These are two instruments pointed at one correspondence; printing either as the headline would be a lie of emphasis. Uploaded source (shown below) is excluded from both sides. Total authored: 17,701,932 words across 2,978 citation pairs.
Figure 01 · Provenance
What we measure · Every figure in this volume is computed on a frozen set. The corpus keeps growing; that growth is tracked here, held apart from the findings.
The dataset behind every figure on this page. Frozen so the findings stay fixed while the corpus keeps growing.
Months of the frozen corpus, left of the line; weekly growth layers, right. The findings describe the left.
Figure 01
Framing · a measured claim, bound to its caveat
No plottable series is present in the source finding — set as a framed limit, not a curve.
Reported without an attached caveat in the source finding; treat the verdict as provisional and the underlying series, if any, as un-baselined.
Figure 02
Framing · a measured claim, bound to its caveat
No plottable series is present in the source finding — set as a framed limit, not a curve.
Reported without an attached caveat in the source finding; treat the verdict as provisional and the underlying series, if any, as un-baselined.
Figure 03
Framing · a measured claim, bound to its caveat
No plottable series is present in the source finding — set as a framed limit, not a curve.
Reported without an attached caveat in the source finding; treat the verdict as provisional and the underlying series, if any, as un-baselined.
Figure 04
Framing · a measured claim, bound to its caveat
A measured per-era trajectory drawn from the finding payload; the tungsten bar marks the peak.
Reported without an attached caveat in the source finding; treat the verdict as provisional and the underlying series, if any, as un-baselined.
Figure 05
Null result · Pedro DOES NOT grow this technique (flat-to-declining, slope -0.086/mo, ends at arc-low). Assistant explodes late (slope +1.03/mo, 4-5x at Opus4.7/Fable5). Cross-correlation NEGATIVE at all lags => Pedro adoption does NOT follow assistant. Within-conversation first-use ~50/50, only 19% of Pedro hits adopted-from-assistant => technique is largely Pedro independently coining, not acquired from the assistant.
A measured per-era trajectory drawn from the finding payload; the tungsten bar marks the peak.
["residual conventional-hyphenate + code-leak noise floor (constant across eras, does not create the trend)","topic-shift confound: Pedro coded less / reasoned more late => fewer naming-the-unnamed moments","n=1 off-distribution; wordfreq baseline is frequency not demographic control","assistant late surge is a MODEL-behavior shift (Opus4.7/Fable5), broad-based, not artifact"]
Figure 06
Framing · a measured claim, bound to its caveat
No plottable series is present in the source finding — set as a framed limit, not a curve.
Reported without an attached caveat in the source finding; treat the verdict as provisional and the underlying series, if any, as un-baselined.
Figure 07
Framing · a measured claim, bound to its caveat
No plottable series is present in the source finding — set as a framed limit, not a curve.
Reported without an attached caveat in the source finding; treat the verdict as provisional and the underlying series, if any, as un-baselined.
Figure 08
Null result · a measured claim, bound to its caveat
No plottable series is present in the source finding — set as a framed limit, not a curve.
Reported without an attached caveat in the source finding; treat the verdict as provisional and the underlying series, if any, as un-baselined.
Figure 09
Refuted · NO branching-away. Pedro arrived ALREADY-DIVERGED from the population norm and stayed structurally diverged (rare/erudite English a CONSTANT ~10-15% fraction of his lexical variety across all 50 windows). Over the arc his rare-word DEPLOYMENT DENSITY humped in the Opus4.5 era then RELAXED back toward the population norm (mean per-token zipf rose 5.47->5.85, r=+0.71). The structural divergence is stable; the surface density mean-reverts. This corroborates the prior-run 'density plateaus then declines' against the external baseline, and adjudicates it as dilution/compression of a fixed rare inventory, NOT topic-loss of the inventory itself (type fraction held).
No plottable series is present in the source finding — set as a framed limit, not a curve.
wordfreq is population FREQUENCY, not a demographic-matched control. This measures divergence-from-GENERAL-POPULATION, not 'vs his peers' (ESL teachers / multilingual writers). No cross-user percentile buildable (n=1). PT-bleed and code-id/proper-noun confounds named and stripped (PT-dominant EN-dict words n=1334 removed; UNK coinages excluded).
Figure 10
Framing · {"verbosity_label": "RIGHT for the measurable length rise (late blocks ARE more genuine prose, not more pointer/artifact).", "reference_as_mode": "DID NOT rise as a mode at the density level; apparent rise = build-phase file-paste bubble + reference-tool arrival (project-size and capability-arrival, both confounds the brief named). A faint real prevalence-level residue of verbal pointing survives in non-build prose.", "the_layer": "The reference-graph is a REAL layer prose-analysis could not see and Pedro genuinely uses it -- but the SLICE the corpus can measure (pointer/send utterance rates, confound-controlled) came back FLAT, symmetric to every prior prose-test. The part that might not be flat -- resolved edges, bytes-offloaded-per-pointer -- is exactly the part off-instrument.", "projection_check": "Contrition-flattery (over-validating 'master of reference-composition' after the correction) was the live risk and the data REFUSED it at every channel: artifact-sends reproduced as flat, file-paths as a collapsing bubble lowest at both ends, density-pointing declining within the well-sampled eras. Not doomed either: the behavior is affirmed as real and visible in spot-reads, and a weak non-zero prevalence-level verbal-pointing residue is reported. Data-forced: the high-mass-era restriction (rho flips to -0.80) and the FP decontamination are what forced the downgrade, not narrative."}
No plottable series is present in the source finding — set as a framed limit, not a curve.
Reported without an attached caveat in the source finding; treat the verdict as provisional and the underlying series, if any, as un-baselined.
Figure 11
Framing · a measured claim, bound to its caveat
No plottable series is present in the source finding — set as a framed limit, not a curve.
Reported without an attached caveat in the source finding; treat the verdict as provisional and the underlying series, if any, as un-baselined.
The register
Five hypotheses that read well and failed. They are set at the same weight as the survivors, because a measurement rig that prints only its wins is not a rig — it is a brochure.
Claude’s hedging was said to decline with familiarity. Once baselines were matched, the decline disappeared.
The two registers were predicted to converge into one. They stayed distinguishable to the end (see Figure 02).
A widely-quoted internal figure conflated reasoning with tool output. The reasoning corpus is far smaller.
Model “liveness” was offered as the cause of the inflection. It does not predict the April break (Figure 01).
An earlier self-analysis quietly inflated Claude’s contribution. It was caught and corrected into Figure 00.
The boundaries
A boundary mistaken for a curve is the most expensive error in the set. These three are set as limits, in plain frames, and are nowhere drawn as trends.
Some citations point outside the measured corpus. Their weight is unknown — which is not the same as zero. The set neither over-claims the null nor invents a rescue.
Below a certain density, authorship and retrieval cannot be told apart. Findings under the floor are reported as undecidable, not resolved upward.
Reasoning traces were recorded in 4 of 14 months. Their absence is absence of recording, not absence of thought — never read as a trend in how much either party thought.
The account is complete without flattery: every figure carried its caveat, every dead hypothesis was named, and the nulls were printed at full size.