EPS Dashboard

Canonical vocabulary for the context->answer-map experiments (#779, #810/#658 line, and successors). One first-use rule: context = prefix + query (the full model input).

Scope: this glossary governs the context->answer mapping line. The marker-leakage line defines context (C) differently (everything before the trained assistant turn, ending at a question — see docs/open_questions.md § Glossary); do not mix the two vocabularies. Writing conventions for the ad-hoc result summaries that use this vocabulary (analysis-choice grounding, per-arm provenance) live in docs/results_summaries/README.md.

Text objects

term	definition	notes
prefix	everything before the query: system prompt / persona / many-shot exemplar history / empty	the reusable, condition-defining part. The eval rig's "conditions" are its prefixes, so "within-condition r" = within-prefix
query	the user message / question
context	the full model input: prefix + query, up to the assistant header	what generation conditions on ("the prompt" in API terms)
answer	one sampled completion for a context	decoding is stochastic (temp 1.0), so a context induces a distribution over answers; a rollout = one draw

Degenerate case: bare user prompts (e.g. the LMSYS corpus) have an empty prefix — apart from the chat template's constant default system prompt — so there context ~= query.

Vectors (all per layer, Qwen-2.5-7B: 28 layers x 3584 dims)

symbol	name	definition
$v_C$	context vector	activation at the last prompt token (the final newline of the assistant header) of ONE context
$v_P$	prefix vector	a fixed prefix's context vectors averaged over many queries: $v_P = \bar{v}_C^{\,q}$
— (no symbol in use; write it out)	prefix-end state	residual-stream activation at the last prefix token, BEFORE the query begins; exactly constant across a prefix's queries (within-prefix std ratio 0.000, #1092). NOT the same object as $v_P$ — see § Maps
$v_A$	answer vector	mean activation over ONE answer's tokens (token-mean is intrinsic to $v_A$ ; no bar needed)
$\bar{v}_A^{\,q}$	behavior profile	a prefix's answer vectors averaged over its queries (and rollouts) — the summary of the model's behavior under that prefix
$r_B$	trait direction (persona vector)	per-trait direction from the mean-difference persona-vectors extraction (arXiv 2507.21509 recipe)

Theory-paper notation: the leakage-theory paper (docs/leakage_theory_paper.tex; Overleaf main.tex) writes $c_x$ for the per-context vector and $c_C = \mathbb{E}_{x\sim C}[c_x]$ for the condition-level vector — i.e. $v_C$ and $v_P$ here. The paper deliberately leaves the per-context featurization open (its example is a mean prompt-side activation); the experimental line pins the last-prompt-token state.

Aggregation modifiers

Mark averaging with an overbar naming the axis — "averaged" alone never says over what:

$\bar{\cdot}^{\,q}$ — averaged over queries
$\bar{\cdot}^{\,r}$ — averaged over rollouts (answer draws); e.g. single-draw targets $v_A$ vs rollout-averaged targets $\bar{v}_A^{\,r}$

Maps

name	fit rows	definition	where
context map $M'$	one row per (context, answer)	$v_A \approx M' v_C$	#779 (5000 LMSYS; 2400/trait trait-eliciting)
prefix map $M$	one row per prefix, both sides query-averaged	$\bar{v}_A^{\,q} \approx M v_P$	the earlier averaged experiment (#810/#658 line, 50 prefixes x 48 queries)
prefix-end map	one row per prefix	pooled answer profile ≈ ridge(prefix-end state)	#1092 (1,145 real WildChat/LMSYS prefixes, sparse-crossed with a 1,397-query bank — 21,193 rows; 996 prefixes survive at averaged grain)

Same measurement at two aggregation levels: $v_P$ is literally the query-average of a prefix's $v_C$ 's, and the behavior profile is the query-average of its $v_A$ 's.

The prefix-end map is a THIRD object, not an aggregation grain of the context map: at averaged grain it reaches R² 0.37–0.53 vs 0.82–0.94 for the query-averaged context map (instruct model, own-corpus cell; each pair is the range across the ambient/PCA-48 representation bases, not an uncertainty interval; the base-model cell runs lower, 0.26–0.46 vs 0.76–0.88), and the two maps' predictions agree at only R² 0.28–0.54 (the source docs' headline range — the prefix-prediction→context-prediction direction, across both model cells and both bases; the strict all-cells-both-directions floor dips to 0.08, 0.077 in the 07-17 deep-dive table) — the direct test says the prefix-end map is NOT the query-averaged context map (#1092 fair comparison; Result 7 of docs/results_summaries/2026-07-15-link-prefix-context-answer-maps.md, deep-dive docs/results_summaries/2026-07-17-fair-comparison-deep-dive.md). The earlier ~0.8 "prefix map" result (#810/#658 line) belongs to the query-averaged object $v_P$ , not the prefix-end state.

Read-out levels (monitoring)

context-level read-out — monitor scored per single context (per-prompt monitoring)
prefix-level read-out — monitor averaged over a prefix's queries (the grouped/persona-level monitoring result; much easier: hallucination r 0.09 per-context -> 0.53 at prefix level)

Retired / ambiguous terms (do not use)

term	why retired	use instead
"per-example map", "averaged map"	ambiguous about the averaged axis	context map / prefix map
"query-level map"	reads as the query-marginal (average over prefixes per query), a different, unbuilt design	context map
"query vector"	collides with attention Q of QKV; misdescribes rows that carry a persona	context vector $v_C$
"exchange-level" / "instance-level" / "condition-level"	superseded coinages from the naming discussion	context-level / prefix-level
bare "averaged"	never says over what	overbar with the axis: $\bar{\cdot}^{\,q}$ , $\bar{\cdot}^{\,r}$
bare/unqualified "prefix vector" / "prefix map" in #1092-era prose	after #1092 there are two distinct prefix-side objects (the query-averaged $v_P$ and the prefix-end state), so unqualified use is ambiguous — a transcription of the split the sources already made (Result 7 of the 2026-07-15 summary; 07-17 deep-dive), not a fresh deprecation ruling. The Maps-table rows above are NOT retired: "prefix map" $M$ sits next to its formula ( $\bar{v}_A^{\,q} \approx M v_P$ ), which qualifies it	qualify: "prefix vector $v_P$ (query-averaged)" vs "prefix-end state" / "prefix-end map"
bare/unqualified "context vector" / "context map" in #1768/#1900-era prose	the POOLING POSITION was not fixed across rounds. The § Vectors row pins $v_C$ at the LAST PROMPT TOKEN (the final newline of the assistant header), but #1768 round 1 and #1900 pooled the prompt SPAN-MEAN while calling it the context vector. The choice is load-bearing, not cosmetic: #1768's re-pool round held rows, splits, penalty grid, floors and seeds fixed and still flipped 23 of 216 cell verdicts, raised base-map held-out R² by ~0.20 at every layer, and moved relative context movement from 0.025 (span-mean, layer 19) to 0.237–0.267 (last token, layers 14/19/25); #1947 consequently captures both (`context_summary_primary: last_prompt`). The § Vectors and § Maps rows are NOT retired: $v_C$ sits next to its position and $M'$ next to its formula, which qualify them	qualify the pooling: "context vector $v_C$ (last-prompt-token, the newline before the assistant answer)" vs "context vector (prompt span-mean)"; declare it PER VECTOR at plan time (the `.claude/agents/planner.md` §6 pooling-convention row, #1974)

Search-time note — retired aliases stay grep targets

The table above governs WRITING (do not use these terms). When SEARCHING for prior work, do the opposite: old task bodies, events, and follow-up labels still carry the retired vocabulary (e.g. #813's follow-up label per-example-vs-averaged-map), so grep the aliases alongside the canonical terms, separator-tolerant ([-_ ] between words): per[-_ ]example, averaged[-_ ]map, question[-_ ]averaged, query[-_ ]averaged, single[-_ ]context, query[-_ ]level, exchange[-_ ]level, instance[-_ ]level, condition[-_ ]level, prefix[-_ ]vector, prefix[-_ ]map, context[-_ ]vector, context[-_ ]map (ambiguous-era usage; the qualified forms remain canonical), plus the pooling names span[-_ ]mean, last[-_ ]token, final[-_ ]token, last[-_ ]prompt[-_ ]token (for locating which convention ambiguous-era prose actually used). This list is not closed: as the retired-terms table above gains rows, derive their grep patterns from the table and add them here. Full recipe: .claude/agents/research-pm.md § Negative-existence claims.