-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathCITATION.cff
More file actions
121 lines (121 loc) · 7.28 KB
/
Copy pathCITATION.cff
File metadata and controls
121 lines (121 loc) · 7.28 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
cff-version: 1.2.0
message: "If you use PRISM in your research or work, please cite it as below."
title: "PRISM: A Framework for LLM Research and Audits"
abstract: >-
PRISM is a methodology for keeping multi-prompt, multi-session LLM research
and audit work coherent across prompts, sessions, and the context limits of
any single chat. It splits a research problem into atomic specialist prompts,
dispatches each where it runs best across Claude, ChatGPT, Gemini, and
Perplexity, and converges their outputs into a single living document called
the Master. v2 introduces the PRISM Lens Library — a reference catalog of
audit-scope lenses graded against the draft Prompt Strategy at Setup to
surface silent omissions before any prompt ships. v2.1 adds machine-readable
frontmatter, a strength × polarity normativity vocabulary, and an
observation-driven appendix indexing vendor parsing behaviour on dispatched
content. v2.4 generalizes the reserved execution-mode flag into three
orthogonal orchestration axes — surface, execution driver, and persistence —
each a closed validated enum; v2.5 lands the repo-backed persistence
mechanics, holding durable state in a GitHub repo with the running
What's-next artifact made repo-resident as the cross-surface pickup point.
v2.6 adds corpus-access dispatch (Phase 1): a first-class investigation
posture for retrieving from named sources and document corpora, kept
structurally distinct from epistemic triangulation, with a coverage fan
that merges returns for recall and a mandatory exhibits manifest carrying
per-item provenance.
v2.7 adds recommended-sources-on-lens: an optional Lens Library field
binding framework-curated external reference sources, each with a
mandatory framing and recency caveat, to a lens's material question.
v2.7's consuming behaviour lands in v2.8 as corpus-access Phase 2: a
lens-anchored auto-trigger that recognizes when a lens's recommended
source answers the question being worked and proposes a lookup for
operator ratification rather than auto-dispatching, auto-populating the
framing and recency caveats so they cannot be silently omitted.
v2.9 declares the orchestration substrate by vendor and tier (Claude,
Opus-class) rather than pinned model versions, and decomposes the framework
into a lean always-loaded core plus on-demand reference bundles, packaged as
a Claude Skill plugin alongside the canonical single-file form.
v2.10 adds the Scope-Integrity Test: a coverage-time falsifier gate that
enforces that a lens's minimum scope binding was actually satisfied, not
waved through, on a rigor ladder from inline self-check to cross-vendor
probe, with an optional per-lens sharpened-probe field; the embedded Lens
Library advances to v0.16.
v2.11 adds an explicit Skill slash-command invocation path on Claude —
`/prism-start <subject>` activates PRISM and begins Setup for the supplied
subject — a Claude-Skill packaging affordance carried by a plugin slash-command,
alongside the unchanged natural-language invocation that remains the portable
form on every vendor.
v2.12 adds a companion `/prism-whats-next` command that resumes an engagement
from the Master's What's-next artifact.
v2.13–v2.14 add the output-discipline family — the uninvited-frame rule
(the Elephant Rule), plain-words / recompute / claims-carry-their-basis /
editions-stand-alone principles, and an Independent Validation Dispatch.
v2.15 ("Foundations") adds two Standing Principles — trust-structure /
self-report-advisory and surface-translation — a retrieved-versus-inferred
finding-basis axis, filename-authoritative vendor reconciliation, Setup
stakeholder-positioning, and a Setup-artifact re-audit.
v2.16 ("Operating-model backbone") adds the lanes / roles operating model —
every engagement runs as parallel lanes worked by sessions in a role ×
context-tier matrix — with the PRISM Desk and PRISM Meta standing lanes and
an operator-facing PRISM UI; an append-only cross-lane open-items inbox with
commit discipline; bundle-load integrity; Setup onboarding (a Setup-generated
per-engagement system-instructions file plus a first-class quick mode); and
M5 self-banding.
v2.17 ("Dispatch round-trip") names the dispatch round-trip as one bounded
construct — the five-stage dispatch lifecycle (build, dispatch, execute,
return/converge, reconcile) — with a late-bound build that requires a pass
spec rather than a frozen envelope, a proactive return-handling card and a
failure leg routed by type, a unique dispatch identifier paired with the
copy-through digest so a return reconciles to the exact dispatch instance, a
fanned Independent-Validation adjudicate-by-re-derivation refinement, and
passive pre-fill of self-report.
v2.18 ("Lifecycle & deliverables") adds a symmetric engagement-closure gate — a
three-layer close sweep bookending three-layer readiness, with an orphan-sweep,
a reconcile-at-close codification sweep, and deliverable polish — and codifies
the engagement deliverable: a comprehensive verdict-first final report, an
optional interactive workbook, a presentation house-style, deliverable
bump-atomicity (shipped equals validated), and an external-share family (one
repo per engagement, a de-coded share archive, and an image-redaction
procedure).
v2.19 ("Core-slimming refactor") slims the always-loaded core by projecting
its phase-, seam-, and condition-scoped mechanics — Setup, currency
maintenance, session continuity and recovery, and corpus-access — into four
on-demand reference bundles keyed to the role that needs them, each reachable
through an in-core trigger stub and the phase → bundle manifest; the canonical
single-file form is unchanged.
v2.20 re-embeds the Lens Library at v0.16 — adding the reader-respect lens
"Does this respect the reader?", which flags filler that wastes the reader's
time and completes the document-review family, and curating external
recommended-sources on the comparison and unit-economics lenses — and names
that lens in the Independent Validation Dispatch default kit.
v2.21 ("External-review hardening") hardens the methodology from external
review — turn-close evaluate-then-emit ordering, a Self-check
visible-recompute line and a first-token Dispatch-ID echo, a clean split
between the orchestration dispatch record and the self-contained execution
seat paste, and a heavy-line-versus-XML dispatch-delimiter rationale — and
adds a runtime front door: a /prism-help orientation command and an
orientation-before-engagement core behavior.
The framework codifies a triple execution contract
(Envelope · Self-check · Output), Monitors, Standing Principles, a telemetric
context-pressure framework, and continuous Master + What's-next state —
distributed as a single markdown file that can be attached to any LLM session
or installed as a Claude Skill.
authors:
- family-names: Kuper
given-names: Ron
alias: Ronkupper
repository-code: "https://github.com/Ronkupper/PRISM"
url: "https://github.com/Ronkupper/PRISM"
license: CC-BY-4.0
version: "2.21.0"
date-released: "2026-06-29"
keywords:
- llm
- prompt-engineering
- research-methodology
- audit-framework
- lens-library
- multi-prompt
- multi-session
- context-engineering
- claude
- anthropic