Proof row
No productivity number is published
The public case labels the result as a measured pilot and explicitly avoids broad ROI or productivity lift claims.
Supports claim discipline and prevents the comparison from being overstated.
Comparison · Measured pilot
Cold start vs. system-backed work
A team wants to know whether a new agent produces better work when it starts from the project system instead of a bare task prompt.
comparison pilot evidence public summary sanitized 3 evidence rows 8 evidence values
comparison diagram
Cold prompt vs. system-backed lane
Prompt Generic plan Repeated assumptions
Project state Decisions Reusable surface Scoped action
HERMES case-study frame
Read the case before the evidence room.
A reviewer should be able to answer seven things before opening the JSON export: what decision is in front of them, what changed against a cold start, which row carries the claim, where the value appears, what source trail exists, what purchase decision it informs, and what the case does not prove.
- 01 · Buyer question
What is being decided?
A team wants to know whether a new agent produces better work when it starts from the project system instead of a bare task prompt.
- 02 · Cold start
Cold run
The agent received a competent task prompt but no project memory, decision trail, or reusable surface inventory.
- 03 · KBPD-backed
System-backed run
The agent read the project state, decision records, and role guidance before producing the same class of artifact.
- 04 · Boundary
Measured pilot
The result supports a launch proof, not a statistical productivity claim.
Best public row no-productivity-number-published Publicly inspectable
- Evidence value
- Claim boundary: pilot Good enough to inspect; not a broad numeric ROI claim.
- Source trail
- Comparison matrix review table · Public projection
- Buyer decision
- Read how the system works The visible difference is not prose polish. It is continuity: the second run starts with project memory and keeps the work attached to it.
- Limit
- Measured pilot The result supports a launch proof, not a statistical productivity claim.
Evidence room
Inspect the evidence, values, and limits.
This is the public review surface for the case. Each row names the observation, the evidence values it stands on, the source trail, the boundary, and the matching structured entry in the evidence export.
Review order Question -> evidence -> value -> source -> limit Start with the case question, then inspect what the claim is allowed to stand on.
Evidence export /proof/cold-start-comparison/evidence.json The same evidence rows are emitted as JSON for independent review.
Claim boundary Measured pilot The page cannot claim more than this status supports.
Source availability
A source trail is not always a raw public artifact. Public projection means the source has been redrawn or summarized; retained source means it is named for audit but withheld until redaction clears.
- Public projection
- 1
- Retained source
- 2
- Pending run
- 0
Public evidence excerpts
Readable snippets tied to evidence rows
These excerpts are the public bridge between retained raw sources and the evidence table below. Each one says what was preserved, what was redacted, and which evidence rows it supports.
- Retained-source summary
paired-scenario-review-excerptPaired scenario review excerpt
Five anonymized work scenarios were reviewed in paired form: one prompt-only run and one project-system-backed run for each scenario.
- Source label
- Comparison matrix
- Redaction note
- Source run text remains retained; the public excerpt preserves the comparison structure without task names or private workspace details.
Supports evidence rowspaired-scenarios-reviewedsystem-backed-outputs-carry-continuity
- Public excerpt
claim-boundary-review-excerptClaim boundary review excerpt
The public case is labeled as measured pilot evidence and does not publish a productivity-lift number.
- Source label
- Public proof case
- Redaction note
- This excerpt is visible on the public page and constrains the sales claim.
Supports evidence rows
- Evidence rows
- 3 Reviewed public observations
- Evidence values
- 8 Metric and run values exposed for review
- Source trail
- 3 Artifacts named for review
- Claim status
- Measured pilot Claim boundary for publication
Claim under review
System-backed runs preserve prior decisions, reuse existing structures, and surface clearer next actions than cold runs on the same scenario set.
The visible difference is not prose polish. It is continuity: the second run starts with project memory and keeps the work attached to it.
01 Five paired scenarios were reviewed Public projection evidence_facts[0]
- Observed
- The pilot evidence set contains five anonymized tasks run once from a prompt-only starting point and once from a system-backed starting point.
- Evidence
- Run metadata lists five paired scenarios; the comparison matrix is retained as the public projection source.
- Supports
- Supports the claim that the comparison is based on paired cases, not unrelated examples.
- Scenario set: 5 Five anonymized work scenarios were compared as cold and system-backed runs. Value handle: data-01-scenario-set
- Claim boundary: pilot Good enough to inspect; not a broad numeric ROI claim. Value handle: data-03-claim-boundary
- Sample: Five paired scenarios Sample size or unit under review. Value handle: data-05-sample
- Provenance: Public projection of internal comparison runs Where the public projection says the evidence came from. Value handle: data-06-provenance
Small sample size.
02 System-backed outputs carried named context Public projection evidence_facts[1]
- Observed
- The system-backed summaries include inherited decisions, named reusable assets, and explicit limitation language that the cold summaries did not consistently carry.
- Evidence
- Before/after excerpts and metric notes on the proof detail page name reuse, inherited decisions, and explicit limitations as the primary reviewed signal.
- Supports
- Supports the narrow claim that continuity improved in the inspected pilot set.
- Primary signal: reuse Cases were reviewed for inherited decisions, named reusable assets, and explicit limitations. Value handle: data-02-primary-signal
- Claim boundary: pilot Good enough to inspect; not a broad numeric ROI claim. Value handle: data-03-claim-boundary
- Reproducibility: Representative excerpts are public; raw transcripts remain internal until fully redacted. What a reviewer can repeat or inspect publicly. Value handle: data-07-reproducibility
- Claim status: measured pilot The maximum strength of the public claim. Value handle: data-08-claim-status
Anonymized scenarios only.
Evidence id
system-backed-outputs-carry-continuity Review URL #fact-system-backed-outputs-carry-continuity 03 No productivity number is published Publicly inspectable evidence_facts[2]
- Observed
- The public case labels the result as a measured pilot and explicitly avoids broad ROI or productivity lift claims.
- Evidence
- Claim status, metric note, and limitations all state that this is pilot evidence rather than a statistical productivity result.
- Supports
- Supports claim discipline and prevents the comparison from being overstated.
- Claim boundary: pilot Good enough to inspect; not a broad numeric ROI claim. Value handle: data-03-claim-boundary
- Provenance: Public projection of internal comparison runs Where the public projection says the evidence came from. Value handle: data-06-provenance
- Claim status: measured pilot The maximum strength of the public claim. Value handle: data-08-claim-status
The result supports a launch proof, not a statistical productivity claim.
Source refs
Evidence values
The values this proof is allowed to stand on
Field Value Meaning
Scenario set Value handle: data-01-scenario-set 5 Five anonymized work scenarios were compared as cold and system-backed runs.
Primary signal Value handle: data-02-primary-signal reuse Cases were reviewed for inherited decisions, named reusable assets, and explicit limitations.
Claim boundary Value handle: data-03-claim-boundary pilot Good enough to inspect; not a broad numeric ROI claim.
Run date Value handle: data-04-run-date June 2026 Date label attached to this public proof case.
Sample Value handle: data-05-sample Five paired scenarios Sample size or unit under review.
Provenance Value handle: data-06-provenance Public projection of internal comparison runs Where the public projection says the evidence came from.
Reproducibility Value handle: data-07-reproducibility Representative excerpts are public; raw transcripts remain internal until fully redacted. What a reviewer can repeat or inspect publicly.
Claim status Value handle: data-08-claim-status measured pilot The maximum strength of the public claim.
Source trail
Where the public proof can point
-
source-01-cold-outputsCold outputs run transcript set · internal source retained -
source-02-system-backed-outputsSystem-backed outputs run transcript set · internal source retained -
source-03-comparison-matrixComparison matrix review table · public projection
Evidence status mix
What kind of support is each row?
- Public projection
- 2
- Publicly inspectable
- 1
Metrics
Signals carried by the case
- 5 Scenario set Five anonymized work scenarios were compared as cold and system-backed runs.
- reuse Primary signal Cases were reviewed for inherited decisions, named reusable assets, and explicit limitations.
- pilot Claim boundary Good enough to inspect; not a broad numeric ROI claim.
Source artifacts
Where the evidence comes from
- Cold outputs run transcript set · internal source retained
- System-backed outputs run transcript set · internal source retained
- Comparison matrix review table · public projection
Case quality review
This case has a cold-review quality score.
This is a cold-review rubric, not a sales claim. It checks whether the case can be understood, inspected, bounded, and acted on.
Overall 4.8 Launch-readable
Cold-audience clarity
5.0A reviewer gets the situation, contrast, and buyer decision without reading the JSON.
Evidence visibility
5.0The case exposes proof rows, values, source handles, and excerpts before the JSON export.
Limitation clarity
5.0The page names what the case proves and what it does not prove.
Visual hierarchy
4.0The route uses visuals, storyboards, or surface frames to break the proof into inspectable layers.
Purchase motivation
5.0The proof sends the buyer to an evidence room, artifact link, pricing path, or local run.
Next depth gate
Visual hierarchy: Promote the proof row into a diagram, table, screenshot projection, or openable artifact.
Open case evidenceEvidence docket
Review the evidence rows behind the case.
- Evidence rows
- 3
- Public projection
- 2
- Publicly inspectable
- 1
- 01 Public projection
Five paired scenarios were reviewed
- Observed
- The pilot evidence set contains five anonymized tasks run once from a prompt-only starting point and once from a system-backed starting point.
- Evidence
- Run metadata lists five paired scenarios; the comparison matrix is retained as the public projection source.
- Supports
- Supports the claim that the comparison is based on paired cases, not unrelated examples.
- 02 Public projection
System-backed outputs carried named context
- Observed
- The system-backed summaries include inherited decisions, named reusable assets, and explicit limitation language that the cold summaries did not consistently carry.
- Evidence
- Before/after excerpts and metric notes on the proof detail page name reuse, inherited decisions, and explicit limitations as the primary reviewed signal.
- Supports
- Supports the narrow claim that continuity improved in the inspected pilot set.
- 03 Publicly inspectable
No productivity number is published
- Observed
- The public case labels the result as a measured pilot and explicitly avoids broad ROI or productivity lift claims.
- Evidence
- Claim status, metric note, and limitations all state that this is pilot evidence rather than a statistical productivity result.
- Supports
- Supports claim discipline and prevents the comparison from being overstated.
Evidence Flow
Read the proof as a chain, not a claim.
Each case has to survive the same path: starting condition, system-backed change, visible result, retained evidence, stated limitation, and a next action.
- Before Cold run prompt-only output
- Change System context enters comparison
- After System-backed run project-system output
- Evidence 3 source artifacts June 2026
- Limit 3 named limits Measured pilot
Claim
System-backed runs preserve prior decisions, reuse existing structures, and surface clearer next actions than cold runs on the same scenario set.
The visible difference is not prose polish. It is continuity: the second run starts with project memory and keeps the work attached to it.
Before / after
Same use case. Different starting conditions.
Before · prompt-only output
Cold run
The agent received a competent task prompt but no project memory, decision trail, or reusable surface inventory.
Generic plan, generic risks, and repeated assumptions. Useful as a first draft, but disconnected from prior choices.
After · project-system output
System-backed run
The agent read the project state, decision records, and role guidance before producing the same class of artifact.
The output cites inherited decisions, names reusable assets, and separates what is proven from what still needs a gate.
Evidence
What this case actually carries.
Run metadata
Metrics
- 5 Scenario set Five anonymized work scenarios were compared as cold and system-backed runs.
- reuse Primary signal Cases were reviewed for inherited decisions, named reusable assets, and explicit limitations.
- pilot Claim boundary Good enough to inspect; not a broad numeric ROI claim.
Source artifacts
- Cold outputs run transcript set · internal source retained
- System-backed outputs run transcript set · internal source retained
- Comparison matrix review table · public projection
Limitations
What this case does not prove.
- Small sample size.
- Anonymized scenarios only.
- The result supports a launch proof, not a statistical productivity claim.
Next